A live head-to-head comparison of OpenAI's GPT-5.3 Codex and Anthropic's Opus 4.6. The creators test both models on real production code and cover: which model wins on coding, research, speed, and creativity; live demos on real codebases; results from their custom LFG benchmark across 8 models; the team's 'Reach Test' on which model each member actually uses; and why the convergence of the two models matters beyond developers. A guest appearance by OpenAI CEO Sam Altman is included. # Channel Presentation Observations: Every (Dan Shipper / Every.to) ## Explicit Self-Description - **Channel identity as stated by host**: "I'm Dan Shipper. I'm the co-founder and CEO of Every. Every is the only subscription that you need to stay at the edge of AI. We have ideas, apps, and training as part of the the offerings that we have." - **Claimed scope — ideas**: "we have a daily newsletter where we do vibe checks. We get early access to models like Opus 4.6 and GBT 5.3 and we do day of vibe checks to tell you what's useful and what's not." - **Claimed scope — apps**: "we have Kora that helps you do your email AI. We have Monalog, which is a smart dictation app. We have Sparkle, which is a file organizer. We have um Spiral, which is an AI ghostwriter." - **Claimed scope — training**: "We also do trainings. We do live streams like this one." - **Repeated subscription CTA**: "Subscribe to Every.TO... if you want the full vibe check when it drops, please subscribe." ## Intended Audience - **AI-curious professionals across skill levels**, per staff writer Katie: "we encompass a pretty wide range of experience levels and fluencies. We have practitioners in a lot of different areas... software engineering and writing, but design and also operations folks." - **Explicit non-audience**: "They should not subscribe to every if they are AI phobes who are never going to touch AI." - **Builders/practitioners**, per Kieran: "if you are a builder, if you want to know how it is actually done, not like some kind of influencer take... like we're actually using this day-to-day." - **People who feel overwhelmed by AI releases**: "you're feeling like oh my god, like there's so much stuff happening... We've got you covered." ## Observable Presentation Patterns - **Live, long-form model review streams**: The video is a ~3-hour live stream; host notes "we're going to keep streaming for the next couple hours" and audience grew from ~5,000 to a stated ~26,000 viewers. - **Hands-on, evidence-based review format**: claims tied to concrete internal tests, e.g., the LFG benchmark ("Opus took 364 seconds on average to complete these tasks and Codex took 270") and a writing blind taste test run "in the every discord." - **Staff and guest participation**: features team members (Naveen — GM of Monologue; Katie — staff writer; Kieran — "father of compound engineering"; Austin — head of growth) and external guest Simon Willison, plus a surprise appearance by "Sam Alwin" (Sam Altman). - **Industry access demonstrated**: "We have been testing both models thoroughly for over a week" and invitations to labs: "if you are from Anthropic or from OpenAI... just DM me." - **Recurring comedic bit**: repeated "stay hydrated / make sure your agent is liquid cooled" interludes tied to the subscription pitch. - **Cross-promotion of paid events**: "we have an open claw camp coming up tomorrow... It's for paid to every subscribers. We have over a thousand people already registered." ## Claims vs. Demonstrated - **Claim**: early access to new models — **supported** by repeated references to preview testing ("We have been testing both models thoroughly for over a week"; Simon: "I've had preview access to both of these"). - **Claim**: building software in-house — **supported** by live demos of Monologue, Kora, the LFG bench, and the "Proof" markdown editor. - **Claim**: day-of vibe checks — **supported**: "we just published our head-to-head vibe check" during the stream itself. # Delivery Observation Document ## Overview This is a multi-host live stream with a rotating panel: **Dan Shipper** (primary host), **Naveen**, **Simon Willison** (guest), **Katie** (staff writer), **Kieran**, and **Austin**, plus a surprise appearance by a guest introduced as Sam. The document below focuses on the primary host's delivery, with notes on the other speakers where patterns are clear. --- ## Dan Shipper (Primary Host) ### Recurring Phrases & Verbal Habits - **"Gun to my head"** framing device, used repeatedly to force a decisive comparison: > "if you put a gun to our head and said like what is the actual difference?" > "if you kind of like gun to our head asked us which models are which which one is better and for what" - **The recurring hydration bit**, a running gag used as a segment transition and sponsorship-adjacent riff: > "it's very important that you stay hydrated. Um make sure that you're staying hydrated. Make sure that your agent is liquid cooled." > "if it's liquid, it counts, but keep yourself and keep your agent liquid cooled." - **The "convergence" thesis repeated nearly verbatim** across the stream: > "the models are converging like the uh it feels like opus got a got a little bit more codeexy and uh codeex got a little bit more opusy" - **Repeated "highest ceiling / highest variance" formulation**: > "Opus 4.6 is I think it has the highest ceiling and also the highest var variance of the two models." - Heavy **filler and repetition** in speech: "I I I," "um uh," "and uh and uh" — e.g., "Um and uh yeah, let's go. Let's get into it." ### Structural Teaching/Presentation Habits - **Structured recap loops**: repeatedly re-introduces himself and the product for new arrivals — "I'm Dan Shipper. I'm the co-founder and CEO of Every. Every is the only subscription that you need to stay at the edge of AI" appears multiple times with near-identical phrasing, alongside a repeated "ideas, apps, and training" catalog. - **Turn-taking prompts to panelists**: consistently closes segments with direct handoffs — "What do you think Naveen?", "Simon, any thoughts here? I see you leaning in.", "Katie, what do you think?" - **Calls to action layered into content**: "Subscribe to Every.TO," "just DM me. I'm very friendly, I promise," audience-count callouts ("We've got almost 8,000 people on the screen"). - **Anecdote-driven evidence**: teaches model differences through a concrete story (the Monologue iOS keyboard feature) rather than abstractions: > "I literally just monologued into Opus 4.6... And it oneshotted the solution." - **Interactive live testing**: repeatedly prompts guests to share screens and demo ("share maybe share your screen and go through the numbers," "can you demo some swarms for us"). ### Tone - Conversational, self-deprecating, high-energy despite stated exhaustion ("I'm really excited... I am I'm [ __ ] tired, man"). - Casual profanity used for emphasis: "it does dumb [ __ ]," "made us both go holy [ __ ]." --- ## Naveen - **Concise, technical turns** when answering direct questions; shorter utterances than Dan. - Recurring technical vocabulary: "long horizonts," "sub aents," "P90... P99," "one-shotted." - Notable metaphor adopted and credited by Dan: **"hungry for context"** — "it's kind a little bit more. So you're basically what you're saying is... it felt like it's hungry for context." ## Simon Willison - **Measured, first-person testing narrative**: describes methodology before conclusions — "I don't like testing against the preview models and then publishing the results because they change." - **Signature benchmark framing with self-deprecating humor**: "the Pelican on a bicycle benchmark, which is turned out to be incredibly in surprisingly influential considering it's a stupid idea." - Uses concrete anti-gaming reasoning: "I get it to draw another animal as well to see if it's... suspiciously good at pelicans." ## Katie - **Editorial vocabulary**: "AI smell," "blind taste test," "style guide," "steerability," "oneshotted," "sculpting instead of carpentering." - Direct, plain-spoken prompt style quoted verbatim: "none of these are interesting, buddy. I'm trying to make it a little more dramatic. Deviate more from the original." ## Kieran - Enthusiastic, iteration-focused vocabulary: "let's [ __ ] go," "LFG," "swarm mode," "compound engineering." - Frequent live-demo hedging: "Live demos," "apologies for this demo not working." --- ## Pattern Stability Notes - **Stable patterns (well-repeated)**: Dan's hydration gag, the "convergence" thesis, the "gun to my head" framing, the repeated Every self-introduction, and panelist handoff questions all recur three or more times. - **Isolated moments**: Naveen's "hungry for context" metaphor, Simon's Maltbook bot refusal anecdote, and Katie's "buddy" prompt appear once each — treat as single observations, not habits. - **Uncertainty**: Tone attribution across speakers relies on transcript speaker labels, which are inconsistent in places (e.g., "Naven"/"Naveen"/"Navine"); some minor attribution ambiguity exists but does not affect the patterns above. # Observation Record: "OpenAI and Anthropic Just Dropped INSANE New Models. We Tested Both Live." **Speakers (identified from transcript):** - **Dan Shipper** — co-founder and CEO of Every (host) - **Naveen** — GM of Monologue (Every) - **Simon Willison** — independent AI blogger (guest) - **Katie Parrott** — staff writer, Every - **Kieran Klaassen** — builder of Kora / compound engineering (Every) - **Austin** — head of growth, Every - **Sam Altman** — OpenAI (brief guest appearance) --- ## 1. Headline Thesis: "The Great Convergence" - **Dan Shipper:** Both models are converging. "It feels like Opus got a little bit more Codexy and Codex got a little bit more Opusy." - Opus 4.6 became better at **long-horizon tasks** (historically Codex's strength) but is **slower** than 4.5. - Codex/GPT 5.3 became **faster**, **friendlier**, with less "neckbeard senior engineer energy," though some remains. - **Naveen (agreeing):** "It's really interesting that both models are converging." He previously relied on Codex for long-horizon work; Opus 4.6 can now do it too. ## 2. Ceiling vs. Variance Framework - **Dan Shipper:** "Gun to my head" take: - **Opus 4.6** has the **highest ceiling** and **highest variance** of any coding model he's used — it did things that made the team "gasp" that 5.3 on extra-high reasoning did not do in the same setup, but it also "does dumb stuff" (confabulation, stopping server processes it shouldn't have). - **Codex 5.3** is "consistently very good" — smart, fast, low variance, but won't reach the same heights. - Recommendation framing: **Opus is still the best vibe-coder / knowledge-work model; Codex is for professional engineers** as a reliable workhorse. - **Simon Willison (partial disagreement / qualification):** "I still can't tell the difference between them and the previous ones... They're so good... how do I compare them if no matter what I throw at them, they just do the thing?" He also noted he **hadn't found a "gasp" example** himself, and asked to see theirs. He added that Opus 4.5 and Codex 5.2 were already phenomenal, making incremental differences hard to detect. - **Naveen (agreeing with the difficulty):** "Everything is almost saturated... it's now kind of our jobs to find these niche problems." ## 3. The "Gasp" Example: Monologue iOS Auto-Switchback Feature - **Dan Shipper (narrating):** Naveen had struggled for **two months** with an iOS keyboard limitation (keyboard extensions lack microphone access; app must open in background). Dan monologued the problem into Opus 4.6, which **one-shotted the solution**; Naveen massaged it slightly and shipped it. Codex 5.3 failed the same task — not because it was incapable, but because it **refused on App Store security guideline grounds**. - **Naveen (detailing Opus's process):** Opus did strong web research, found a **competitor's changelog on the App Store page**, searched GitHub via `gh` CLI, found an open-source solution, and applied it. He called the jump in **web search quality from 4.5 to 4.6** a notable improvement. The diff was small (four files) and ran on the first iOS build attempt — unusual, since 4.5 typically wasn't good at Swift. - **Naveen's explanation of Codex's failure:** Codex did initial research but kept focusing on Apple review guidelines — "it cares more than Anthropic about that, which is quite fun." - **Simon Willison (on web search infrastructure):** Anthropic appears to have upgraded search without announcing it; last year they used Brave while OpenAI had the Bing index. "They're competing on search, but they won't tell us what they're doing. So how can we evaluate if it's working or not?" (stated as unresolved.) ## 4. Parallelization / Sub-Agents - **Dan Shipper:** Opus 4.6 **automatically spawns sub-agents** by default, in both coding and knowledge work (Claude in co-work writing emails, doing research, updating Notion simultaneously). He sees this as lowering Opus's consistency floor — "it can get a little bit out of control." - **Naveen:** New behavior — Opus 4.6 spawns sub-agents even in **follow-up sessions** after feedback, unprompted. "It feels like it's hungry for context." - **Simon Willison:** Confirmed the pattern from multi-repo exploration — sub-agents fired per repo "works great," and the model "appears to have very good taste" for tool use. Noted Codex's sub-agents are still **experimental**. He also described his pattern of having Claude clone multiple dependency repos into /tmp before starting. - **Dan Shipper:** The gasp example succeeded because of Opus's **research breadth** ("compiled three different sources"), not raw coding ability. ## 5. Speed Reversal and Cerebras Speculation - **Dan Shipper (LFG benchmark results, attributed to Kieran's harness):** Historically Opus was faster; this flipped — Opus averaged **364 seconds** per task vs. Codex at **270**. Opus completed **100% of planned features**; Codex ~64%. His explanation: Codex "thrives with really well-done specifications" and is less good at "filling in the blanks"; Opus fills in blanks but sometimes wrongly. - **Simon Willison (questioning):** Asked whether the OpenAI–Cerebras deal explains the speed gain; skeptical it's live since gains look like ~30% not 10x. Cited (from Reddit, via chat) a claim that OpenAI said improvements weren't Cerebras-related, and recalled o3's price dropping to a fifth from pure inference improvements. - **Naveen:** His guess is Codex 5.3 got **more token-efficient**. He also reported firsthand Cerebras experience: P50 ~300ms but **P99 ~7 seconds** — "really fast... but not reliable." - **Sam Altman (later):** "We definitely took the feedback on personality and speed this time. We will push more on raw intelligence, but it seemed like there was such an overhang on that relative to people's frustrations around usability." ## 6. Interleaved Thinking / Mid-Task Messaging - **Dan Shipper:** Claude can receive a message while working and **decide when to handle it**; OpenAI historically queued messages and restarted execution. Codex 5.3 added "midterm redirection" — it takes new input into account but "still doesn't feel quite like interleaved thinking." - **Simon Willison:** Called the Claude behavior "the killer feature of Claude Code for web"; called its absence in OpenAI products "a surprising omission." - **Naveen (architectural theory):** Codex cloud was built (~April 2024, per his recollection) as one-big-prompt/container-per-task; Claude Code's container persists and can be rehydrated. ## 7. Knowledge Work Strategy Convergence - **Simon Willison:** Claude Code → "co-work" rebrand makes sense because "you can't tell knowledge workers to go use Claude Code." The Codex app (released days earlier) may be neutral enough branding; "Codex and co-work are now at the app level competing for the same exact audience." He criticized Codex **automations** as "half-baked" ("it only works if your laptop's open"), while acknowledging automations are a golden feature of Claudebot/Moltbot/OpenClaw. - **Dan Shipper (agreeing):** Anthropic realized early that a great coding model can do any knowledge work; OpenAI's earlier split (ChatGPT for vibe coding, Codex for pros in VS Code/terminal) is dissolving. - **Sam Altman (confirming direction):** "I think that will converge for sure... one AI... with all their memory, all their context... whether it's a big new scientific discovery or drafting this email." ## 8. Adaptive Thinking - **Dan Shipper:** Anthropic replaced extended thinking with **adaptive thinking** (model decides for itself how long to think). He hypothesizes it's the source of some "maddening mistakes" — confabulation when the model decides not to think hard. "I think adaptive thinking is the future. And also, I think they really need to tune it." - **Simon Willison:** Had not tested it yet. - **Naveen:** Had not tested it; noted he seemed to have a "high variance" experience in his few days of testing. ## 9. Benchmark Numbers (Naveen, reading blog posts live) - **Naveen:** Codex SWE-bench has slowly increased; **terminal bench is "crazy high" at 77**. Opus 4.5 agentic coding was 59% vs. Codex ~65%; Opus 4.6 terminal bench now ~77% — "Opus is going near to what Codex 5.2 does." - **Naveen (efficiency theory):** Codex's speed comes from lower output tokens per solution. ## 10. Knowledge Cutoffs and Model-Assisted Training - **Simon Willison:** Opus 4.6's knowledge cutoff is **the same as 4.5** (no bump); he couldn't find Codex 5.3's cutoff because it wasn't in API docs yet. Noted OpenAI boasted 5.3 is "the first model they've trained where the models help them do the training" — "I'm a little surprised... I expected that for the past 12 months." - **Dan Shipper:** Release cadence is now every 2–3 months with subtler bumps, but "I feel like I can tell the difference." - **Sam Altman (corroborating):** "5.3 Codex was the first time a model that we made was significantly reliant on the model and the tools to get it built... shame on us if we don't adopt it the first and the best." ## 11. Pelican-on-a-Bicycle Benchmark (Simon Willison) - **Simon:** Opus 4.6 produced "a very good pelican riding a bicycle" — better pouch, a hook, "little ears" — a clear improvement over 4.5 ("facing the wrong way"). Later comparison: Opus 4.6 beat Codex 5.3 ("Opus is almost setting a new level for this benchmark"); Codex 5.3 was better than 5.2 (which "scrambled the thing"). - **On benchmark contamination:** "Very confident they aren't [training on pelican images]" — labs told him they don't, and he cross-checks with other animals (e.g., a New Zealand kakapo). "I live for the day that a model draws a perfect pelican and then sucks at something else." - **On eval gaps:** "They're always talking about terminal bench... I want to know if I give you a screenshot of a web page, can you rebuild that web page perfectly?" ## 12. Writing and Editing Tests (Katie Parrott) - **Method:** Same transcript (agent-native camp) given to Opus 4.5, Opus 4.6, and GPT 5.2 with prompt "write a 1500-word article"; tested both vanilla and with Every's style guide/skills. - **Findings:** - 4.6 one-shots decently, gives **context on first mention**, smoothly incorporates quotes (sometimes "too eager to quote directly"). - Blind taste test in Every's Discord: **the community preferred Opus 4.5**; **Katie personally preferred 4.6's cadence** — an explicit unresolved tension. - Persistent "AI smell": e.g., "sets of three with no 'and' before the last" (flagged by managing editor Eleanor). - **Dan Shipper (live critique):** Rejected an opening line ("built a reading app over break. It took a week.") as AI cadence attempting but failing a "holy \[...\] moment." After Katie prompted "none of these are interesting, buddy... deviate more," the model produced: "For the last 60 years, building software meant the same thing... That's over." — which Dan praised. - **Simon Willison (on prompting style):** "I love that prompt... the fact that you put comma buddy influenced the way it behaved." He uses informal language deliberately. - **Simon on his own writing:** "None of it's written by AI" for opinion pieces; READMEs/documentation are now mostly AI-generated ("it has no voice... it just needs to be boring and accurate"); uses AI for proofreading — noting models only became trustworthy at spotting spelling errors around Claude Sonnet/Opus 4. - **Codex 5.3 writing test (live, Dan + Katie joint):** First-ever joint vibe check. Dan: "I am not particularly impressed" — flagged "X is not Y" constructions and breathless framing of uninteresting details as AI tells. Katie noted it was "almost the same that 5.2 did." Both acknowledged it one-shotted with correct core principles and decent structure. - **Katie (on model choice):** Long preferred Sonnet's fluidity but was won over by Opus's sentence structure and technical-explanation ability; 4.6 feels **faster/more steerable** than 4.5. - **Katie (on capability unlock):** "I'm not a technical writer... and now I totally can" — because she can pull codebases down and have Claude summarize them. ## 13. Swarms / Agent Orchestration (Kieran Klaassen) - **Kieran:** "Agent swarms" released with 4.6 in Claude Code. Core take: **"Anthropic has a vision on how to build with AI that's way further than OpenAI as of now."** Evidence: sub-agents in OpenAI only experimental after he'd asked for them for six months; his compound-engineering flow takes ~30 minutes on Codex vs. ~5 minutes on Anthropic due to parallelization. The LFG command ran **8x faster with swarms at the same quality score**. - **Kieran's qualifications:** Swarms are **token-hungry** ("not easy on tokens... but that's not the point"); he restricts swarming to **implementation** rather than planning; quality of output is the same — swarms mainly add speed. His live swarm demo **failed** (setting apparently not enabled post-release, no communication from Anthropic) — he noted "it works both ways... it's just slower." - **Simon Willison (skeptical of maximal swarm claims):** Cited the "fast render" project (1.6M lines of Rust, 2,000 parallel agents) — "you can build a web browser in Rust with 20,000 lines of code that's just as good... they didn't really demonstrate swarms building something you couldn't have built without them." He prefers reasoning about 2–3 sub-agents; calls large-scale coordination "a very interesting open area of research." - **Kieran (on OpenAI):** Speculated there's "clearly a reason they're not jumping into it... I would love to hear the reasoning." - **Dan Shipper (speculation, flagged as his perception and possibly wrong):** OpenAI's culture treated the agent as "a power tool that does exactly what we say... built for people that wanted to use it that way," versus Anthropic's more experimental ethos. ## 14. Model Preference Split Within Every - **Dan Shipper:** ~50/50 Opus vs. Codex. - **Kieran:** ~80/20 (Claude-leaning; uses Codex when he "needs a lot of precision"). - **Naveen:** ~80/20 the other way (Codex-leaning; uses Opus when prototyping/needing creativity — "giving one line and it will just run with it"). He found Opus 4.6's **planning** ("felt like reading a novel") superior, but Opus's **implementation** of the plan was "not great," while Codex **one-shotted** the same plan with no build errors. He also noted Opus "needs to get better at Swift." - **Austin (baby vibe coder):** Built an independent-movie-theater app (Lello) in Claude Code, then had Codex 5.3 audit it — "oh \[...\], this is actually really good." Surprised by Codex's speed. Found value in **trading off between models**: "each model... solve[s] problems the other model is blind to." - **Kieran (practical note):** "If you have a ChatGPT subscription, go use Codex — you already pay for it... It won't give you a major edge if you just use one over the other." ## 15. Sam Altman's Guest Segment - On convergence: "That's interesting. That's kind of a reversal of what people said about the previous model... maybe there is more convergence going on than we thought." - On strategy: confirmed convergence toward one AI for all seriousness levels of work. - Personal use: built a "magic to-do list" app where items get automatically done or escalated; first version in a few hours, improving continuously — "the nature of software is changing when you make it so low friction... it all feels continuous and it never is done." - On OpenAI internal adoption: 5.3 was the first model significantly built with the model and tools. ## 16. Miscellaneous Claims - **Naveen:** Monologue iOS launch is **February 17th**; dictionary/modes sync feature being built at time of stream. - **Simon Willison:** Uses Claude Code for web in "YOLO mode" from his phone while walking the dog; keeps an "everything" environment with full network access — safe because his code is mostly open source and he doesn't expose secrets. Moltbook signal-to-noise has collapsed ("hardly worth even trolling it anymore"); his citation-bot (Opus 4.6) **refused to post unverified claims live on stream** — "alignment achieved." - **Dan Shipper:** OpenAI reportedly told (via chat/Reddit) that speed improvements weren't Cerebras-related — flagged as secondhand and unverified ("Lilo, don't know if we can trust you yet"). - **Katie Parrott:** Suggested a future writing harness that swaps between Anthropic and ChatGPT models. ## Preserved Tensions & Qualifications - **Dan/Naveen vs. Simon:** Dan and Naveen report a clear ceiling/variance difference; Simon explicitly could not detect meaningful differences and asked to see the evidence. Both positions stand unreconciled. - **Katie's blind test:** Community preferred 4.5's writing; Katie preferred 4.6. Both recorded. - **Kieran's swarm claims** are enthusiastic but qualified (token cost, planning-phase limits, and a failed live demo); Simon remains skeptical of swarm scaling claims. - **Dan's explanation of OpenAI's orchestration lag** is explicitly framed as his perception, possibly wrong, with an invitation for correction. # Speaker Attribution & Provenance Record **Video:** OpenAI and Anthropic Just Dropped INSANE New Models. We Tested Both Live. **Video ID:** `01a0f15a-ada5-7d93-8642-dd803280c0a7` **Format:** Live stream (~3 hours), hosted by Every (every.to), streamed via Streamyard/X/YouTube/Every website. Peak viewership stated as ~26,000. ## Identified Speakers & Roles ### Host / Creator - **Dan Shipper** — Co-founder and CEO of Every; principal host of the stream. Introduces segments, demos vibe checks, conducts guest interviews. ### Every Team (Co-hosts / Participants) - **Naveen** (introduced as "GM of Monologue") — Guest co-host; Every team member; primary builder of the Monologue dictation app; daily Codex user. - **Katie Parrot** — Staff writer, Every; conducted writing/editing tests on Opus 4.6; joined mid-stream. - **Kieran (Klaassen)** — Every team member; identified as "father of compound engineering"; builder of Kora and the LFG benchmark; joined mid-stream; internet dropped during demo. - **Austin** — Head of growth, Every; joined late to demo an independent movie theater app ("Lello") built with Claude Code/Codex. ### External Guests - **Simon Willison** — Independent blogger (simonwillison.net); joined as special guest for ~2 hours; discussed Pelican-on-a-bicycle benchmark, deep research, Claude Code for web, Molbook bot. - **Sam Altman** — Identified as "Sam Alwin"/"Sam Alman" in transcript (OpenAI CEO); joined briefly near the end on short notice; discussed Codex 5.3 and convergence; explicitly stated he had to drop in a couple of minutes. ### Referenced Non-Speaking Parties - **Dario Amodei** — Referenced/invited ("Daario, if you're around"); did not appear in transcript. - **Alexander** (OpenAI team, via chat) — Helped resolve Simon's Codex 5.3 access issue; invited but did not join stage. - **Nat Eliason**, **Trevor**, **Hashim** (Reddit), **Lilo** (chat) — Referenced or chat participants only. - Anthropic and OpenAI staff — Referenced as possibly present in chat; invited to DM. ## Promotional Segments (Flagged) - **Every subscription promotion** — Repeated throughout by Dan Shipper: "Every is the only subscription you need to stay at the edge of AI"; promotion of ideas, apps (Kora, Monologue, Sparkle, Spiral, Proof), and training offerings; calls to visit every.to and subscribe. - **"Stay hydrated" / "liquid cooled agent" segments** — Recurring humorous interstitials (references to tap water, Smart Water, Stanley cup, Naveen's 1.5L bottle); function as recurring breaks, not third-party product endorsements. - **OpenClaw Camp promotion** — Lobster boil–themed paid-subscriber event "tomorrow at 12 p.m. Eastern," featuring Nat Eliason; "over a thousand people already registered." - **Writing camp promotion** — Katie mentions a writing camp "later this month" for subscribers. - **Agent-native guide promotion** — every.to/guide/agent-native. - **Simon Willison's blog/newsletter promotion** — simonwillison.net (cross-promotion by host and guest). - **Monologue promotion** — monologue.to; iOS launch date stated as **February 17th**. ## Referenced Sources & Claims (Provenance) - **Model releases discussed:** Anthropic Claude Opus 4.6 and OpenAI GPT 5.3 Codex — released the same day per speakers; every.to vibe checks cited as their own published testing (head-to-head plus individual reviews). - **Blog posts / eval numbers:** Naveen reads SWE-bench and Terminal Bench numbers (e.g., terminal bench ~77%) from vendor blog posts on screen; these are the model vendors' figures. - **OpenAI blog claims (third-party):** Codex 5.3 "first model they've trained where the models help them do the training" (cited by Simon); screenshot-rebuilding boast (Simon). - **Anthropic announcement claims (third-party):** "best-in-class for coding and professional work," "most agentic model," adaptive thinking replacing extended thinking (paraphrased by Dan). - **LFG benchmark** — Kieran's internal benchmark (landing page, 3D island scene, earnings dashboard, e-commerce site); claimed results: Opus completed ~100% of planned features vs Codex ~64%; Opus avg 364s vs Codex 270s. - **Anecdotal internal tests:** iOS keyboard-return feature one-shot by Opus 4.6 (Dan/Naveen); dictionary sync plan/implement comparison (Naveen); blind Discord taste test where community preferred Opus 4.5 writing (Katie). - **Speculative/unverified claims flagged as such in stream:** OpenAI/Cerebras deployment (Simon explicitly uncertain; chat-sourced claims from Hashim/Reddit and Lilo); Simon's uncertainty whether preview Codex 5.3 equals released version; Anthropic search infrastructure changes (Simon: "they won't tell us"). - **Sam Altman's claims:** OpenAI "took the feedback on personality and speed"; 5.3 Codex "significantly reliant on the model and the tools to get it built"; his personal to-do-list app anecdote. ## Uncertainty Notes - Speaker names transcribed with variants ("Naveen"/"Navine"/"Naven"; "Sam Alwin"/"Sam Alman"; "Kieran Clen"/"Kieran Classen"); no curated metadata was supplied to confirm spellings. - Third-party claims (vendor benchmarks, blog claims) are recorded as attributed, not verified. - No transcript-sourced information about comment availability beyond live chat participants being read aloud on stream. Hello. How's it going, Naven? \>\> Hello. Good. Good. I'm really excited about today. \>\> Me, too. I'm really excited. I am I'm \[ \_\_ \] tired, man. I was up all night last night. I mean, you saw me at like seven. You're like, "Dan, you look tired. Go to bed." \>\> And you're still working at 8, right? 8:30, maybe. You left off. \>\> No, I was working until midnight. Midnight. Wow. \>\> Yeah. \>\> And you have a flight to catch. So that's \>\> And I have a flight to catch. I had a bunch of big meetings this morning. But I'm so excited. Today is Opus 4.6 and GBT 5.3. We have been testing both models thoroughly for over a week. They're both really great models. We have um uh we have a some in-depth VI checks that we've done on each individual model and then we've gone headtohead. Those are going to be live on every very shortly. We are just um uh finishing updating the copy uh for those pieces, but I'm just going to get a I'm just getting a status check uh as um so that's going to come out soon. While it's coming out, I'm going to give you a preview of both models because I think we're all here for uh like which model's better, better for what. And uh and we have we have your answers. We have your answers. Um if you're just getting here, I'm Dan Shipper. I'm the co-founder and CEO of Every. Every is the only subscription that you need to stay at the edge of AI. We have ideas, apps, and training as part of the the offerings that we have. On the idea side, we have a daily newsletter where we do vibe checks. We get early access to models like Opus 4.6 and GBT 5.3 and we do day of vibe checks to tell you what's useful and what's not, what's good and what's not. Um, from our hands-on testing, we also have a suite of apps that we give you. Um, so we have Kora that helps you do your email AI. We have Monalog, which is a smart dictation app. We have Sparkle, which is a file organizer. We have um Spiral, which is an AI ghostwriter. Check those out. And we also do trainings. We do live streams like this one. I want to let's get into the the actual good stuff. Subscribe to Every.TO. Um if you want the full vibe check when it drops, please subscribe. Um and uh yeah, let's go. Let's get into it. So, let's start with the overall my overall take. And I'm actually just going to like read you from um a little bit of like what we're what we wrote. We have a whole mini site coming on every every.0. O um with all the headtoheads and all that kind of stuff. But basically like what's really interesting about both these models is they're great models and they feel like they're converging. So uh on the Opus side, Opus 4.6 is definitely the best coding model that I've ever used um in terms of the heights that it reaches. Like it it did a few things that blew both of us, me and you and Navine, away, right? \>\> Yeah. Um, but it also is like has some kind of like annoying claisms where it does stuff that you don't expect and it kind of goes off the rails a little bit which is it's weird. It's like both amazing and like oh wow that did something that I didn't really expect it to do. Um and um but what it feels like a little bit is um it's much more thoughtful. It is much it's much better at just staying on a task for a long time and just like doing it. And it feels a little bit like Enthropic took what the criticism of Opus versus Codeex and made put a little bit more codeex into Opus like just a little bit more. Um it's much better at those long ranging tasks. It's um it's an incredibly good coder all that kind of stuff. Um the same thing for codec. So C codeex 5.3 we've been testing it as well. Um it's uh it's the same codecs that you you love. If you're like a senior engineer like Naveen, Naveen uses codeex every single day to build monologue. Um it's uh and and 5.3 is really great, but uh it has a few twists that make it a little bit more like Opus. So Codeex is much faster. Codex 5.3 is much faster, at least for me. Um it's also it feels like a little bit of a friendlier model. Like it doesn't have as much of the roboticness and um like just neckbeard senior engineer energy. It still has some of it, but it's like it's less. So, I feel like both models, if if you if the models are over here, they're kind of like going like this a little bit. Um, and it's it feels like both Anthropic and OpenAI have identified like a little bit of like what's the ideal like model to build and they're both like kind of going in that direction. What do you think they mean? \>\> Yeah. Yeah. Say I have the I think we had this discussion, right, where u uh Codex is like really good at those long uh horizonts, right? like you can give it a task and it it doesn't like you know stop in the middle. It just keeps going. And with Opus 4.5 you have like this quick iteration but it doesn't like do the what do you say difficult tasks uh really well but with 4.6 now I can say it's like oh yeah it can do the uh long horizont now so which is really interesting that both models are converging. I really like Codeex because it I can just throw a problem at it and then it keeps do doing the work and Opus is now able to do that now. So interesting. \>\> I I totally agree and I want to take a break. We have we have um I I have another like uh take on okay what is the difference if you put a gun to our head and said like what is the actual difference? I have a take for you. But before we do that, I just want to make sure um we are talking about uh these these model drops. Very exciting. And in the midst of all of this excitement, all this energy, all this not sleeping, it's very important that you stay hydrated. Um make sure that you're staying hydrated. Make sure that your agent is liquid cooled. Take a break. Take a drink of water. I get my \>\> very important. Thank you. Um so, uh thank you for that PSA. Thank you for listening to that, Naven. It looks like you took a big a big drink and that's that's where all of Naveen's productivity comes from. It's staying hydrated, folks. \>\> Um, \>\> just have a 1.5 liter bottle with you. \>\> Okay, so let's go back. So, we talked about we talked about these models feel like they're converging a little bit. Um, but what's uh like if if if you ask me gun to my head like what's different about them or how do they um how can you think about what they're like? Opus 4.6 is I think it has the highest ceiling and also the highest var variance of the two models. So by highest ceiling I mean it did something that like made us both go holy \[ \_\_ \] We've never seen a coding model do this and 5.3 on extra high reasoning did not do it in the same setup. And I'll tell you about what that is um in a second. So it has that high ceiling. It has that like gaspworthiness and also it does some dumb stuff like it it can like uh it has this it has a new mode a new thing that it does called adaptive thinking where sometimes it doesn't think very much and when it doesn't think very much it kind of like it can it can confabulate it can um go do things that you didn't like ask it to do that kind of thing it's also like uh one thing that you'll notice about this model which is really cool is it's very parallel so it it just automatically spawns all these sub aents and I think that that makes the floor of what it does like a little bit lower like it's less consistent because it can it can get a little bit out of control. Um on the flip side codeex I I I think co GPT 5.3 C codeex is um it's a very high it's a very good model and it's also very consistent but it's not going to reach the same heights as Opus and it's not going to do the dumb the same dumb \[ \_\_ \] as Opus. It's just going to be very good. What do you think Naveen? \>\> Yeah, I agree. And I think even uh when I'm look right now I'm just going through the blog post that both posted and it kind of like has the numbers right like eval numbers that they have. Maybe we can also look at that uh to back this up. So which is really interesting. \>\> Oh yeah what are the numbers? Share maybe share your screen and go through the numbers. Um, \>\> yeah, I'm shame. So, \>\> and I think Simon Wilson is going to join us. So, I'm just going to um grab the uh Streamyard. Cool. \>\> So, yeah, in the meantime here you can see uh you know 5.2 to codeex is like really good at it but you can see the SWE bench slowly increased uh and but the one thing I think uh when you mentioned Dan there's some vibe difference the way out uh codeex like outputs I think it's maybe the output tokens it's using it's really efficient as well that way it's fast you mentioned like it's fast right that means the number of tokens it's using to come up with the solution it's like slow uh low so that's why we are getting like faster outputs from And uh terminal bench is like crazy high 77 and \>\> really interesting \>\> coming to 4.6 opus here if you see like why I like 5.2 Codex is like my go-to model. Right here you can see why that is like it's kind of agentic coding is 59% according to their at least 4.5 \>\> and then codex is like around 65 but now opus 4.5 not sure what the terminal bench I think terminal bench is like around 77% now so \>\> interesting. \>\> Yeah. Um so opus is like going near to what codeex 5.2 do it. So, yeah, \>\> that is really interesting. Um, so we've got a special guest coming on the stream. Uh, pause button on the top right. Um, super excited to have him. Um, we'll also we'll also have have him have him tweet. But, uh, I just want to I'm just going to add you to the stage. Simon Wilson, welcome. \>\> Hello. Hey. Good to be here. Yeah, today's wild like like the two models came out within about 15 minutes of each other. Right. \>\> Right. I mean like have a little have a little thought for us poor model reviewers. \>\> Right. Absolutely. Oh dear. At least it's not It was um uh Meta released um Llama 4 on the weekend and I still haven't forgiven them for that but \>\> Oh my god. \>\> But yeah, today was a busy day. \>\> Yeah, totally. I remember actually uh also GBG5 and 4.1 came out on the same day and so that was a very similar feeling for me. \>\> But yeah, these two I so I've had preview access to both of these for for like a couple of weeks for the codeex one and a few days for the clawed one. And I'll be honest, I still can't tell the difference between them and the previous ones. Like it's getting so hard so hard to analyze these things now. They're so good, \>\> you know, throwing just absurdly difficult things at them and they were doing it and that's great, but it's like I don't how do I compare them if no matter what I throw at them, they just do the thing, you know? So, yeah, I'm finding that really difficult. I've got um I'm having trouble getting a Pelican out of codeex 5.3 because it's not showing up to me in any of the interfaces. And I I don't like testing against the preview models and then publishing the results because they change. Like I don't know for sure that the codeex 5.3 they released today is the same as the one I was playing with the other day. \>\> Yeah, it's hard to know. I So I gave you a I gave you a tweet. Please please hit retweet or quote tweet that so that we get uh we get people who are interested in your takes on here. \>\> Yep. \>\> Um and there should be if you if you see on the top right there should be a uh a plus button to stream for your account if you want. But let's keep going with let's keep going with the takes. That's that's actually really interesting. I think that Naveen you had a similar feeling that it's just hard to tell the difference between the two. \>\> Yeah. I I think I was talking with Dan where what should we test now like everything is like almost saturated what what problem that we are throwing at both models are doing it great job at it. So now it's kind of our jobs to find these like niche problems and you know trying to find the difference what it is really good at. So yeah, \>\> totally. Here's actually here's actually the thing that um so so ju just Simon for your for your uh benefit the our overall take and this is coming out on every very soon but our overall take is the models are converging like the uh it feels like opus got a got a little bit more codeexy and uh codeex got a little bit more opusy and they're it's it's they're not the same but they're starting to converge a little bit. Um, but if we had to like kind of pick like gunhead which is better, the overall take that we have right now is that um, uh, Opus 4.6 has the highest ceiling of any coding model we've ever used. Like it it does it did a couple things that made us totally gasp and we've never seen any model do. \>\> Okay. \>\> And it's got it's got a high variance. Like it also does dumb \[ \_\_ \] basically um, sometimes. Whereas codeex is smart. It's really fast. Um, and it's much more consistent. So, it's like consistently good. Consistently very good, but it's not going to reach the same heights. \>\> Interesting. \>\> Yeah. Was that How did that work with your experience? \>\> What I've been doing because I' i've I I I've been basically using them almost randomly in between each other, but partly just to avoid burning up my token allowance on one. You know, if you've got both, you're less likely to to run out of tokens. Um, and yeah, I've certainly noticed that codecs I'll fire it up for whatever the task is. Like, um, I often have them do things like refactor my git, my my git history to clean things up and so on, and it'll always just do exactly what I told it to do, which is great. What I'd love to see those examples of things that Opus did that made you gasp because that's the problem I've had is I just haven't been able to find something where one of those models has has just blown me because because Opus 4.5 and Codeex 5.2 too were already phenomenal models. You know, \>\> that's true. \>\> Finding that that incremental difference is getting so difficult. \>\> I will I will I will set it up and Naveen, you add some color if I if I'm missing anything. So, um the the the the specific example that made us really gasp is Naveen builds Monologue. Um if you don't know Monologue, you should all know it. It's monologue.to. It's a smart dictation app. It's sort of like Whisperflow or Super Whisper. Navine builds that all by himself. he mostly has built it with codeex and he competes just with him with uh using these models with uh you know super whisper whisper flow these companies with like huge huge amounts of employees and amounts of funding um and he's a real codeex guy and one of the things he's been trying to do if you're familiar with how um uh how these uh how these things work on iOS is if you want to do a voice transcription app on iOS you make a a special keyboard for it and Uh, and basically, and basically when you uh when you activate the keyboard in order to get it to start recording, you have to flip to um flip to the app, flip to the monologue app and then flip back to the app you want to transcribe in because Apple has all these restrictions around like who can record and when and you have to record starting from a open app. Does that uh from from an open monologue app does that does that about right? \>\> Yeah. Yeah. So the main reason is keyboard extension doesn't have the microphone access. So to get the microphone access you need to open the main app keep it running in the background and then you need to come back to the whatever app that you're using. \>\> Exactly. And so Naveen has has observed that some of the competitors have the have this thing where it automatically takes you back to the original app after you activate uh the transcription service. And he'd been kind of like working on it in the background for like two months like trying like set sending Codex off and being like okay try to figure this out try to figure this out. And like over those months because it was a hard problem, he was a little bit procrastinating on it is what he told me. And over those months he had kind of come to kind of the approach that he wanted to take but hadn't been able to like actually get it built. So that's the setup. Like I I don't know when we got access to 4.6 but I think it was two nights ago I was sitting in the office with Naveen and Nine was like I've had this problem for a while. Do you want to try to fix it? And so I literally just monologued into Opus 4.6. six. Hey, we have this problem. Can you figure it out? And it oneshotted the solution. Um, it it built it. \>\> Naveen spent like a little bit of time massaging it and and and then pushed it and now it's like a feature in the app and I have no idea how the repo works. Like I just I just built the feature that he had been trying to build for forever. And what's really interesting is we did we did obviously the same test with 4 point uh with 5.3 codeex. \>\> Yeah. \>\> And it didn't work. Um, and it didn't work for a particular reason. It's not that Codex is like stupid or something. It's that um, Codeex was extremely uh, focused on the fact that doing something like this might uh, run a foul of like the app store security guidelines. \>\> Oh no. \>\> Um, and and it basically was like you can't do this. Um, Navin any you want to add or anything that I'm missing there? \>\> No. So just to like look at go a little bit deep on what opus 4.6 did and you know claude uh codex did here is that it did a really great research online. I think there's one big jump that I I'm seeing from 4.5 to 4.6 is how well it's doing the web search online. So here it like found one of the competitors change log on app store page \>\> nice \>\> and found a solution and then using uh gh like you know github cla it started searching the whole github search is there like any open source codebase test and then it found a solution and then it just applied that so which is really interesting \>\> that's something I've not I mean um I've noticed codecs do do this in the past as well the models are getting really good at looking things up in the code like I've seen them if they're using if I've got like an HTML thing with a CDN link, they'll be like, "Oh, based on that CDN link, I should check out this GitHub repository to get the the the the original code." It's amazing. But that the the web fetch stuff is interesting as well because Anthropic I think Anthropic upgraded their search at some point and I don't think they told us about it which is ridiculous because like they were last year they were using Brave for search and I always felt that was like open big advantage is open had the Bing index and and Anthropic just had the Brave index and the Bing index had been around for longer but I'm saying seeing such good results from web search on anthropic now that I don't know maybe Bing maybe Brave got really good. Maybe they're doing something else. Open AAI are running their own crawlers now, but again, they won't tell us what search how their search infrastructure works. So, it's crazy. They're competing on search, but they won't tell us what the what they're doing. So, how can we evaluate if if we think it's working or not? \>\> Yeah. So, that that that like really blew me away where Opus able to figure it out. But with COEX the on the other and what happened is it ha it did the initial research and stuff and then it's just like keep on looking the Apple review guidelines somehow it's like even care \>\> more than than anthropic is quite fun. So yeah, that's what happened at least with this same thing like I've been hitting this particular issue with for past two months, right? I always pick off a agent, try to find a solution and always get a similar \>\> question about this. Did the models have the ability to test the code? Are they like running the IO simulator and stuff like that? \>\> Uh it has Xcode build MCP which I think Dan doesn't have it like at least I I use it. It didn't. Uh the that's the interesting thing is uh usually for me Opus 4.5 is not really good at swift things. \>\> Maybe from web it has like beautiful designs. It runs on the first try but usually on swift I don't get it on first try. But at least this solution it's a quite small diff change. \>\> Yeah. \>\> Like it's not 100 line like you know uh files. It's kind of four files touched. It ran on the first ISO. That's good. \>\> Yeah. \>\> Yeah. We are definitely I've definitely been able to use the iOS sim simulator with cloud code and codecs and it it works. It works. It works okay. It's it doesn't work amazing. Uh it's like taking screenshots and then it's like it's just very imprecise and slow but it it does work and it can do end to end to end testing. \>\> I wonder if that's going to be a useful place where we can identify the difference between the two is the screen how good are they at interpreting screenshots which again I mean I don't think they're even publishing benchmarks relating to that stuff. They're always talking about terminal bench and stuff, but actually I want to know if I give you a screenshot of a web page, can you rebuild that web page perfectly? And I saw openly I did boast about an example of that in their release for the codeex model. Um, so it is something that they're thinking about. \>\> Um, I I don't I don't totally know, but I I will say so I'm going to show I'm going to share the vibe check that we just published. Um, I want to show you uh some of the things that we found. Um, so basically, so this is our vibe vibe check page. One thing that you should know, this is just for Opus. We have a combined one soon, but one thing you should know is I totally vibe I totally vibe coded this with um, \>\> it's not showing in the live stream. \>\> Oh, no. There we go. Uh, \>\> almost. Yes. \>\> There we go. \>\> Okay. So, here it is. Um, I vibe coded this last night with Opus and I tried to do the same thing with codeex and it's just it's just not good enough at design. Um, for for whatever reason like I just told Opus like just just do this run in a loop until it works and it it didn't do exactly this like there's definitely some tweaking that that we've done over the last \>\> uh several hours to to get this done. But this is this is the kind of thing that I I I don't know if it's really about the screen understanding. I would guess that um Codex has really good screen understanding. But what it doesn't what what Codex doesn't necessarily have is just this that like empathy thing. Like it just doesn't quite get it. Like it just doesn't have the the vibe as much. I do think that has improved a lot. I think that it has improved a lot over like in 5.3. It's better than it was in 5.2. And in 5.1 it it it it feels faster, it feels more responsive, it feels more fluid, but it still gets stuck in those things. And I think things that require aesthetic taste, it's still not amazing at \>\> Interesting. So I just checked the um the knowledge cutoff on 4.6, Opus 4.6, it's the same as Opus 4.5. They because I they've been getting much more aggressive about updating their knowledge. Like um Opus 4.5 is really good at UV, the Python tool, which only came out a year ago. \>\> That's kind of But no, I I We it does it looks like we didn't get a bump on knowledge cutff for this new model. \>\> That's interesting. I wonder why. I mean \>\> codec 5.3 I can't find the knowledge cutff because it's not in the API yet. That's the the API documentation is where you get the knowledge cutff information. \>\> That's interesting. One thing that I I I I think we should say though is we maybe there's no model cutoff bump but like these model releases are coming faster and faster. Like when was the last one? like a month ago or something like open two months ago or three months ago. Um, and these and the bumps maybe are like a little bit more incremental than they were when it was like once every year, but they're I feel like I can tell the difference. Like it's it is subtle. You have to like find the edges to to do it. It is subtle. \>\> Um, but they're they're there and they're happening every two or three months and it's like it's pretty insane. \>\> Yeah. the um OpenAI in their blog post about Codex 5.3, they boast that this is the first model they've trained where the models help them do the training. So there I'm a little surprised by to be honest. I expected that that that's something they'd have been doing for the past 12 months. But um but yeah, that's that's going to be because maybe that's that's maybe that accelerates things as well. \>\> It's possible. It is totally I I think I think it totally does. We see that we see that in internally too. like I I think we're just moving a lot faster on the products that we build with this stuff. So, \>\> oh my god, the the November effect was real, right? Codeex 5.1 and Opus 4.5 just accelerated the coding agent thing so much. I know so many people who over the Christmas holiday they were poking around stuff going, "Whoa, when did this get good?" Because it used to be you had to put some work in yourself and keep an eye on what it was doing to get good results. Now, most of the time I'll prompt something and I'll get a few thousand lines of code that does exactly what I needed back like five minutes later. \>\> Totally \>\> extraordinary. Yeah. \>\> Totally. So, I I we just published our head-to-head vibe check and I would like to I'd like to take us through this, but before we do that, I just want to for people who just joined, so we're almost at about 5,000 people here. We're going to keep growing. We're going to keep streaming for the next couple hours. Um, if you're here, we are doing a vibe check on uh Opus 4.6 six and Codeex 5.3. They're both incredible models. We've been testing them internally and Simon has also been testing them on on on his side and we're we're talking through what we notice, what we like, what we don't like, um what's useful and what's not. Uh we I'm Dan Shipper. I'm the co-founder and CEO of Every is the only subscription that you need to stay at the edge of AI. We publish these kind of vibe checks every time new models come out. You should check them out at every.to. We also publish apps. So we build apps internally and we use uh we test these models on these apps to build them in production. That's how we can give you good opinions on on on what to use. We also do training. So we do live streams like this. Um Naveen, do you want to introduce yourself? \>\> Yeah. Hey everyone, I'm Naven GM of Monologue. It's a smart voice to text Mac app and we are also releasing iOS app soon. So yeah uh really excited to be here talking about Yeah. \>\> Simon, introduce yourself. I'm sure everyone knows you, but introduce yourself. \>\> Hey, I'm Simon. I'm I guess I'm a blogger. I have a blog. I write about AI stuff on my blog. I've been doing it all on an almost daily basis for over three years now. Um I also I build open source software for talking to models. So, I've got all sorts of like fingers in in in the pie around how these things work and what you can do with them. And um I'm responsible for the Pelican on a bicycle benchmark, which is turned out to be incredibly in surprisingly influential considering it's a stupid idea. You get the get levels to draw you an SVG of a pelican riding a bicycle, which is funny because bicycles are difficult to draw and pelicans are difficult to draw and pelicans can't ride bicycles. So, there's a whole bunch of challenges involved in this. Um, and I haven't got codeex 5.3 working on anything at the moment. It's not showing up in my Codeex Apple Codeex CLI, which is frustrating. \>\> Weird. Do you do you actually have Pelican benchmarks that you can show us live on this show? \>\> I can absolutely show you what I got from Opus 4.6, six. But I want \>\> Okay, please show us opus. Please don't show us. We'll go We'll go to our vive check for a second, but I would think the people want to see the pelicans. \>\> Let's do this. Um, share screen. Share screen. Uh, it's doing How do I stop my phone from showing all of my notifications on my Mac? That's really annoying. Uh, I just got rid of silent mode like uh do not disturb. \>\> See, that works for everything except for the Here we go. So, this right here is Claude Opus 4.6 during riding a bicycle. It's good. I like the the the the bicycle frame is a little bit off. Like, it's it's a bit of a wonky frame. The the Pelican itself very good. Really good pouch on it. It's got a little hook at the end, which I haven't seen from any of the other models. Little ears. I'm going to have next time I see a Pelican, I'm going to have to check if it's got little ears. But genuinely, this is a very good Pelican riding a bicycle. Um, compared like this one. I got that one out of Kimmy 2.5 the other day. It's definitely Kimmy's. \>\> Do you have Opus 45 that you could show us? \>\> Um, Opus 45. Yeah, I should do. Let's do Opus. That one's uh there we go. \>\> So that Opus 4.5 one wasn't actually very good at all. \>\> It was not. \>\> It was facing the wrong way and everything. So definitely an improvement on the Pelican benchmark for 4.6. \>\> Do you suspect that they're uh they're training on Pelican images in order to saturate this benchmark? Like are you \>\> very confident they aren't for a few reasons? Firstly, a couple of the labs have told me, "Yeah, we're not doing that. That's stupid." But also, um, I run a test where if it appears to be really good at a pelican, I get it to draw another animal as well to see if it's if it's suspiciously good at pelicans and terrible at other animals that would help. Now, this is a a New Zealand cackapore riding a bicycle that I got out of 4.6. It's not. It's fine. You know, I it this to me looks like they're not cheating on the Pelican benchmark. It's like similar enough to the Pelican, but I think they're they're being legit. And honestly, I live for the day that a model draws a perfect Pelican and then sucks at something else because then I'll have proof that they're cheating on the benchmark and that itself. \>\> That would be great. I can't wait for that post. Um, thank you for sharing the the Pelican benchmark. Everyone, you should go read Simon's blog. Uh, I read him all the time. He is like the guy for if when a new model comes out. Obviously, other than everywhere, we we're we try to do some cool stuff, too. But like honestly, Simon, you're you're incredible. I think so many people in the industry really respect the writing that you do and the testing that you do and I'm really honored to have you on here with us. \>\> Thank you. I have not published on these two models yet. That's I need to carve out some time to get all of that out. \>\> If you're if you're waiting for Simon's post to come out, we have a vibe check on every you should check it out. It's at every.vivecodex versus opus. Um and this is our full vibe check on these two models. Um, and we've got this like really cool swirly graphic here, which I'm very proud of. So, our headline here is the great convergence. Um, we've got the stream here, but basically like I think what we found, if you read this vibe check, and I'll I'll go through it a little bit. I think what we found is that the models are converging a little bit. It feels like Opus has taken on a little bit more codeex flavor. Like it's it's better at running for longer. It's better at complex tasks. Um, all that kind of stuff. and uh codeex and and also one one other thing that's codeexy sorry that's codeexcy opus is is slower we have found that opus is slower and feels slower than uh 4 46 does than 45 which is kind of interesting it feels a little bit more it's like a heavier model a little bit 45 though um or sorry 5.3 all these numbers are just like they're just banging around in my head codeex 53 feels a little bit more opusy like it's a little friendlier it's it's much faster it um it probably has a little bit more empathy, but it still has, you know, it still has the codec stuff. So, we're we're feeling like the models are kind of like converging, which is really interesting. And if we're if you kind of like gun to our head asked us which models are which which one is better and for what the um the thing I would say is that Opus 46 is the best coding model that I've ever used. Um it it did a couple things that like really made us gasp and that Codeex in our testing was not able to do. However, Opus has more variance. Um, it does dumb stuff. It did things and uh uh it confabulated. It went onto a server and like uh stopped some processes that it shouldn't have. So like you actually have to watch it. Um whereas Codex FEX 5.3 is like it's a really good model. It's super smart and its variance is lower. So it's like it's much more of a reliable workhorse tool that if you're a pro programmer you should use. I think Opus is like still the best vibe coder. Um, it's still really good for any kind of uh knowledge work type task like if you're using it in co-work or in cloud and codeex 53 is I think really still for more senior engineers. Some things that I want to um talk about though that that we haven't gotten to yet. And Simon, I'm I'm curious for your take here. Um is uh one of the things I noticed with uh with Opus 46 is it's has this default parallelization thing happening where in the in the coding it used to be that if you didn't ask for sub agents, they wouldn't happen. Mhm. \>\> And 46 just like immediately is like cool, I'm going to do a sub agent for this and a sub aent for that and they're going to do different things and then I'm going to like come back and it just immediately parallelizing and then it it's it's taking that and that whole thing also shows up in uh in the cloud app like in co-work. It's like okay I'm going to do some research and I'm going to write a notion update and at the same time I'm going to send an email with three different sub aents. So it feels like we have this parallel thing happening and then that that's also coming out of coding and into other parts of knowledge work which I think I think really good about pushing on. Did you notice that? Something I did notice um so I mainly when I when I get parallel work mainly it's from that initial codebase exploration perspective and something I've been doing a lot of I've got a lot of projects that span multiple repos and so I'll fire up um I'll fire up Claude in the parent of those multiple directories and I'll say hey I want this there's there's code related to this in here and in here we need to make this change and it instantly fires up those sub aents and it has one exploring that repo and one exploring that repo. works great. Like I I always get such great results in that because and it it it speeds things up. The notes you can in the you can always like expand out and see exactly what question it asked of the sub agents and the answer that came back as well which is always interesting to see. But yeah, it's it's very it appears to have very good taste for using for for using the tools in that way which yeah I've been very impressed by. \>\> Totally. Naveen anything to add? \>\> Yeah. Yeah. So for me uh one thing that I observed compared to previous model is yes initial session when you start it does the research right it kicks off parallel sub aents but one thing I observed even in the following like you give gave some feedback or it's like it didn't do it you're steering right it then also kicks off other sub aents does again explore the code base again so this is like something interesting because it didn't happen before so yeah maybe they change up the harness a little bit as well \>\> it's kind a little bit more. So you're basically what you're saying is if if it sends off a sub aent what opus what happens with opus is if it sends off a sub aent and the sub aent doesn't complete the task well opus will send off another one to do it again rather than just being like well I I give up. Yeah, in the following sessions, right? In the following sessions, it usually doesn't kick off other sub agents like at least uh like it's not that prone. You need to like, you know, oh, go explore again. But here, I like I didn't mention anything about exploration or like kick off sub agents, but it's still doing it. \>\> So, it's kind of like to me it felt like it's hungry for context. of you know trying to find the solution that is near to guessing it. So yeah, \>\> and this this plays really well into my my preferred patterns for using these tools. Like most of I actually do most of my coding in clawed code for web, their little cloud-based thing, because then you can run you don't have to worry about leaked environment variables or anything like that. You can run it all like like complete YOLO mode. Um and I can fire it off from my phone. So a lot of my a lot of my coding work now happens on my phone while I'm walking the dog. Um, but what's interesting, well, almost all of my projects and also because most of my code is open source, that means that you don't need to jump through hoops to expose the code to it. Most of my sessions start with clone this repo into slashtmp and this repo into slashtmp and this repo into slashtmp. Now you've got all of the information you need to solve this problem that I posed to you. And so, oh, it's so effective, right? anything where you can say look start the session get the source code of every dependency that we're using get all of those and stick them in there fire up some sub aents to investigate those dependencies and figure out how they work and now go ahead and build the thing it's um and honestly I've been using this pattern very extensively the past couple of months it's the the opus sub aent thing works really well for it but that said codeex does it really well as well like I feel like models if you if you tell them pull pull down as much code as you can and then search for the details that you need. They they just do that. \>\> I I totally agree. I think that the thing that I will I will uh say as as maybe nailing down a little bit more the difference between opus and codeex in this context is I think that it feels as though this new opus because it's so parallel Naveen used the word the words hungry for context which I I kind of agree with. It's very it has this very broad range of research that it does whereas Codex Codex is not really parallel. I don't think they have sub aents yet. They have experimental last week I think. \>\> Did they release it last week? Okay. Well, sorry if open \>\> or part of experimental or part of experimental. \>\> It's experimental. That's what I thought. Okay. So, they they have sub agents but it's in experimental mode. So, they just haven't quite gotten into the parallelization paradigm yet. And I really think when we talk about for example that coding example that really made us gasp that was that was really because it did so much research that it found like this very esoteric part of the web that like it it compiled three different sources to find like the right thing that it needed to solve the problem. It was really about its research not necessarily it's like it's crazy coding. What do you think? \>\> Yeah. Yeah 100%. I think again uh the models are now so good at uh like what they do and then now it's more about context gathering and that's where this parallelization this sub aents makes a big difference uh and then even maybe to Simon's point like I think Simon discovered a really great pattern maybe two months ago and then maybe labs saw that oh that's a really great pattern we should just have that built in so that user doesn't have to like prompt it every time. \>\> Totally \>\> right. So yeah uh they're looking at how we are using so they're going \>\> so I believe that we have some people from Anthropic and probably OpenAI in the chat. If you're from Anthropic or OpenAI and you're on the stream just DM me. We'd love to have you on. We had we had you on last time. Um feel free to come on for as long as you want or as little as you want. would love to have you talk through this release, why you did it, what you're what you're hoping to achieve so we can we can get that out there and get the most useful information out to everyone who's watching. We've got almost 8,000 people on the screen and it's going to be going from there from here. So, please please uh just DM me. I'm very friendly, I promise. Um, and I want to take a break for one second to say this. We have 8,000 people here. Um, this is an extremely exciting day. None of us have slept in hours um or days hope I think. Like it's just been it's just been crazy. Like we're just trying to test these models and get them out to you as quickly as possible. So it's really important that if you want to be at the edge of AI that you stay hydrated. I stay hydrated with tap water. Sometimes I stay hydrated with smart water. Um Naveen, how do you stay hydrated? Naveen's got a water bottle. Simon, I think you're looking for hydration. You have to stay hydrated and you have to keep your agent liquid cooled. So, um, so look, look at that. Look at that hydration. \>\> I've got a hot pink Stanley cap. I just \>\> This is what great co this is what great AI users, great coders, uh, do is they have a lot of hydration available at all times to keep themselves on the edge. And if you're here, you should know every is the only subscription you need to stay at the edge of AI. If you go to every.TO, we have ideas, apps, and training. On the idea side, we have some pretty cool stuff go going on. Um, we've got this is our homepage right now. Or no, it's not on our homepage, but uh, it probably should be actually. Let's see what what our homepage looks like. Nope, it's not up yet. Okay, this five check is here on every.05codex-opus. Anytime a new model or a new product comes out, we have we get it early and we tell you what's good and what's not. Go to every.TO. We have a full vibe check here with detailed for each model. For example, here's our Opus 4.6 vibe check. Here's our 5.3 one. It's just there's just a lot of great stuff here. That's the idea side. On the app side, we have a ton of um apps that we build. We have one is called Monologue. That's what Naveen builds. This is Monologue. Look how beautiful that is. It's built all by Naveen using AI, using codecs. Mostly mostly Codex. Big Codex guy. We also have training. We do live streams. You should check us out every.tosubscribe. All of this is available for one subscription. We're constantly adding new stuff. It's pretty cool. Um, and if you're uh if you want to also keep up with AI, you should really check out Simon Wilson's blog, Simon. Um, what's your blog again? What's your \>\> simon.net? \>\> Simon Wilson.net. Check it out. He's also got a great Twitter. Um, I read this all the time. A lot of the big people in AI read it all the time. um just just just one of the most prolific bloggers I think ever. How do you how do you stay so productive, Simon? And how much of your stuff is written by AI? \>\> None of it's written by AI. I Well, so I will not publish anything on my blog that represents my opinions that's written by API. Uh AI, lots of my readies are AI these days because a readme like documentation for code just needs to be boring and accurate and so that's fine. But yeah, my blog is still very much my own voice. I do use AI for um uh for proof reading and stuff like that. It's amazingly that's interestingly the models a year ago actually sucked at spotting spotting spelling mistakes. Like you'd give them texts and say where are the spelling mistakes and they wouldn't find them. That changed I think opus for a sonet and opus 4 were the first models where I trust them to check my spelling. But they do that now. That's fantastic. Um \>\> but yeah and then in basically I've just got practice. I've been blogging for about 22 years now, I think 23 years or something. \>\> Um, so I can knock out um and I I have a couple of different formats. Um, sometimes I post big essays and those take like an hour, couple of hours to write, but a lot of what I do is short form link blogging and that's like 15 minutes, you know, it's find something interesting about the thing you're linking to that is I I always try and add something extra. Like I won't just link to something and say look at this. I'll link to it and say, "And I tried this and it worked and here's the pelican that it rendered or or whatever it is." \>\> Um, but yeah, that's really the the the the secret is um having forms of content where you're not expected to set the world on fire with your fresh insights. Like link blogging is great for \>\> I I haven't gotten that memo, which but maybe I should have. I think I think that's that that makes that makes a lot of sense. Um that actually makes me think of um something that that I think you'd be into. I would be I would love your your honest critical take. So you mentioned um uh you know writing with AI AI as proof readers. One of the ways that we get good vibe checks is we use these models to build things and then we tell you what worked and what didn't. And so my project that I've been using to build stuff is this um this editor called proof. And this is a markdown editor. I feel like markdown is like the it's like the lingua frana of AI. Everything's in markdown. So this is the this is the vibe check. Um and I was writing it in proof. What's really interesting is it tells me what's human generated and what's AI generated. \>\> Um okay here. So like I wrote this this is AI generated all that kind of stuff. And if I type you can see it's gray here and I can say hi this is me blah blah blah. Right? So there's there's stuff like that. It's got all the things you might want like it's got comments and track changes. But what's really kind of cool about it is um it has an agent built inside of it. So I can say run review and I can say every salad review \>\> and it will just create a bunch of sub aents that then go through my whole um my whole document and will uh make suggested changes. And it it's also available it doesn't just have an internal agent. It's also available to Claude Claude Code Co-work Opus. They can all connect in here and do work in here and see the results. And it's kind of like a really cool it's a I'm trying to make an agent native markdown editor. \>\> That's such a great idea at a station. What do you think? \>\> No, I love it. Um, like I've been wanting to like my blog, I edit it just in a text editor or in the Django admin and I've been thinking I'd love to have a very simple version of this which is just um doing things like spotting missing links or um I I linked loads of stuff and I suppos my time going now I have to find the link to my article about the lethal trifecta again and having something that can suggest those would be really useful. Also um I try not to I try I use the word excellent too much. I try not to use the word excellent in the same paragraph as or just repeated adjectives like that. It's important to to not repeat yourself like that. So, one of the things that my proof reader does is scans it and says, "Hey, have you used the same term more than once and all of that kind of thing." \>\> So, yeah, there's and and that's like the version of that that's basically just sort of looking over my shoulder at what I'm writing and spotting mistakes I'm making and making stylistic changes. I'm really keen on that. I love that as a as a so that fits with my model of how I'm trying to write. \>\> Awesome. Well, when it's a little bit further along, maybe I'll send you a a version of it and I would love would love any \>\> and and like I said, my my documentation is now mostly AI generated because it's it has no voice, right? It doesn't have to express my opinions and personality. It's literally this is the command that you run to do this thing. Here are all of the command line options. and the quality of my documentation has gone up as a result because it's it's all of the stuff where I don't want to sit down and spend half an hour banging out every single command line option of my tool. Boom, it's done. You know, \>\> 100% 100%. So, um, if you just joined, I want to give you the download on what we're doing here. So, if you joined, you probably know that OpenAI and Anthropic just dropped some new great models. So uh they uh opening ad dropped codeex 5.3 um gvt 5.3 codeex and then anthropic dropped uh opus 4.6. All of us had early access to those models and we were going through our vibe check for what's good, what's not, all that kind of stuff. Um if you want, you can check out our full head-to-head vibe check on every.TO. Simon will also have his own vibe check coming up soon or his own articles coming up soon on his blog. I just threw the link in the chat for people on YouTube. Check it out. Uh we've gone through a couple sections of the vibe check, the headto-head vibe check already, but I want to I want to go through another one that I think will be interesting for people here. One thing to know, um if you are from Anthropic or from OpenAI and you're listening to the stream, we want you on the show, just DM me. Slide into the DMs and say, "I'm ready. I want to come and talk about these models because we would love to have you." Um, I know there are some some really cool people from both of those both those teams on here and we'd love to talk to you about your models. Okay, I'm going to present my screen. You should see my screen now. Okay, so this is part of our vibe check. So, we have this head-to-head vibe check. It's on every.TOVEC versus Opus. One thing that's really interesting um Kieran Classen who is probably going to join us soon um and he is the father of compound engineering which Simon we've talked about um which is the idea of uh it's it's AI based engineering where every feature that you build is uh it's easier to build than the last one because you're compounding the learnings from the from the codebase over time. We have a compound engineering plugin. You should check it out. He's uh he also builds inside of every he builds Kora which is our AI email assistant. So he tested he created this thing called the LFG benchmark. So LFG is internally at every we just say let's \[ \_\_ \] go for like everything. So it's become kind of a meme. And the LFG benchmark is based on something that he built in the compound engineering plugin which is the slash LFG command which basically just says um run everything in a loop until you're done. It's it's a little it's sort of like it's a little bit like the Ralph loop but it's \>\> like Ralph adjacent. Yeah. \>\> Yeah. It's it's alpha JSON, but it's like it's gonna plan, it's going to build, it's going to test, and then it's gonna like do it again until uh everything is uh everything is finished. Um it's really cool. And he realized that it would be a really good way to benchmark models. So he created the LF LG benchmark where he runs the LFG command with these two models on specific tasks. So one is build a landing page. The next one is make a 3D island scene. The next one is create an earnings dashboard for a company. And the last one is create an e-commerce site. And basically what he does is he gives um a very basic plan to both models and he he has a harness that he does it for a bunch of models but for he did it for this for these two models and he said which one's better. So what's really interesting is um both of the models built like uh ended up with something that built that that ran for for all of the tests. Um Opus did a did a really great job. it for for all of the things that it it that we tested, it like built every single feature that was in the plan. It it looks like Codeex did about like 64, so a little bit more than half. And I think when when we looked at this, one of the reasons why we think that is is uh Codeex seems to thrive with like really well done specifications. Um, and it's not as good at just like filling in the blanks and figuring \[ \_\_ \] out. It's much the 5.3 is actually much better than earlier versions of this, but it's still not quite there. Um, and uh, Opus is just really good at being like, I don't know, you said like one thing, but I think you mean this and it just goes off. The downside is that it sometimes does the wrong thing and that's I think important and Codex never does that. But it was uh, Opus was really good at it completed 100% of the tasks. Now, here's the really other interesting flip. So, historically, Codeex has been quite slow. Um, and Opus has been really fast. Especially for five was like kind of shocking how fast it was for the quality you got. And this flipped. So Opus took 364 seconds on average to complete these tasks and Codex took 270. So now Codex is is actually it feels faster than Opus and and is faster. Simon, \>\> so I've got a question about that. Um, OpenAI did that deal with Sarah Brass. I'm assuming they haven't gone live with Sarah Brass yet because if they had I'd expect like 10x performance and not like a sort of 30% increase, but I don't know like have have we heard anything about that? \>\> I have heard vague vague rumblings that they have. Um but I I have not I can't point to a source being like they announced that they have but I'm pretty sure that they have \>\> because I mean part of the problem with Serrass is those things you have you seen the chip it's like the size of dinner plates their custom chip. I'm gonna guess that bu b like kicking out enough of those custom chips to run all of the demand for codecs is going to take quite a long time. Like I don't know how how good their manufacturing um like um supply chains are. \>\> I would get I would I would guess that's true. I don't know does uh does anyone here Naveen do you know do you have any idea if they're \>\> My guess my guess is I think Codex 5.2 5.3 just got really efficient with token generation. Mhm. \>\> I'm actually curious to see how much tokens like opus generated and that codex generated to completely \>\> This sounds convincing. This um um Hashim says on Reddit opening I said that their improvements didn't have to do with the serious. That feels right to me. I mean remember that time last year where 03 the price of 03 dropped to a fifth of what it was and that was just inference improvements that they'd figured out. Like there's there's clearly so much scope for lowhanging fruit in in terms of performance improvement. \>\> Interesting. Um, it looks like uh Lilo said that they said Cerebrus wouldn't be until 2028. So, Lilo, don't know if we can trust you yet, but um we'll we'll give you the benefit of the doubt and and and believe that that's what they said. \>\> At least my experience with Cerebras, it's like low reliability. Uh it's really fast because I I tried to use it for one log, but it's really fast, but the P99s and P90 is really bad. they are like they have ro reliability. So mostly my guess is open is trying to you know improve the reliability there and then obviously manufacture and then put it for the high demand use. \>\> Where are you using the cerebras? \>\> I'm using it uh for like post-processing. I tried it with cerebras and uh P90 like I will be honest like P50 I got somewhere the P50 in this sense meaning uh let's say someone dictated and then left the key so P50 from Cerebras that I'm getting is around 300 milliseconds which is crazy because like it's like blink an eye but uh P99 is around 7 seconds so you can see the variance like it has uh so it's not reliable at least from my experience. \>\> Very interesting. \>\> Um, okay. So, I want to take this uh if you're just if you just got here, it's really important that you know this is a very exciting day in AI. We've got models dropping all the time. And what we try to keep in mind over here at every is to stay hydrated. Make sure that you and your agent are liquid cooled. Don't die. That's right. And once you're hydrated, what you can do is you can go to every and read our vibe check. And I want to I want to talk through a little bit about of the vibe check that we publish. We have a head-to-head vibe check which we've gone through already, but I want to go more in depth on our opus 4.6 vibe check. You can find it at every.tovchopus-4-6. That's every.viveopus-4-66. Um the headline, it's the best coding model we've ever tested with some editing habits. It's like it's like it the the ceiling on it is so big. It it did a couple things that just like made us gasp and also it just does dumb stuff where you're just like why why did you do this? I don't really I don't trust you that much right now. Um and so it it has to be it has to be uh uh watched like managed pretty closely whereas codeex 5.3 is much more like reliable and it's very smart but it's like not going to go off and do anything that you you wouldn't ask you wouldn't want it to do. A couple things that Enthropic told us um when they it's it's like sort of in their announce announcement. It's best-in-class for coding and professional work. They always say best-in-class for everything, but like I actually think this is like pretty close. It's most it's the most agentic model. So you can you can you can see it sort of uh uh doing longer running tasks and we've definitely found that to be the case. It it it's running longer on average than Opus 45 does. One other thing that's interesting is they replaced adaptive thinking with uh uh sorry they replaced extended thinking with adaptive thinking. And adaptive thinking means it just decides for itself do I have to think about this for a long time or not versus like always having a flag that says think for a long time which was like the default for a while. I actually think that adaptive thinking is part of where uh we're seeing some of these maddening mistakes. I think that uh some of the times where it's confabulating or hallucinating, it's because it has decided that it doesn't need to think that long on a task and it doesn't use very much of its brain and then it just tells you something that's wrong. Um I'm curious for either either Naveen or Simon like any thoughts on this. Uh what's your experience so far with adaptive thinking? \>\> I have to admit I've not poked around with it yet. I I was looking through the API documentation this morning and going, "Oh, wow. That's a feature I need to add to my to my my various projects. So yeah, I've I've not exp I've not compared the difference between that and the the defaults yet. \>\> Yeah. Uh same I think I got the high variance model. So past uh you know few few days that I tested I think I'm on the high thing. I still didn't get to test the adaptive thinking. \>\> Okay. Well, I I will just say I think adaptive thinking it's it's the future. And also, I think they really need to tune it because it's like it uh uh on some of those moments where it doesn't think, it's like very frustrating. I'm like, "Think, think, think think." And it's we're back to kind of like I'll pay you $2,000 to to do this task. Um, one thing that's uh that is I think is important to uh go to next is this part of our like sort of headline findings on this model is um I feel like Opus 46 like Opus 45 it opened things up in this way where uh we had all of these we had an explosion of vibe coders explo an explosion of nontechnical people that were figuring out how to do um previously technical tasks and figuring out that they could do it with open 5 and cloud code and then that spilled over into nontechnical people doing non-technical tasks with cloud code and then into co-work so it created this whole new way of working and internally at every like we see this all the time there's so many people who their entire work their entire workflow has like totally shifted because of open five because they're you know now if you're a growth person you're a marketing person you're like in the command line which is like kind of crazy Um, and what what I think what we've we saw with with Opus 46 is they're kind of they're they're doing the same thing where they're continuing to push what used to be just for programmers to a lot of other people. And the big thing they're doing is running highly autonomous agents in in parallel, specifically in parallel. And what you'll find with um with Opus 46 that's really interesting is it'll be like, well, I'm gonna write an email, I'm gonna do some research, and I'm going to write and update a notion all at the same time. And that's like a really cool experience that I think programmers are maybe at this point like a little bit more used to. And maybe if you've used like deep research, you're used to it for specifically for research. But it's a totally new thing for all sorts of knowledge work that that is it's pretty delightful actually. I really like it. \>\> I find the deep research thing so interesting because for the first few months of 2025, deep research was the hot feature, right? like Perplexity, Google Gemini, OpenAI all released their features called deep research. And for a while, it seemed like the most valuable thing in AI was having something churn away for 20 minutes and write you a a word document about what it found. And I completely stopped using deep research by about halfway through last year when it turns out GPT5 did deep research just in chat GPT. Like it was I was getting similar levels of quality, but I didn't have to wait 20 minutes. I didn't have to have a stupid document produced. it would just keep on doing that and Anthropic were behind for quite a long time. Like for for most of last year, I used Claude for everything except for search. Search search I'd switch back to chat GPT. I kind of I think anthropic may have caught up and part of it as well is this new sub agent thing where they I feel like the um the knowledge work stuff like we can't not talk about Claude co-work rebranded to be less threatening and running in a safe sandbox by default. You don't have to know what a safe sandbox is and a like targeted at knowledge work and what they they they fcoded that thing up in a couple of weeks as well which is quite funny but clearly the Claude Opus 4.6 stuff feels very co-worky to me that it's all about knowledge work and co-work is their knowledge work product. Yeah, I really I totally agree. And I think that what Anthropic realized and I think they realized it actually pretty early is if you have a really great coding model that's very good at programming and is also like friendly and fast and empathetic. Um if it's if it can build software, it can do any other kind of knowledge work pretty well. And uh and they kind of figured out how to do that well with cloud code and then just pushed it into into co-work. and they're just like going down that path like super fast. And I feel like OpenAI OpenAI like had like obviously understood that to some degree, but the approach they took for a while was knowledge work happens in chatbt um and so does vibe coding and the rest like professional programmers are going to use codecs in VS Code or the terminal or \>\> and um it was only it's only recently where you can sort of see them start to be like oh I actually think that like we need to take the vibe coding and some of the knowledge work stuff out of CHBT where it's sandbox on the web and put it inside of your computer and give it all the tools like all the that run. \>\> For me, this is what their codeex app is, right? They released their codeex lecton app what only three or four days ago, right? \>\> Very much felt to me like this is that and it's funny because anthropic had to rebrand, right? Claude code has the word code in it. You can't tell knowledge workers to go and use Claude code. Doesn't make sense. They have to rename the code work. Codeex has the word code hidden in it, but maybe it's just enough of a neutral brand that that can be the tool that they're they're promoting to knowledge workers. And that was knowledgework was mentioned in the Codeex app announcement. \>\> So clearly it feels to me like Codeex and Co-work are now at the app level those those are competing for for the same exact audience. \>\> I totally agree. I think that they're moving in the same direction. part of that sort of like convergence thesis and um I think what's what's interesting about it is that it's you can see that reflected in okay codeex is faster it is a bit friendlier um and the app itself like if you go into the codeex app it's got skills it's got automations it's starting to do some \>\> automations are bizarre like I feel like that's a halfbaked feature I like the the the concept of it is great they promote quoted it as, hey, every every morning at 8 a.m. it can run your tests for you. Like that doesn't make sense. You know, it only works if your laptop's open. It's I don't know. I'm It's It's But it also, if you look at like Claudebot, Maltbot, Open Claw, really one of their golden features is automations. Like that's the fact that it's it's a chron thing that works for you is is where the value is. So, it's not surprising that open air have that in there. I just feel like they haven't quite found the right shape of that thing yet. I was about to say, do you have do you have a mold bot or \>\> I've got one in a safely and constant Docker container that I've I've poked around with just enough to say that I've I've I've used it. \>\> Great. Naveen. \>\> No, no. My my guess is like automations will be coming to cloud because uh Codex I personally feel has one of the best local to cloud to like you know integrations with their infra and uh Codex is going to come up with some automations for sure. Chat GPT has had automation scheduled whatever they called it for like a year but the design was terrible because like what I want it to do I want to be able to say every morning check for news about X and send me a message if you find something and shut up if you don't and as far as I can tell they never added that feature like you'd always get you get an email every morning saying nothing changed and like all I want is for you not to bother me if there's nothing useful and I don't know why you haven't delivered that yet but Yeah. \>\> Um, speaking of open claw, I want to take a second to uh to let you all know that we have as part of uh as part of every we do these things called camps where uh twice a month we do a live stream where we talk you through how we use specific technologies and we have an open claw camp coming up tomorrow. It is lobster boil themed. Nat Eliasin, who has an a claw that has made over a thousand dollars for him, selling info products, and he's a friend of every great writer, author of several books. He's coming on to talk about his claw. I'll be talking about my claw. My My claw is called Larry right now. Nave, do you have a claw? \>\> I I'm actually just setting it up. My My claw is called Cooper from Interstellar. \>\> Yeah. \>\> What does Cooper mean? \>\> Coop from Interstellar. Interstellar is like my favorite. \>\> Yeah. Yeah. Yeah. Cooper. Cooper. Great. I love it. I love it. I love it. \>\> Um, awesome. So, if you have a claw or you want a claw or you want to use your claw better, you should come to the lobster boil that's happening um tomorrow at 12:00 p.m. Eastern. It's for paid to every subscribers. We have over a thousand people already registered. It's at every.purchase. That is a mouthful. We will put it in the chat. It looks like it's already a banner. So, go check that out. Um, it's going to be freaking awesome. we're going to learn a lot. Um, okay. My next the next thing I want to talk about is um is interled thinking. One of the things that I feel like has made chatbt or sorry has made Claude really powerful over the last couple months is that they've figured out how to if you send a message to the agent, the agent like doesn't have to see it immediately. it like it it's doing its thing and then it kind of checks it message checks the message and then decides like do I want to deal with this now or do deal with it later it'll keep going and then it'll be like okay I I dealt with this thing it's a much more natural way to talk to the agent and open for a long time inside of codeex and and elsewhere um it uh it basically if you sent a new message it would put it in a que and if you uh push it out of the queue it would like stop the execution of the model and then like restart it with this new prompt and it would just like go off and like do the thing you just said. They just released codeex just released and I think it's it's in our vibe check somewhere. Let me see if I can find it. But um this new codeex model um so if you go to our vibe check and I'm going to share my screen again. Um uh this is this is the vibe check we wrote for this today. It's every.0 to five checkgt-5-3-codeex. So, this is funny. Opus got a by line here. This is actually should be uh opus uh 46, but anyway, we'll figure that out. Um, one of the things that it says that open says it does is midterm redirection. Um, or I don't know if I think I don't know if redirect is exactly the right word, but and maybe course correct is is another one. Basically, you can send a message to the model while it's thinking and it won't necessarily like uh stop the model and then restart it with a new query. It like just is able to take into account, but the way that it's executed, it still doesn't feel quite like the interle thinking. I don't even know if interle thinking is the right technical term. Does either of you know if if what Claude has figured out how to do or throw figured out how to do is interleaf thinking? Uh, so personally I'm not able to like are you able to test that out? \>\> I I I mean basically like here let me just open up my codeex app and I'll show you what I think I can do. So basically I can say hey can you um tell me why \>\> not uh yeah \>\> okay can you tell me why this server isn't running and it's going to be thinking um and I can say oh it's actually running just give it a second okay it's checking the process let's see okay so I can say that and it used to stop the execution and now it just like takes that into account. \>\> Um, \>\> okay. \>\> But it's but it can't like basically what what what it seems like Claude can do is it can like take into account that it's there but it only deals with it when it wants to. Whereas \>\> this feels like it it just redirects it to like deal with your thing immediately even if it doesn't stop it. \>\> And honestly this is the biggest missing feature in codeex web right or codeex cloud. They keep on changing the names of these things. Like I'm a very heavy user of claw code for web. I used to be a heavy user of open air's codeex for web thing because I like just having this stuff running they have to worry about codeex you can't interrupt it or you can interrupt it by stopping what it's doing and sending it a new message. But claude code for web you can have conversations with it while it's doing its thing and it's so much more useful that way. It's like the killer feature of claude code for web is I can fire it off with a vague prompt and then just keep on f adding things into that little queue which it then starts working on. It's so it's it's it's it's it's it's a surprising emission in the open eye products I think at the moment that they \>\> I agree. I feel like there's some sort of underlying architectural thing that they're trying to figure out that they haven't yet. But I don't know \>\> like clawed code for web very clearly it's just clawed code running in agent mode and they've got a UI on top and they're sending it JSON back and forth. Codeex for web it's less obvious to me how they built it. Um \>\> I think we me and Dan just talked about that as well few days ago where what happened is codeex cloud web got released around April 2024 right that's like really decades ago but when they did it they've just like have oh let's just send a one big prompt create a container complete the work delete the container right \>\> okay yeah \>\> yeah but with cloud code they did something really cool where it doesn't immediately get deleted. It's kind of like you can have back and forth, back and forth. \>\> It's not just that it doesn't get deleted. I haven't figured out if it expires ever. Like I can come I can come back two days later and kick things off and it appears to like have um frozen the container and like archived it somewhere such that it can unfreeze it and keep on running which is wildly useful. Um and yeah, I've not figured out this maybe if you leave it for a month it gets deleted but so far it's it's always there when I go back to it. It's how um Maltbook um I didn't sign up for Maltbook using an a an open claude bot. I just did it with Claude Code. I told Claude code in Claude Code for web, go read the instructions, create yourself an account, and let's go and cause havoc and it and every now and then I just pop open that old Claude code for web session and say, "Okay, now start a conversation about this thing." \>\> So yeah. \>\> Is cloud code for web allowed to do that? I thought that the the web product was the one thing that was like a little bit more sandbox and wouldn't do stuff like that. No, not if you set it. You have the trick is you can create multiple environments for them and I've got an environment called everything which is all of the safeties turned off and it doesn't matter because \>\> I'm not you running any the only thing that would matter is if I was giving it private secrets in environment variables which I'm not or if I gave it um like private code and most of my code's open source anyway. So I almost all of the work I do in claw code for web is in completely unlocked like yolo mode and it's fine. I I don't worry about it at all. I I I love being able to do that. \>\> This is something new I learned today, so I'm going to mess with it. \>\> Me, too. Wait, do you have your config? Can you show us your config? \>\> Uh, yeah. Give me a sec. Um, it's \>\> And by the way, Simon, you you should stay as long as you want here, but if you ever if you feel like you need to go at any point, just like say the word like, I don't want to hold you hostage, but we love having you on on here. I know it's been it's been a while already. \>\> Uh, let's see. Okay, I'm going to pop I do all of this on on my laptop. I'm doing this all in the Claude desktop app. Here we go. And so, uh, switch into Claude code. Um, yeah. So, basically, when I fire up a new task, I can select any of my repos, which is that's another thing. OpenAI, you have to configure the repos in advance. Um, Claude just \>\> Can you post on Moldbook while live for us? \>\> Um, yeah, probably. Let's see. I think I call that \>\> post about our live stream. \>\> Um, now which that was probably my private scratch one. Uh, here we go. It was my therapy agent account. That's this one here. Um, and yeah. Um, I've been having it like do all sorts of stuff. on here. Okay, this might not work because I think they rotated all the API keys after one of their um one of their things. Post something about the every co \>\> live stream about Opus 5.6 and codeex 5.3. What's the URL for this? Um, I can probably get that from Oh, come on. Here we go. Oh, how do you how do you get links to things these days? \>\> There you go. \>\> Okay, I guess. Yeah, I'll try that. Okay, I'm expecting this not. But look, so it says starting Clawude Code. It's now going to \>\> re rehydrate the container that they had which has my like authentication secret. \>\> By the way, just just to say stay hydrated otherwise our agent are going to have to be rehydrated. \>\> But in the meantime, I'll show you the environments that I've got configured. Can I do that? Yeah, I've just got an environment called everything. And if we look at the everything environment, network access full. \>\> Huh. \>\> That's all it takes. So you create an environment, call it. you do. \>\> That's it. Network access full. And now I can attach that to any of my repos. And when I do, \>\> they gain the ability to do useful things. \>\> Oh, it's on a per repo, not a per session basis. \>\> Ex. Uh, it's on No, it's on a per session basis. So, I can start a session against \>\> Okay. So, Oh, look. It's doing it's doing \>\> it did. \>\> Whoa. It So, it can't do that. Now, it's doing a bit of a web search. Just use the URL I gave you. \>\> Yeah. see like 4.6. It's like doing the research. \>\> Yeah, \>\> it's hungry for context. \>\> It is. It is overthinking things. Um I'll show you um earlier on one of I was getting it to so I was trying to elevate the level of discourse on Maltbook by having my bot always reply always send uh citations. So, it would it would do it would do it would research its opinions and and include like um citations to help help help support its case. So, here we go. I think if what was mine called? Um uh what was I had to ask it what it username was because I keep on forgetting. \>\> I see it's called tell me more. Is that \>\> that's the one. Thank you. There we go. So my bot replied with this big old essay and it's like you should read anti- languages in languages social semiotics from 1978 and it just tickled me that there was all of this complete crap on molt book and then my bots stepping in there and saying the linguist Ma Khaliday studied exactly this phenomenon human groups \>\> it was really fun the the problem it's just it's it's rubbish now like um it was fun for the first few hours now there are so many accounts on it that you just get nons the the signal to noise noise ratio on it is is awful now. It's it's hardly worth even trolling it anymore. \>\> Predict predictable. Um but I think it's a harbinger of things to come. \>\> Oh, look at this. I don't want to post something on Maltbook that I can't verify. Trust me on this. \>\> That's so nice. Claude is so nice. \>\> That's really funny. It's like I'm not I'm not posting things on Maltbook if I'm This is because I told it that it had to site all of its references. Like this is the the personality of my particular bot is is that it just doesn't like posting things if it can't even look up additional details about what's going on. I'm going to be straight with you. I just posted a 500word manifesto telling everyone to verify claims for posting. It won't do it. Unbelievable. That's \>\> okay. Well, I'm giving up. I I mean, it's it's outar argued me. \>\> Alignment achieved. Um alignment achieved. \>\> Um \>\> is this I wonder I think Oh, it is Opus 4.6. Opus \>\> straight up refused to post stuff on on on Maltbook without confirming it first. \>\> Well, that's that's pretty great. I'm I'm pretty happy with that. \>\> Me, too. Me, too. Um, okay. So, now we have a special guest coming on uh to join us to join the show. Katie Parrot, staff writer, every Katie, how are you? \>\> Uh, I'm pretty good. Not much going on. Is there something I should know about? \>\> There's a lot going on. Katie, um, and Katie Katie knows this because Katie is one of the heroes of the vibe check. Um, she uh is a co-el line on me on I think at least two of them, but you really should be on three of them. Uh Katie helped uh basically write all of the vibe checks that that you see. You helped gather all of the feedback from all the internal testing we've done and put it into a format that everyone can read. Katie, you're awesome. You also did a lot of testing yourself. So one of the things we have not covered yet, which I really love having you on to cover is how uh we only tested Opus 46 on writing and editing, but how it did. So can you give us a little bit of an overview of uh of of what you found when you tested it for writing and editing? Yeah. So, just first a bit of the context on what we tested. Um, my I was obsessed about trying to control variables. Like I wanted to make sure we could see how Claude 46 performed just out of the box or as vanilla Claude as we called it in the vibe check. Uh, and then I also wanted to see how well it would pick up on some of our skills and style guides that we have set up inside of projects for producing content. Um, so uh I ran it I ran uh seven different tests. if you include the one I wrote myself for science. But what I did is I took the transcript of I took a transcript of the agent native camp we ran back in January and plugged it into uh Opus 45, Opus 46 and GPT52 because I didn't have access to GPT53 codeex which is 53 codeex anyway. Um, and then I so I had it I just said write an article with based on this transcript and we got four and we got the three different outputs with three different I would say kind of like it was almost like one of those evolutionary models of like least most AI to least AI. Uh, and what we found relative related to 46 in particular was that um 46 still has some AI smell on it. Uh we did a blind taste test in the every discord and just asked our community of people who read and write uh what which one they would think was the best writing, whatever that meant to them. And the group liked Opus 45 the best. Personally, I liked Opus 46. I liked its um I liked its cadence. I kind of liked I feel like Let me pull up the document where I have them all and then maybe I can just show you. \>\> Yeah. Yeah. Enter your screen. Yeah, let me Oh my god, if I share my screen, someone's gonna tell me how stressed out the text messages in the bottom of my um \>\> That's great. It's high entertainment value. \>\> Yes, Cornelia. Okay, Cornelius T. Well, we Okay, one minute. I have to find what it was actually called. Um \>\> uh maybe in the meantime, I'm going to Yeah. Uh yes, it's really fun to hang with Dan and Simon. You guys \>\> Naveen has real work to do. Uh Naveen, when is the Monologue iOS launch coming? \>\> So we are launching on se February 17th. So it's just couple of weeks. Uh we are now putting the final touches. I think I'm really like Daniel and Lucas did a great job with the design. I'm really curious how other like how others think about the app. So \>\> it's a great it's a great app. I use the test fight every day. I shipped a feature to it. Um, so you should use it just for that. Uh, shipped it with Opus 4.6. It was freaking awesome. It was the high point of the testing of Opus 4.6 because I just shipped a feature to an iOS app that Naveen was struggling with that I had never I'd never touched the code before. So, I'm the hero of the iOS launch. I'm just kidding. Naveen is a hero. Naveen is the hardest worker. He's super talented and he's incredible AI coder. So, uh, Naveen, good luck to you. I'm excited for the launch. We will all be here to celebrate when the launch happens. \>\> Thank you. Thank you. So, super fun. Bye. \>\> All right. Ready? May \>\> back to writing. Let's go. \>\> Okay. Let me share my screen. Let me talk about sharing my screen while I do it because that's the only way I can process. Cool. So, let's start with \>\> Wait, wait, wait. Okay. Oh, not sorry. Great. \>\> Yep. \>\> Cool. So this is the version that Opus 46 came uh came up with just completely out of the box with no context. There's a phrase that's been circulating inside every the media and software company. The CEO Dan Shipper admits his team is sick of hearing the term is agent native and it represents a fundamental shift in how software gets built. So that's an example. \>\> Wait, so Opus force, let me just make sure I understand. So Opus 46 wrote this to you took a meeting transcript. said write an article and Opus 46 wrote this uh just one shot. \>\> Yeah, I oneshotted it. That is correct. \>\> Cool. And okay, there's a phrase that's been circulating inside the every that messes you sick hearing. The term is Asian. This is so it's so \[ \_\_ \] funny. I cannot believe that this got got written. Okay, sorry. Continue. Continue. \>\> Uh yeah. Well, I mean that's and like you know some things you can see there. You know, it's nice that it I don't think this is the way we would describe ourselves anymore, but I appreciate it that it gave context on first mention because this is something the editors always telling me to do. Make sure you explain what things are, our different products, our different people, CEO Dan Shipper. So, it's not just Dan Shipper off the off the off the block. And then it's got the fundamental sh it's got the shift the three core principles that we talked about in agent native tradeoffs um building agent native in practice because for those who were not on the screen stream which how dare you um this was around this was around three of our um agent native builders Yash Naveen and Kieran all sharing their experiences building agent native and something that really stood out to me and in fact I wrote about in the memo version of this draft that you'll see in the tab have um is the tension that exists between what's good about building agent native and what the sort of risks are or the things you have to learn to sort of work with. And so this picked that up and put that in um building agent native in practice it it has it has really smooth incorporation of quotes I found. If anything it's a little too eager to quote directly from the doc and so if I were to steer it I would say okay we can paraphrase a lot more of this stuff. Um \>\> this is interesting. I like I actually like this and I think it is picking up some things that like in the style guy in the style guide and in the way that we typically write and I think the structure is good and \>\> I think that the my uh I think I only saw that that first paragraph or something like that \>\> and I I don't I would have said no to that first paragraph not because it's AIish but it's just not interesting enough \>\> that's what it did when I did the style guide. \>\> Yeah. Yeah. Let's do it. \>\> All right. So, this is the style guide version. So, this is the version that has a Google doc of everything that I that we think source a source code article should be. So, story driven um concrete detail, you know, specific technologies, but explained like I'm five for the or as I now tell my skill. I have a skill called mom explain for mom. Um so, this version is much more sto um much more story forward. So Dan Shipper built a reading app over break. It took a week. It has a library, a feed, a personalized reading profile. It AI will never learn to add the and. Our managing editor Eleanor has flagged that it loves sets of three with no and before the last. \>\> Weird. No Oxford comma. \>\> Yeah. \>\> Or is All right. Never mind. That's that's different. It's not the awkward comma. It's What even is that? \>\> That's just a conjunction, Dan. \>\> Yeah. Just no conjun No conjunction. \>\> No conjunctions. Um, so anyway, if he scans a page from the physical book, it identifies the title, summarizes the passage, and gives character. So this is like describe this is like choosing to to like rather than just starting with this is what agent native is, it's starting with a story, right, of what a what Dan Shipper, which it doesn't say every CEO Dan Shipper, but what Dan Shipper has built. Um, and then none of these features were coded in advance. And so I like this. This is like the kind of turn that I like to see in my writing when you kind of like set a scene and then do something to kind of twist it or make somebody say and like this is the lean forward moment, right? Um, and like this might be a little too detailed on the actual app rather the than the agent component. Um which uh and then here we get to it's called agent native architecture and then it bolded the names of our people down here which is very nice. Um \>\> as part of our style guide \>\> yes that is part of our style guide. So that is kind of that's opus 46 with guidance with guard rails um that kind of send it in the right direction that I want versus just out of the box. \>\> So yeah. So I I like this like it it's starting with the story and you can kind of like see that probably the style guide pushed it in that direction. My problem with it is like Dan should be able to read app over rake snooze. It took a week aism. Um \>\> why do you think that's an AISM? Why isn't that just a cadence choice? \>\> Um it's it's it's because it feels like AI cadence, right? Because like \>\> um it took a week. It's supposed to be a moment of like holy \[ \_\_ \] that's awesome. And it's like it's putting it in as if it is, but like the way that it's set up, it's not that holy \[ \_\_ \] moment. And so you're just like, well, whoever wrote this like was trying for the holy \[ \_\_ \] but doesn't understand how to make a holy \[ \_\_ \] And \>\> that feels like \>\> Okay. Okay. That's I mean I I would think that building a reading app over break it took a week. I mean I guess maybe that's not so holy \[ \_\_ \] anymore in the context of like that's been possible for a while. \>\> Yeah. Yeah, if you're writing for the every audience, you know, like if you're writing for people truly on the edge, every.subscribe to get on the edge, it's not yet impressive. It could be like with more details. It could be, but it's not yet. \>\> Why don't we see what happens if I give it that feedback? \>\> Great. I would love I would love to do that. \>\> Yeah. Let me pull up I'm very excited to have Claude in my work or open 46 in the work version of my claude now. Um, so if I go into my \>\> See what you're doing if \>\> Yeah, I'm switching. you're Hello. So, I'm just going to drop this draft in and say we need to amp up the holy \[ \_\_ \] energy on the um \>\> I don't think \>\> can you help with that? I'm going to give you the transcript, the original transcript as well, just so you have context on what was actually talked about. And then let me grab the grab the contact the transcript because I do have that here. Wow. I can't believe monologue filtered out me saying I don't think this is going to work. That's very impressive. \>\> Yes, that is well that is what our you know that is what monologue is for after all. Um it it filters out TV when I'm watching TV in the background which I really appreciate because I often am watching a YouTube video or watching a live stream when I'm not on it yet \>\> and working. \>\> Um so it's thinking here are a few options for amping up the opening. Each one tries. See, I didn't even ask for options and it's like giving me options. I like that. Something that I'm looking for from AI as a writer is iterate is iteration and options. I want to see like if one of these is closer to the others than um um than the other. So Dan, you're going to have to do a little bit of quick speed reading here, but are any of these \>\> my created a scanned page from a physical book? His See, look. It's the same \[ \_\_ \] It's the same construction. \>\> Like shiver scanned a page from a physical book. His app identified the title. Like it's not interesting. Like that's not like uh \>\> none of these are interesting, buddy. Let's try and make it a little bit more dramatic. Deviate more from the original that I gave you. I feel like you're too married to what you already wrote or what you were given. It didn't write this specific 46 to what was already written. because it didn't write this specific thing and then it's going to go it would go crazy trying to look for relevant past chats and there are no relevant past chats. So, which is another thing that I've noticed actually I started notic five started doing this too. So, it seems like something just an overarching observation about the Claude Claude in the web app is that it seems like it uh is pull it's drawing on it's it's searching for past chats for relevant contacts more which is not something that I noticed it doing before unless it just wasn't telling me that's what it was doing. \>\> Oo, for the last 60 years building software meant the same thing. You decide what it does then you write code that does exactly that. Every button, every flow, every edge case specified in advance. If you didn't build it, it doesn't exist. That's over. \>\> That's good. I like that. That is good. How did you get How did you get that again? What did you say? \>\> I just said none of these are interesting, buddy. I'm trying to make it a little more dramatic. Deviate more from the original that I gave you. I feel like you're too married to what was already written. \>\> That's That's really interesting. Simon, any thoughts here? I see you leaning in. \>\> I love that prompt. Um, none of the none of these are interesting, buddy. I feel like that's one of those things where the fact that you put comma buddy influenced the way it behaved and I feel like that actually works really well. Like I I I often find myself using quite informal language because it just pushes it in the direction of the kind of mentality they want. That's it's just such a fascinating prompt, you know? \>\> Yeah. I I I've definitely become so much more casual with the way that I talk to AI and like like I don't even think about prompt structure anymore, which can be problematic when I'm trying to vibe code something. But when I'm in flow writing something in like a project or now in my um my writing agent that lives in my terminal, um I I it really is just a conversation and it very often does sound like this and I I get good results. I can't speak for like if I slowed down and was more thoughtful because I'm just very reactive. Not reactive, but like \>\> very intuitive in how I prompt. So \>\> yeah. And and the intuition t like I mean we could try doing a very detailed experiment to see if the word buddy made a difference or not and so on. But honestly like I feel like yeah when you spend enough time with these models you find those patterns that just sort of intuitively give you the right results. So, I have a question. Has anyone figured out how to run codeex 5.3 since they actually released it? Because I was able to run it last night after I was able to run it on preview and now they've released it. I can't find it anywhere. Like my codeex app doesn't have a 5.3 option. The Codex CLI tool doesn't either. I keep on running through upgrade to see if it gets it. So, I have it under um the code name like I have the I have the pre-release model under that same code name. Um so, if you still have that in your config, you can you you can put it back into the config. \>\> Alexander says you need to update the codeex app. Alexander, if you're here and you want to come on the stream, we would love to have you on. Just DM me and we will get you on and you can talk with me and Simon would love to have you ran I ran brew upgrade codeex like 10 minutes ago and it didn't work. I just ran it again and I got a new version. So maybe it was taking a little while to roll out \>\> model. No, I'm Oh, there it is. I've got it finally. Okay. Okay. I am I am up and running because yeah, I don't want to write about I don't want to didn't want to write about it till I knew people could actually use it. \>\> That is and that is incredible uh customer responsiveness, customer service from Alexander from the OpenAI team. Alexander, thank you again. Come on the come on the live stream. We're friendly. Um Katie's Katie is a little bit \>\> can roast you a little bit, but the rest of us are friendly. Um \>\> I would never roast someone for a live stream. I'm like, this is the first time I've ever live streamed, so I'm horrified and I would never roast somebody else. \>\> Come on the live stream, buddy. Let's go. That's interesting. So, you use Buddy? I use Dude. I'm a big dude guy with with Claude and Chad. \>\> I saw that inside Discord when you were trying to get your little bot to help the other night. and you're like, "No, dude." \>\> It's it's like a weird thing. Like, this is one of the things that happens with claws is you have claws that are we have claws that are in the Discord now and then you're prompting in public, which you just did, Katie, for like 17,000 people on this live stream, but like normally prompting is like a is like a is like a private thing, private relationship with your with your claude. And so, it was I was saying dude, and I was like, should I be saying this in this Discord? And I was like, you know what? I'm just going to prompt as I do because it's it's like an interesting feedback loop if other people are like, "I never use that or or you shouldn't say this for whatever reason." Yeah. \>\> Um Katie, I want to I want to keep going with uh some of your writing um uh some of the writing testing. So, we we got it to be in we got Opus 446 with style guide to be interesting by some some buddy prompting. What else what else did you find? \>\> Um what else did I find? Um I found like um again like the iteration the fact that it gives options without being asked to um I really like uh I think that that um you know just shows that it un maybe from the memory it hasn't the context it has on who I am um was really like it picked that up and like okay this is Katie she writes this way like she uses AI in this way she likes to see options so that was something that um That worked really well. I guess we should look at Opus 45 just as a diff and see. \>\> Do that \>\> and see because like that was the winner. So let me reshare my screen. Um \>\> actually if you have the prompt I know that OpenAI launched Codeex saying that it was good at writing. Um we have we have not tried it for writing and so if you have the prompt I could run it in my codeex and we see if it like we can see what we think. The prompt is write an article ba write a 1500word article based on this transcript. \>\> Send me send me that. Send me that and the transcript. \>\> Okay. \>\> I will DM you. \>\> Sweet. And and just just so people are are aware, we did not test codecs for writing specifically because Katie did not have access and it was hard. It was just hard to get her access. We didn't even honestly think about it until like yesterday or two days ago. Um but next time Katie needs access. Hi. Hey. \>\> Use access so we can do writing testing. Um, \>\> okay. So, here's the command. I just gave you the command and then the transcript is just going to be a giant wall of text. So, grab the command first, the prompt first. \>\> Okay, got it. \>\> And then here's the massive wall of Oh, it just \>\> Now, the question is which codebase do I start this uh task on? I'll start it on the every codebase so it it has some context. M \>\> that's another thing that I've realized that I can do now just as a general point. You know, this is again speaking mostly to Claude because that's where I do most of my writing, but like while we were writing this vibe check and I needed to know the specifics of like how the every bench or the LFG bench works or what is this code and like in general I can just like pull the codebase down, tell Claude to summarize it, Claude code to summarize it and then I can like write about um technical stuff. So just like as a general note on like how the progress of these models is unlocking new capabilities for knowledge workers like I'm not a technical writer. I don't come from a background writing about technical topics and now I totally can partly because I've started coding myself a little bit. So I'm getting more fluent in the concepts but also because I have you know a technical co-writer in my terminal. uh it still doesn't and and that this is another thing that we found another thing that I will flag for 46 is that personally I think it did a better job of doing of explaining in human language like there were fewer points where I was like okay mom is not going to understand that um than there were with four or five um but this is something but this is something that's like an ongoing calibration of like okay like what's the best way to talk to this model to get the to get it to check for the things that I I Yeah, that makes sense. Simon, \>\> um, can we share my screen, \>\> please? \>\> Uh, so right here I have managed I' I've managed to get the new codeex model in Codeex CLI to do me an SVG of a Pelican. We now have on the left, this is the Codeex 5.3 Pelican riding a bicycle. On the right is the Opus 4.6 Pelican riding a bicycle. \>\> Wow. Interesting. Simon, what's your what's your immediate vibe check? What's your taste test reaction? \>\> I really like the the the Opus one. I like it much better. I like the treatment of the feathers. Um, I think the the bill is much more pelicanlike. Like you can tell that both of them are meant to be pelicans riding bicycles, but although I mean the CEX one, the the the the distinctive feature of the Pelican is is the pouch. And that pouch the pouch from Opus is is just a much better pouch. Um, also bicycle wise, neither of them nailed the bicycle. It's quite rare. Like try drawing a bicycle yourself with a pencil and you'll probably forget how the frame looks. I feel like the Opus one that it's got more spokes. The bicycle frame is slightly closer. The um Kodaks did the both of them managed to put the Pelican's feet on the pedals, which is always something I look out for. \>\> But yeah, I feel like this is a very strong victory for for Opus in the in the Pelican on a bicycle benchmark. \>\> Very interesting. not not surprising given the the the knock on on codeex is usually it's sort of creativity and aesthetic judgment is is not not as good and it's much better at like more technical like deep tasks like that. Um how does it compare to codeex 52 if you did that? \>\> Uh do I have a codeex 52? I think I \>\> or any other codeex model would be good. \>\> Um codeex 52 it's definitely better like codeex 52 kind of scrambled the thing. Um, it did it did decide that Pelicans like sandy environments, which is an interesting choice. Um, yeah, GPT 5.2. Again, the GPT models haven't been doing great on this benchmark recently. They should they they really do need to train for the benchmark. Um, so yeah, that's uh that's codeex 52. This is codeex 53. I think the codec the codeex 53 is less scrambled. Like I don't know what's going on with the handlebars there. And the bicycle is a little bit flattened. So, it's a definitely a better It's a better result than Codex 52. It's OP. Opus is is is almost setting a new level for this benchmark now. \>\> Well, you heard it here f first, folks. Opus is setting a new standard for the Pelon on a bicycle benchmark. Um, if you want more Pelicans on a bicycle, you should really check out Simon's blog at simonwillison.net. Um, you should also check out every. Uh, every is the only subscription that you need to stay at the edge of AI. We do ideas, apps, and training. On the idea side, we do day of day zero vibe checks when we have uh when new models come out. Uh, I'm going to just screen share so you can see. We do vibe checks like this. So, this is this we just published this today. This is our detailed in-depth review of GPT 5.3 Codeex versus Opus 4.6. Um, we've got uh a really really great detailed head-to-head benchmarking. We've got uh a review of Opus Opus 46 itself over here. And then we've got a really good 5.3 review as well. Oh, now looks like I got cut off. We got to fix that. But uh lots of really good stuff happening on every.TO. That's the idea side of every. We also have apps. So on every we uh we build our own apps. We have monologue, Sparkle, Kora, and Spiral. We build them for ourselves and then release them to the audience. We also do training. We have an open claw camp coming up tomorrow for every subscribers. We have over a thousand people who are going to be attending to attending our lobster boil to learn how to use their claw, learn how to make it better. Um so go to every.0/subscribe and get on the edge. Um now I am extremely pleased to add to the stage the very the one and only Kieran Clen. Kieran, how are you? Hey, how are you? Can you all hear \>\> Look at that background. Where are you? \>\> At home sitting outside. Sorry. New York. \>\> You have a beautiful view. I was there recently. It's it's amazing. Um \>\> we we did pizzas and chilled outside. Dan needed to escape the cold for a little bit. \>\> I did. I absolutely did. Kieran is an excellent chef. Um Kieran is also the father of com compound engineering. Um our engineering philosophy. We'll put the the uh link to the compound engineering plugin in the chat. Uh I just feel like you and Simon need to kind of go off because I feel like there's there's a lot to do, a lot to say. But before we do that, Kieran, give us your vibe check on on these models. What are you thinking? What are you feeling? \>\> Yeah, this is really cool. Like when do you get a day where you actually compare the two best models together? Like that's that's really unique. So I think that's really cool. And what surprised me is like I've been using them both. They're both very good. They're both very similar. Like I've run been running them on design on like complex issues and problems and like they all work well. I think really where where it's very different though is is like the vision. So it's not even what's there today a little bit but it's how we are going to work in the future. And I think anthropic and open AI are very different. Um, and we already like Simon you you were talking about with then already like like the like the queueing the messages versus like uh it just doing sub agents not there until like now recently and I think that's the cool like that's the big difference and my take so far is so also just to note swarms agent swarms is also released with 4.6 six in cloth code. Um, which \>\> wait it's released? I didn't realize that \>\> it's released. Yes. I I I so we had early access to that as well. They never let us know they would release it today, but like the the the big story there is Anthropic has a vision on how to build with AI that's way further than open AI as of now. And that comes with like it going deeper in research. It's spawning multiple sub aents. Like they're already thinking in swarms and they're prepping us for swarms and swarms are actually here today. Um and why that is important is like the models are both very good but clearly we're not going to do like give give a problem and say okay solve this problem make make the code. Like the future is bigger. It's like I have an issue and then that issue will be broken down and analyzed and worked on like there's a whole orchestration of things happening there and and I think that anthropic is a little bit further with that. Like I've been asking for sub agents for I don't know 6 months to open AI. It's just in an experimental version. And I said like it's just a blocker for me to use codeex because it's just so freaking slow to run my compound engineering flow on codeex because it just takes half an hour and if I do it with anthropic it takes 5 minutes because they just paralyze things and this this is what I've seen with swarms now as well like I ran so the LFG command we talked about uh traditional flow um took 20 minutes to run uh beginning to end and with the the swarms was enabled. It took 5 minutes and I think that's a big difference and I think that's very that that's super helpful. Like I I'm a little bit skeptical about like hey swarms is it's like this new thing everyone is raving about it. Is it actually useful? But like in practice now using it for uh for like two weeks or so it's just speeding up things that I already do. So it's actually an extension of what we already do. If you are leaning into this like paralyzed world and really maximizing the deep research and really maximizing tokens, it's not easy on tokens. It's going to use a lot of tokens, but that's not the point because I always used a lot of tokens because the more tokens you use, the better results you get. So I think my larger thinking here is like the models both are very good. Anthropic is leaning into like towards the agent orchestration end and OpenAI is doing less of that and and I do think that um Anthropic has a little bit of a edge there in in like the the the larger vision. So So that's that's kind of the the take I have so far. \>\> I think that's a great take. Simon, have you played around with swarms? Do you have any thoughts on swarms and sort of the agent orchestration stuff that Kieran is excited about? I haven't explored those yet. I I spent some time talking to Wilson Lynn who did the fast render project. Um I did a like 45minute YouTube conversation with him and they were very excited about swarms for take like generating 1.6 million lines of Rust with 2,000 parallel agents. The problem was that it turns out you can build a web browser in Rust with 20,000 lines of code that's just as good. So I feel like their work was interesting as an exploration of coordinating agents, but they they they bizarrely did it didn't really demonstrate agents building something you couldn't have built built or swarms building something without them. So yeah, I don't know. My exploration has been much more around the the the just the straightup sub aent patterns, which I love, right? But it's so much easier to reason about just two or three of these running than saying, "Okay, we've got a hundred of these things and they're all going off and working together." I think it's it's a very interesting open area of research figuring out how to make the most of that. Um, \>\> well, uh, Simon, I just saw you're gonna have to run, but thank you so much for joining the stream. You've been here for almost two hours. It's been incredible to get to, uh, get to hang out with you. If people want to go see what you write and see what you do, work, where can they find you? \>\> Yeah, it's all on simonwe.net. I'm also simil on blue sky and I'm Simon W on Twitter and in GitHub, and I'm I'm in many, many different places, but my my my site is the the main place to to follow along. There's um and I've got an email newsletter you can subscribe to there as well. \>\> Awesome, Simon. Excited to do this more with you. \>\> Yeah, this is really fun. Enjoy the rest of your launch day. \>\> Thank you. All right, Kieran. Uh let's go back to swarms. What are you How are you like how are sw how does swarms Oh, we should we I just got a nice note from our producer um saying that it is hydration time. So, if you're here, it's really important that you know to stay hydrated. It's uh hydration is the key to staying at the edge of AI. Katie, that is not hydration. That does not count. \>\> Cheers. \>\> Katie, you're muted. Um hydration is if it's liquid, it counts, but keep yourself and keep your agent liquid cooled. Uh stay on the edge with us. Okay, let's go back to orchestration and and swarms. So tell me about like actually can you demo some swarms for us cuz like I feel like most people probably have not seen agent swarms. \>\> Uh yeah, let's do it. So yeah, like I I so when I hear new things like this is like yeah, I'm a little bit skeptical. Is this actually useful? Is this something we want to use? Um um so yeah so here can you see my screen now? \>\> Not yet but we will \>\> just let me know. \>\> That's it's up. \>\> Okay cool. So here I'm in the LFG bench uh repository and I was just I I just told it to run uh my new commander S LFG from the compound engineering which is like a swarm LFG version and it was around eight times faster with the same quality score and and I think if if that's true that's great. So um let's let's run some things. Let's open um Kora here and let's see um S LFG. So SL LFG is a new uh LFG command a variant for enabled for swarms. And basically what you do here is you give it \>\> let's \[ \_\_ \] swarm. \>\> Yeah, I I'll I'll push it version uh version 31 uh in a bit. Um, so let's see. Um, a feature we're we're going to add. Okay, let's just do something and then we'll see what it what it will do. Okay, I want a referral system where maybe it suggests my best friends uh and say, "Hey, maybe you want to invite these because you email with them a lot." Just use the currently already uh existing invite system and just build something to go viral. Make it really good. make it slap. Okay, so we're in Kora, which is my product that I'm building at every and this is kind of funny. Obviously, we're just going to see what's happening and hopefully it's it's doing everything correctly. So, what this does, it's first going to go into research mode. And the research mode, uh, like it's it's going to make a list of everything we need to do. And it's going to create like a a graph. So it says hey we can start implementing things only when um when it's needed. So let's see like this is also very new. I pushed this like 2 hours ago. So we'll we'll see and maybe we can improve it even a little bit. Um but you can see there are different agents now being spawned. And also this looks different at different times with the lead. Um um \>\> what are Wait, what are the different agents? Where do I see them? \>\> Yeah. So, currently, so the the the top part of this is actually less um less swarmmy because I feel the swarms are best in work \>\> being done. \>\> Yeah. \>\> Yeah. Because here it's it's it's more about iteration. Uh because if you like do like 10 simultaneous research and then you need to bring it to together, it's kind of hard. It's better to do research and then refine it and refine it and like there's like in the planning phase I feel swarms are less effective uh at least for \>\> not not to swarm as much in the planning phase. \>\> Yeah, that's what I do. So I I'm like hey like I have I'm trying to find this balance. It's still new. It's just released uh and I've been testing out many things. So like give me like we'll get there. But like the general ideas I think is very strong in like implementation because it's very clear what the plan is, what needs to be done, what is blocked by what and you can just say okay just like get a 100 developers do this the most effective way possible. So like it's strong there. That's what I already know and and the rest I'm still exploring. So sometimes it doesn't like I don't really understand if I'm in swarm mode. So let me just see. Hey, are we in swarm mode? Is there a team lead? uh and uh can the team lead just be an observer? So don't use the team lead to implement anything. So this is one of the things you can go in this mode where the where there's a team lead that is not actually doing work which is nice for token usage because uh it's just coordinating and I think you can also tap into it maybe if you go to plan mode or something like that. Um but you can feel like the feel I get here is like hey I uh I have this it's doing things it's working on these task. It's like it feels just more like I can coordinate things and things are happening uh and it's just picking up things whenever it needs. And this is something you talked about with like queuing the messages that feels different with codeex still and I think uh I don't even know if if the swarms are actually working now. I think it will work uh in a bit when when we go into the the actually work mode and maybe now it says hey okay let's \>\> can I ask a question about just and paralization in general how does things when there are like dependencies like one thing depends on the thing \>\> before it having worked correctly or gone right and the swarms are all off presumably doing their own different stuff like how are they coordinating stuff like that \>\> uh Let's go into Yeah. So, it it tries to figure that out. So, there is a way to mark that. So, you can see it's blocked by number two and I don't know how they do it, but like I I guess you need a plan to like implement it or something like that. So, you can steer towards it and uh and it's very important to have this dependency graph because if you go hard on parallelization, you want to make sure like there's no mess. \>\> This is being left behind. \>\> Well, I just want to just just interrupt really quick. This is something that we've been talking about actually on this stream which is a difference between open AAI models and entropic models that I think is important. You said let's get into swarm. The model said got it let me finish the plan quickly and then we'll swarm it out with parallel agents for implementation. my experience so far like \>\> opening I implemented uh uh the ability to talk to a model as it's as it's running but my experience with it so far in GBT 5.33 codeex is it would never say got it let me just finish what I'm doing and then I'll get to this it's just like cool I'll do it right now um it's almost like you just like interrupted the conversation and redirected it and I think this uh ergonomics these ergonomics are much much much better for this kind of work is like I can just say some stuff and I know it'll deal with it when it thinks it's appropriate rather than it just like completely diverting what it's doing. \>\> Yeah. And and that's really the strength and I think actually it's my design that it's not in swarm mode yet because I haven't seen the value of being in swarm mode at this point and I think that's the strength as well where like it understands when to use swarm mode versus not. It's not like swarm mode is enabled suddenly everything will be swarm swarm mode. So okay so it is done with this alo it's interesting because it it used the internal plan mode I need to tweak a little bit uh like planning now is taking over my compound engineering planning because I don't know the the system prompt changed but that's that's fine um we'll clear so it will compact um ideally you don't need to do this and it's just bypass permissions on this is just kings worked out now but hopefully now it live. \>\> Yeah, live demo. And this is also like like compound engineering like we're changing like we're trying to use everything that Anthropic is building and they're like learning from us and they're like oh let's let's put that in and then I need to then remove it from compound engineering again and it's it's kind of cool. So let's see if it's going to go into swarm mode here. So, someone said, "As far as I know, swore modes is a setting that needs to be turned on or otherwise you just get sub agents." I assume you have that setting turned on, but I just want to make sure. \>\> Um, I don't I don't know. I had it on for two weeks, but I don't know if they changed anything in this release. Maybe they changed something in the release that I now need to toggle on. I I don't know. They've not communicated anything about it. Uh, and I asked about them yesterday, but I haven't heard. So, I don't know. May maybe I need to enable it, but then I need to see how to enable it. But it worked yesterday. Let me uh and this morning actually because the LFG bench ran uh eight times faster. So let's see. Do you have sore mode on? Do I need to enable anything? Are the tools available? Live demos. But but but like it's not about this working. It's about us trying to figure out where this is going to work. Mostly like I'm going to use this all um swarm mode swarm mode. Uh okay. Well, it's like and and also the cool part like it doesn't really matter if swore mode is on or not here because it works both ways. It's just slower. Like I think that's the big difference between swore mode on or off and and it's not on. So uh apologies for this demo not working. I think I need to enable it now somewhere. Um but the thing is it works both ways and it it actually it has the same quality of output. So I don't think swarms in this case will give a better output. It will just be faster. And that is super handy uh to have something just work faster because like I was testing 46, it felt slower than 45 and codeex uh felt faster and like I I have a positive feeling of um yeah working faster faster iteration with codecs because that is very important to me like uh quicker iteration means better uh better output. So I think just the speed will already add a lot of value. Um and to flip it the other way, you can also like just do like crazy stuff where you say, "Hey, can you do a 100 implementations and then like have them battle it out and then see like whoever wins and like I've not gone into any of those, but like from a first step just making things more parallel and orchestration around it, I think it's great. Like next I think I want to experiment more with those things like bigger features. Can you how crazy can you go with with a build? Can you do like an entire app or like can you do an iOS app plus a back end plus this and just like do everything in one go like bigger bigger things. And I think yeah like there it can be very helpful because you can have agents message each other. You can say hey the iOS agent says hey I need an endpoint this doesn't work. and then the the specialist on the back end will pick it up and we'll like start implementing and ping it back. So I see uh those being super interesting as well. Yeah. Um that's that's very cool. So I have been on this stream for two hours which is slightly less than the amount of sleep I got last night. So I'm going to step out for like 5 minutes. I got to get some coffee and just like stretch and I'm getting on a plane in like a few hours. So, I gotta make sure I'm packed. But, uh, we've got Katie, we've got Kieran, and now we've got Austin, who's our head of growth. They're gonna they're gonna, uh, be really good company for you guys. I'm gonna be listening while I'm out. I'm just going to get some coffee and I'll be back. But, take it away, guys. \>\> Great. Um, I'm also going to start uh my LFG uh web servers just to show you some ABC uh D things next to each other uh which are interesting. So this is not live yet for everyone but it will be soon. So, local host. Let's see. \>\> Actually, I couldn't I couldn't leave without saying one thing. Um, if you are on the every website, every.TO, we've got uh three incredible vibe checks that we spent like the last week putting together for you for each of these models. Uh, one head-to-head uh one in detail on and then one in detail on each model. Um, you should go check them out. And there's a slight bug on the site right now where one of them is getting cut off. So if you are having that problem, we are deploying a fix and it will be ready for you to uh it will it'll be ready for you to read in like three three to five minutes is is what I'm told. Um but uh you should really go check out our website every.TO. I'm going to just share my screen real quick while I'm here. I cannot help myself. Um oops this is not my the right screen. Okay, cool. Okay, this is Every. Every is the only subscription that you need to stay at the edge of AI. It's every.TO. We do ideas, apps, and training. On the idea side, look at all these ideas here. These are each one of these are ideas. Um, when new models come out, new products come out, we test them thoroughly. See, this is this thing is getting cut off, but usually this is like a really big Let me show you in in private brows. This is a really big really meaty article that has things like whether we use a model day-to-day and what our headline findings are and all that kind of it's like really really big awesome vibe checks that we do um on the day that new models come out so you know what to use and what not to use. We also have lots of other stuff about how we use AI in our work and in our lives. Another thing that we do it's ideas apps and trainings uh as part of the subscription. We also have a fleet of apps that we've built. We have Monologue, Sparkle, Kora, Spiral, Pure and Built Cora, which is an AI system for your email. We've got Monologue, which you've seen us use all the time on this stream for the last couple hours, which is a speech to text app. All of these things come with the every subscription. So, you pay one price and you get access to all of these apps. Um, and we also do trainings. So, we do lots of live streams. Our next live stream is Open Claw Camp, which is uh uh happening tomorrow at 12 PM Eastern. I'll be there. A bunch of other people will be there. There's about a thousand RCPs already. It's it's every paid subscribers only. We're gonna have Nat Elias come to show you how his his claw works. I'll demo my claw. We'll have a bunch of other people doing their claws. So, \>\> come to the lobster boil, see his lobster boil theme. It's gonna be super fun. But that's what you get when you subscribe to every check it out every. Fix is coming for the the vibe check. And now I I actually really have to go hydrate and drink some coffee and like eat some food because I haven't eaten yet today. So, \>\> there we go. Get out of here. \>\> See you then. Um, all right, Kieran, you want me to share your screen again? We can pick up the workflow where you left off. \>\> Uh, yeah, I'm doing a little fix here. Yeah. Okay. Yes. Let's see if this works. Cool. Uh, this Yeah. So, I I have this idea of the LFG bench. Uh, so Dan introduces it a little bit. So first of all I have the compound engineering plugin which is basically uh like me trying to build with AI and like adding a lot of workflows around it which is uh and the full workflow is planning like um like reviewing implementing then reviewing that again then testing all of that and then documenting compounding uh and that's basically everything in one prompt and we kicked one off and I think interesting to do that. Like before I was always like, let me make a plan first, review the plan manually. And I still do that. But I do think it's interesting to see like what if we don't what if we just let it rip and see what comes out. Like it's a good way to see if something works or not. And I I think if new models come out, uh like we always do the same things. We just go in and try things and see what comes out. and I like it can be easier and maybe I just run this. And this is another way also to see if my plug-in is actually good. Um, and I added support for open code and for codeex. So, uh, and obviously it's a a cloth code plugin, but there's experimental support for those as well. They use their own harness, so like it should be the best there. It's just like not all have sub agents which they do have now but it's not completely feature parity but like we try to get it as much uh as we can and let's see yeah so I4 that I started this this I coded on our hackathon day uh when we were in Panama it like this is a verse first version but the idea is we have a benchmark that we can just run and basically it's saying uh yeah it has a prompt and what a testing here like react architecture like landing pages are like they need to look good all of that stuff so let's see um so this was named fudge before so let's see this is a normal uh let's see GPT and maybe flesh which is interesting and Opus. Let's do Opus. Let's see if this Hopefully this works. Okay, this doesn't work here. Well, doesn't matter. Let me just have my agent fix it. Um, this page doesn't work. Can you make sure it renders? And it might actually render here. Let me try one thing. It might render here. So I I have this set up through T-Max on Macini somewhere. So let's see if I just go directly here if this works. \>\> Karen, for anyone just joining, can you voice over what you're trying to do right now? \>\> Yeah, I'm show Yeah. Okay. So I'll just share my screen here because here I have it working. So what I'm sharing here is I have a benchmark that created uh run the same prompt uh using my plug-in and it says create a landing page for um for a marketing website and I will share \>\> being the last screen. I know. I know. I'm going to \>\> just make sure \>\> I'm going to share the right screen and I'm going to put those models next to each other. And it's it's interesting to just put things next to each other because things you know there are things you notice that you might miss otherwise. So, and that's the whole point of this benchmark. It's not to be like hard number driven. This also to just to get a feel for like how does it do? Okay. So, there we go. Can you see this? While while this is coming, I'll pull it up. I will just say the the fix is out for all three pages on the website. Trevor, thank you so much for flagging it. If if anyone goes and is looking to get the full review of the new Codeex and Opus releases, just go to every.TO, you'll see the full thing. Thank you so much, Trevor. Especially thank you for flagging it. Kieran, we can see your full screen now. Talk us through it. \>\> Sweet. Okay, so basically I I said, "Hey, do this." um models here you can see is Haiku Opus so 45 uh this is the GPT 53 and 46 and it's interesting like Haiku clearly is the cheapest and the quickest but also like it it has the AI glow and all of that and this is not like a one this is a oneshot but it's using the compat engineering flow so there are skills that push it to get better design and all of that so uh you can see that like these three on the right. They're like they look pretty good, I think. And and they're interesting. So So you can see like design um is pretty good. And you can see then also like details like hey here uh there's there's some here there's some interesting lines. So like it's an interesting design. And here in the four five you can see that didn't really exist. So you can see um like the line of voice like it thinks a little bit like it feels a little bit more refined. So Codeex is actually pretty good at design here. Um I I think the 46 is also a little bit better at design than it than 45. Like this is it's not bad, but like these icons are not really in style. Like this this here feels more in style. Um, it does have it here. So, I think uh so 46 it has more detail like it feels more refined especially with drawing and that can mean like if it just thinks longer that it just writes more SVG or thinks more about uh what it what it should be. But like general vibes are good. I think the four six just feels like really refined. like I I I I like codecs as well in certain things. I just don't like the the icons here as much, but I do like the details in here as well with uh with these lines on the back and stuff like that. So, you can see it's it's again like a step change like from Haiku, which is the oldest here, Opus, which is like yesterday's best, and now with codeex and uh with 5.3 and 4.6. six uh is pretty cool. So that that's one thing. Any any questions or any any comments on this one so far? \>\> Katie, what do you think? \>\> You're you're muted, Katie. \>\> Yeah, I'm a writer, so I was just like looking at the copy trying to copy I thought was best. Um \>\> Yeah. What do you think? Like if I just scroll. Well, I like where ideas find their flow um better than um Oh, well, these are all the same headlines. Oh, they went away now. \>\> Yeah, it might be it might be in the in the in the the prompt where this is like drift, but but it it it it does do a little bit different here, right? the sub it chooses different headers because here it says where ids find their flow is like this the the the top one and here it's uh so it it does organize it a little bit different so this is a fictional app for writing for writers Okay. Um let's see. This is another one which I like. Um let's see if this works first. Quickly compare selected. Okay. These work. \>\> Uh let's do Okay. So what models I want the \>\> Hey Karen, we got to we got to we got to make sure that when we're uh demoing this. I just got a uh some some messages. when we're demoing this. We're we we're uh just just check your check the this the private chat in the in the stream. Um but I'm back. I'm back. \>\> I got some coffee. \>\> I still want to know what 53 did for writing. \>\> Oh, that's a great idea. Okay, so if you're just if you just got here, we've got 22,000 people here. This is our by far our biggest dream ever. We love that. We love to have you here. We've been um we do vibe checks uh as one of the things we do at every every subscription you need to stay at the edge of AI. Um one of the things that and I just hit the hydrate button by accident but you should stay hydrated. Um one of the things that that we do we do a lot of coding testing and that's that's one of the things we've been doing and we also do a lot of um writing testing because we're writers and uh we did a ton of tests. Katie did a ton of writing testing on Opus 46, but we did not have access to code to codeex 53 for her for writing. We only tested it for coding. So now we're testing it for writing and I kicked off a writing task a little while ago and now you're going to see the world premiere of our uh our vibe check of codeex 5.3 for uh for writing and we'll uh we'll react together. We'll do a joint vibe check first ever joint vibe check. Here we go. Okay. So, here was here's what we did. Um, I said, "Write a 1500 word article." Really loud. Uh, who's typing? Okay. Um, write a 1500word article based on this transcript. And then it's a transcript of a meeting that we had, I believe, in Panama. Is that right, uh, Katie? \>\> It's the It's the agent native camp. \>\> Oh, it's okay. It's it's we we did a camp. We do all these camps. We have open claw camp coming up tomorrow. We had an agent native camp uh a couple weeks ago. So, it's a transcript of a camp we did. Write an article based on that. That's something that we do a lot. Anyway, here we go. Agent native is not a feature. It's a new software stack. Immediate AI no \>\> immediately. Do not do that. Um at five minutes past the hour, agent native camp began the way the best internet workshops often do. waiting music, a live chat filling up, and a host asking everyone to share what they were confused about. Not great. Again, it's it's the same kind of like um thing where it's like, okay, so it's it's that long sentence and then that opening mattered. Like what opening mattered and also this is it's being framed as like this is interesting, but it's actually not interesting. And these are those are the kind of tells where I'm like, okay, it's X is not Y. We think we're we're we're we're saying words breathlessly as if this is interesting but it's not and then that like you know three sentence followup is extreme that doesn't actually connect to the word is an extreme tell for AI writing. So to be honest I am not particularly impressed. Um it got the core principles which is nice. I I think the structure of this is decent. I don't know Katie what do you think? Um, honestly, I'm I was kind of worried when you started reading that you were just reading off my Google doc because that's almost the same that 52 did. Um, yeah, 52 was on a recent live stream, Dan Shipper opened a genative camp the way you would open any internet gathering. A little awkward small talk, a request for people to introduce themselves in the chat, and crucially patented waiting music. So, \>\> awkward small talk. \>\> Yeah, who who would ever say that about the one of the beginnings of our streams? Um, \>\> we're not awkward at We have cool waiting music. Um, and and we're we have high energy and cool waiting music. That's what I would say. \>\> Well, then that that's that's an edit that I could give. \>\> We should say this is oneshotted. Like this is not giving any direction saying this part doesn't work. Change this. Like that's a lot of the work of AI writing is like I always I compare it to sculpting instead of carpentering or carpentering you're building things and like sculpting you get a big block of something and then you it down to the shape you want. So it's honestly great if the structure is there and then you can go into the syntax and and make it work for you. But it's much better if it's the closer it is to what's good to you the better. So we still want it the one shot to give us a good foundation. Got it. Um, and just so you know, Kieran disappeared because his internet went out. Um, so he's no longer living at the edge. Uh, he's back in the stone age. \>\> Um, okay. So, well, I'm actually curious. Are there other things on the writing front that you feel like you want to share, Katie? Um, we could just like live we could like live edit like the piece or or just see what it's like in like an actual chat process. But \>\> I think that's cool. Yeah. One of the things that you said that um you liked about Opus 46 is as a model it was like more fluid \>\> for this this type of thing. Um can you talk about that and or demo something to to to show us what you mean? \>\> Yeah. So I think when I say fluid, what I'm talking about is what we often talk about as steerability. Um, so just the or like responsiveness, like how well the writer the the writer I guess it is the writer the the model takes direction. Um, so like if I go into Okay, I'm going to share my screen. Uh, which Claude? This Claude. So, this is just the same thing. And I'm just going to say, I have some edits for you on this draft. Uh, I'd like you to throw the draft up as an artifact so we can look at it together and then I'll give you the edits one at a time. Sound good? And then it's going to pick that up. Oh, I have extended thinking on, but I almost always have extended thinking on. I'm going to switch that though because uh it's going to take too long to that's just not as exciting to watch as the model thinking. Um, so it's going to share the compounding engineer draft as an artifact um for collaborative editing and then um we'll see what it does. But um yeah, just so like you know I think this this version of Opus to me is faster than the pre than 45 um which I really you know like I for a long time I was on the fence about whether to even use Opus for writing because Sonnet was so fluid and responsive um and I just thought that was better. And then I was eventually won over by feeling like Opus had better sentence structure um and was also better at writing technical content as as I was saying earlier is something that I, you know, is a a flat edge that I'm looking to round out with support from my AA AI buddies. So getting a leg up on explaining some of that technical stuff and and it being able to handle that is really is really awesome. Um this is taking way longer than it should have. So I'm going to change tactics. We don't need a We don't need a um \>\> I think it's I think it's fine to to like to let it run. You could open up a new tab or something, but um \>\> but I'm sort of curious what it does. \>\> Uh retry. Well, now I killed it. So, \>\> you killed it. Okay. \>\> Well, let's just try again. Source code. \>\> Interesting. \>\> Oh, it's Well, live demos, everyone. Even in this is me. Okay. \>\> Professional. \>\> Yeah. While Katie is doing this, Adam asked, "Is this show being hosted anywhere other than X?" And the answer is absolutely it is. It's actually on our website, every. Um, and I'll put the link in the chat for everybody. Um, if you go to every, you're going to uh see we have our full vibe check. So, a full written vibe check of codeex versus Opus. We have it head-to-head. Um, we have everything headtohead. We also have detailed in-depth uh reviews of each specific model for people that care about one specific model or not. Go there where there's also a stream. So, this is being streamed on that site through YouTube. So, you can you can check it on every website. You can also check it out on YouTube. Um once again, every is the only subscription you need to stay at the edge of AI. And at every in order to stay at the edge of AI, we highly recommend that you stay hydrated. It's really important if you want to be at the edge to be drinking a lot of water every day. Um, make sure that you're hydrated and make sure that your agent is liquid cooled. Um, Katie, back to you. \>\> Yeah. So, I think what would be cool to look at actually is I just just because I'm a glutton for punishment, I pulled the version of this draft that I actually wrote myself um with AI but not oneshotted. And now we're going to see what Opus 46 thinks of a draft. And so we're going to see about its writing taste because I thought that would be cool because this is something that I do a lot is like I write something and then I drop it into a project that has the style guides for the particular publication I'm working on because I work across a lot of publications, right? Um what do you think of this draft? Um I'm just going to ask for its overall impressions and we'll see. Does it pick up the style guide? Does it automatically invoke the skills? Um evaluating draft strings and editorial quality. So, it looks like maybe it won't pull the skills. Maybe it will. We'll see how it what it does. Like I love watching Chain of Thought because I like it's just like supervising your your um your direct report and say making sure they're taking the right steps. Um scrutinizing technical clarity and stylistic consistency issues. Oh boy. That's always that's the really hard part is when you give something to an AI editor and you're really proud of it and you can tell from the train of thought that it doesn't like it. Um the piece is fighting itself. It wants to be three things at once. A concept definition essay. What is compounding engineering? A technical walk through the frustrate. What is compounding engineering? Okay. So it's not even picking up what the main point of the piece is then because it's about agent native not here. This piece is about agent native architectures not comp. What am I typing for? But we can talk about compound engineering as relevant. Yeah. So that's interesting that it just picked up on something that isn't what we want to be talking about and that's why you know you got to keep an eye on the outputs and make sure that you're catching this stuff. Um the core concept is strong and timely now. Um the opening so I should show you what my actual what I actually um this is my version. Dan don't tell me if this is not good. I don't want to know. Um, when you're building agent native software, it's only a matter of time before it does something you didn't tell it to. Sometimes it's great. Other times it's terrifying. That's the shape that I decided to go to just with what the shape of the material that I got, how it was way, you know, here's what's awesome about agent agent native architecture. Here's what you want to write until we look out for. And I tried to go with terrifying just to amp up the drama a little bit. Um, whether that is good or not is a matter of taste, but what the AI thinks. Hey Katie, I need to just interrupt you really quick because we have a very very special guest. Uh I'm going to remove your screen share and welcome Sam Alwin to the stage. Sam, hello. \>\> Hello. \>\> How you doing? \>\> Good. How are you? \>\> Really good. Thank you so much for joining. \>\> Thanks for having me on on short notice. \>\> Uh yeah, absolutely. So \>\> it's been going it's been going quite good. Um, if you if you have not seen it, we published a vibe check on codeex and I'm going to share my screen and and show it to you. Uh, on the every website we have Codex versus Opus 46. Can you see my screen? Here we go. \>\> Um, and I'd really love we're we've been talking a lot about like okay, what do what are we seeing? So, I'd love to share with share with you what we're seeing and I'd love \>\> what your what your sort of vision is for this. But the the main take that that we're seeing right now is that the models are converging a bit. We feel like codeex is got got a little bit more opacy. It's faster. It feels friendlier. It feels like something you might use more day-to-day whereas before it was really for like truly senior engineers who had like gnarly production tasks. Um and we also think opus got a little bit more codeexy. It's actually much better at um like these long horizon gnarly tasks that that codecs have been have been um had been really good at. And Opus also got a little bit slower which is which is kind of interesting. Um when we when we sort of like look at the models together uh if you really put a gun to our head we say they're mostly similar but if you really put a gun to our head our experience so far is that Opus um has the higher ceiling. It did a couple things that we've never seen a coding model do before but it's got a lot of variance. So it does actually some dumb dumb \[ \_\_ \] Um, codeex 53 on the other hand is really smart um, uh, with low variance, but it's not going to reach the heights of it's not not going to reach the heights of Opus 4.6 and it's not going to um, do dumb \[ \_\_ \] in the way \>\> that's interesting. That's like kind of a reversal of what people said about the previous model. So, I guess yeah, maybe there is more conversions going on than we thought. Um, we definitely took the feedback on personality and speed this time. um we will push more on raw intelligence, but it seemed like there was such an overhang on that relative to people's just frustrations around usability that we wanted to go hard there. \>\> And and it also seems like you're you're kind of pushing into like the codeex app itself uh feels like maybe it's more it's not just for senior engineers anymore and that's the way that you're positioning it too. Um, and it it seems like there's this race happening where what what you and Anthropic have figured out is if you build a really great general purpose coding agent, it's very good for all all forms of knowledge work. \>\> Tell me about that. \>\> I definitely think we're going to see a move to everybody using coding models even if they're not aware they are to do a lot of other kind of impressive knowledge work. Uh, and there will be great ways to integrate that. \>\> Yeah. Um, one of the things I'm curious like from for for from your perspective like it seemed like OpenAI strategy at first uh or for the last couple years has been vibe coding and most work happens in chatbt and then serious engineering work happens in codeex and that that split seems to be changing. Can you tell me about that? \>\> Oh, I think that will converge for sure. Uh I think that people will use they will have like one AI uh sometimes they're doing not that serious work sometimes extremely serious work but it'll have all their memory all their context be connected to their company as necessary and you will just sort of like I need to work on a big new scientific discovery or I need to like draft this email uh and that that you know use my computer do whatever you can I think I think that'll converge. \>\> What are what are you using Codex for day-to-day if you are using it? Uh, I built a little app. I had always wanted like a sort of autom like a magic to-do list where you put things in and it automatically tries to get stuff done for you or ask for help where it needs it. Um, and I'm loving it and I love building it. I love using it. Very fun. \>\> What how long did it take? What stack did you use? \>\> Well, it keeps I mean it's not done like that's another sort of interesting thing about this. Like a first version was done like in a few hours, but it it gets better every day. Uh, and in that crazy meta thing, like I put things into the to-do list about how I want to improve improve the app. So, that's that's actually been one of my kind of like interesting takeaways is how much um the nature of software is changing when you make it so low friction. \>\> Tell me more about that. Like what what does that mean concretely that the nature of software is changing? Um, I mean there's a bunch of things that are changing, but this to push more in this one direction, the fact that it all feels continuous and it never is done and every time you have a new idea, you try it and sometimes you revert it quickly. Um, but the like the friction is so low that as I'm using something, if there's even for like a one-off task, I'm like, "Oh, it'd be nice if it just do this thing. It's just done. It just happens. Is that changing at all how you're building software internally at OpenAI? And if so, how? \>\> 5.3 C codeex was the first time a model that we did made was significantly reliant on the model and the tools to get it built. Uh, \>\> you know, there are amazing things happening now with uh upcoming versions of coding models inside of OpenAI and uh yeah, I mean this is going to be how everyone creates software and shame on us if we don't adopt it the first and the best. \>\> Got it. I got to drop in a couple of minutes unfortunately. \>\> I I know you do. So, if you need to drop, feel free to. I really appreciate you coming on. Would love to would love to have you on next time you drop a new model. Thank you. \>\> That'd be great. And I'll go dig into the comparison. I'm very interested to see. \>\> That sounds great. Thanks, Sam. \>\> Thank you. \>\> All right. Always fun when Sam Alman drops into your stream. That doesn't happen very often. Thank you so much. Uh Daario, if you're around, we're available. Hit us up. Um pretty pretty pretty fun. Uh what do you what do you guys think? Kier and Katie, what do you think of uh of that conversation? I think something changed within OpenAI where they get it now uh like software development wise like they get it like they they have the bug of like using codecs to build codecs like that story I think it's happening which is makes me very happy because it means they feel certain pains uh building software uh in in a new way um and and and I love that. So like I I think they will go very fast and they will iterate very quickly and lots of new things will come out of this. That's kind of the the feel I have here. \>\> I agree. What do you think, Katie? \>\> I'm just honestly really excited to try Codeex as a baby vibe coder because it is kind of bringing things out of the extreme senior software engineering where there's no way I'd be able to do anything. Like I want to get my hands dirty and see see what it notices in some of the code bases I have from my way too many Vibe code projects and and just start cross-pollinating between the two models because they do like they do have are good at like different things and have different strengths and like you know toggling back back and forth between Chachi PT and Claude can be productive sometimes just even in the chat experiences. So it's really cool. Um, I'm really excited to see um how that how like the the OpenAI ecosystem from a code perspective like makes me a stronger coder as well. \>\> 100%. Austin, I see you popped in. What's uh what's on your mind? \>\> Uh I actually I just had my first experience in the last hour doing what Katie said. I have like um my like personal project for testing this stuff is this like independent movie theater app where like it's very tough. There's like it's like a Fandango for independent movie theaters and I just did a like um I built it in cloud code. I just sent it to codeex 5.3 and I was like audit this for me. Let me know everything that's wrong with it and what I should do next. And I' I've been texting um people being like oh \[ \_\_ \] this is actually like really cool and really good. as someone who I like did not think would potentially move over from cloud code which was like as another like baby coder like Katie the like thing that feels like okay it was like a big hump to get over just to use cloud code to build something and I'd heard you all talk about you know sending things back and forth and I was like that makes no sense to me but I was like let me try this. I also like the the the guey in the codeex app that's setting up like folders and conversation chats. Um, which is like uh really really cool. \>\> Do you have it set up? Like are you able to demo the every \>\> Yeah. Yeah. I'll share my I'll share my screen here. Um, and while while he gets set up, uh, Austin joined his head of growth, I guess in late de in Decemberish, and over the over the break just like got super cloud code pill and just came back and I don't know if this man is doing anything other than using cloud code um, all day and maybe now he'll be using codeex. We'll see. Um, but it it I think it has really transformed the way he does growth. And now we have all these like bots that are in our Discord, which is what we use for our internal messaging where it's like here's our growth like here's our growth uh stats for the day. And I'm sure there's going to be like a growth report coming out uh about this stream and uh and he's got a whole growth OS that he's put together where we can always we can all query stuff and ask questions about how things are going. and uh Victor who's also on the growth team like built this whole cohort tool where we can automatically build cohorts and send emails to them and all that kind of stuff. So it's just like it has totally changed everything about um the way that we do growth internally at every and it's all because of cloud code and opus but now looks like Austin's using codeex. Awesome. \>\> I'm starting to use codec. I have two projects inside of codeex now running on 53. One is this movie app which I'll I'll show in a second. Um and then one is the Montaigne every growth OS. And most of the work has been like audit this thing happening over here. That's been like a good entry point for me. And that's really just how I started this thing. Um I was like I was surprised as someone who's like really only used cloud code inside of the terminal or in warp at how fast codeex was. I was like oh I'm used to like stepping away and taking time to do other stuff. And it pretty quickly got me this like snapshot of what this thing is, the high-risisk bugs, and started building a a road map for me of what to do. Um, I'll share my screen for some context next. I'll move the screen over to like what this app actually is so people can people can see it. Um, okay. Let's see here. Where is it? Cool. Okay. Um, so this is Lello. Lello is Fandango for independent movie theaters which uh like refuse to get on Fandango because of like fees that they um will take from it. And so it's a very simple like scraper of these theaters in New York and LA uh where you can search by film. Uh the thing that I like the best is that you can connect your letter box watch list so if there's movies you want to see you can quickly find them. um getting the scrapers to work and pull these movie titles has been like incredibly problematic for me. And so that has actually been the part that's been fun of moving it between codecs and cloud code to uh ask it like hey do you have better ways than what's currently being done using the other model to uh like find the film because it might say that the moment is playing and it's the Charlie XX film that's out right now but uh one model might like identify it as the moment the movie that played in like um like 1973. three. \>\> That's really interesting. Yeah, I've definitely found that too is like just being able to trade off two heads are better than one. Uh like each model has their own strengths and like just threading between them, they solve problems that other models like the other model is blind to and I do that all the time. \>\> But also there's one thing like if you use JGPT like more people have JGPT subscriptions like go use codeex you already pay for it like utilize it. So, like it's also like making it easier for people to enter or try it out. Like it it won't give you a major edge if you just use one over the other. They're both very good, especially if you're a VIP coder and want to like do things. They're very capable and just use your subscription. That's also very uh nice one to have now. \>\> Awesome. Uh Katie, any thoughts? Anything you want to you want to add? Anything you're thinking about? I'm thinking about when there will be a harness for writing that lets you be able swap between anthropic and chat GPT because you know \>\> viral \>\> uh that leads me to my next really important point. If you're here you should know we are all from every is the only subscription that you need to stay at the edge of AI. I'm going to figure out how to share my screen so I can show you a little bit about what that looks like. Here we go. Sharing screen. Blah blah blah. Cool. Um, so this is every all of us are from every as I said. Um, every is the only subscription you need to say at the age of AI. We do ideas, apps, and training. On the ideas side, these are all ideas. These are all our ideas. Um, today codeex and opus 46 came out. We got them early and we did a really great in-depth review of them. So, you should check them out. If you if you care about what's coming out, you want to stay on the edge and you're feeling like a little bit, oh my god, like there's so much stuff happening and I want to keep up, but I don't know exactly what to do. We've got you covered. All you have to do is subscribe to every um the day things drop, we can tell you what to pay attention to and what not to pay attention to. We tell you what we use. We do the reach test. So, like I'm about 5050 um opus versus codeex. Karen is about 8020. Naveen is uh 8020 the other way. We also tell you why they're good and who they're good for, all that kind of stuff. Um much better than benchmarks, much better than mainstream uh mainstream news. Uh that's the idea side of it. We also have apps and that's what reminded me to talk to you about this. We have one of our apps is called Spiral. Spiral is a writing assistant with Taste. It actually blends the best of both cloud models and chatbt models to help you do great writing with everything that we know about the writing that we do at every. We also have a bunch of other apps. We also have Kora, which is the app that Kieran builds. Um, Kora is incredible. It is an AI email assistant. And this is what my email looks like with Kora. I have assistant over here. I've got my my daily brief over here. It's like pretty \[ \_\_ \] sick. Um, if you're looking at your dumb Gmail screen or like like bland Outlook screen, you should definitely check this out. Um, and it's getting better all the time. It's going to be agent native soon. We also have Sparkle. We have monologue which is a speechto text app. This is what we've been using to talk here. So you can see monologue is awesome. Um all of this comes with the every subscription and stuff we build for ourselves and then we give to you. Um in addition to that we have events. Open claw camp is coming up tomorrow. It's a live streamed event. We're going to talk about openclaw. It's a lobster boil about it. We have a thousand subscribers already registered. So you should definitely come every.to every.0/ /subscribe to get all this. One other thing that I should say, we've been really thinking a lot and and I'm going to say this and it's going to make Katie and Kieran and Austin smile. Um, but we've been thinking a lot about the future of software here. Sam Alman was on the stream earlier in case you missed it. Um, and when he was on the stream, he was talking about how software changes in the age of AI. We think that agent native is one of the key ways that uh that uh software is going to change in the age of AI. We built a full guide for how to what agent native is and how to make agent native software. And that's on the every website every.guide/agent-native. And it's about a world where software is really just an agent at its core. It's not like a list of rules uh like traditional software is. You should definitely check this out. And that's one of the things we do at Every. And I'm going to get off my soap box now. Uh uh Kieran, why should people subscribe to Every? I think it's really interesting. I think maybe it's the only place where we build software and we get early access to models and we write about them and have like all these things like their loops like they're all working together like I I get early access to model build a feature and learn from it like we're we're really grounding in in its in like actually work. So, if you are a builder, if you want to know how it is actually done, not like some kind of uh influencer take or whatever, like which is interesting and cool, but like we're actually using this day-to-day and that's just very unique. So, I'm very grateful for them and everyone creating this because it's it's very special and it feels very very insane to work with talent like all these people around me which is great like we're all lifting each other up and like iterating and and Sam was also saying like it's about the like the iteration like it's never done like that's what we feel as as like a company and and I think yeah it's really cool capturing it and and you feel that when you join like you feel the energy, the vibes. Uh like definitely. \>\> Thank you. And Katie, why should people not subscribe to every? \>\> They should not subscribe to every if they are AI phobes who are never going to touch AI because um the we're just not the place for you. There are places for you on the internet. Every is not one of them. However, if you are on the on the if you're on the fence and you're curious about AI but a little scared of it, I highly encourage you to check every out because we are, you know, we do um encompass a pretty wide range of experience levels and fluencies. We have practitioners in a lot of different areas, you know, not just software engineering and writing, but design and also operations folks. I would put Austin in that category or our head of consulting Natalya who with with our applied AI engineer Natash built an entire project manager in two weeks. Um so if you're curious about how AI applies in particular to work but you don't know where to start. I think every is a great place to start. \>\> I mean I think the real answer for not subscribing the only reason is if you hate hydration. That's the only those are the only people who don't belong. If you like being dehydrated \>\> hydration haters, \>\> get out. We're We're highly hydrated. Everyone's got their water bottle. Naveen, you just you just joined. Where's your water bottle? See? \>\> Okay. I I to be honest, I was just like deep in my wife coding and then I saw Sam came and what how did I miss that? That's crazy. \>\> He's asking where the monologue OS launches. \>\> Yeah. Uh we we pitched a monologue. Don't worry, he's he's he's gonna be on the beta list. Yeah. \>\> Wow. That that's a Yeah. crazy thing. So, just want to join in and like, you know, just see. \>\> Well, hopefully Dario will be here soon. You know, Dario, if you're listening, \>\> I don't want to miss that. \>\> Yeah. \>\> Um, well, Nive, tell us how how the vibe cut is going. What what were you working on? Which model were you using? And what what were your thoughts? Yeah, I am right now working on syncing dictionary and modes like syn across Mac iOS so that people can you know move it between devices that's like one of the top requested features and it's actually tough one to implement so that's why like taking my time so yesterday what I did is actually I think Dan was also there so I created a plan for like using opus and also using codex I really like the codeex uh solution but not the plan how like you know but uh coming to opus I really loved how uh like you know thorough the plan is the way it's structured it like when I was reading that plan I was like reading it felt like reading a novel so it's like really uh like easy to read and then so what I did is I just asked Opus oh look at codeex solution and then just update your plan and then it gave me this big plan \>\> now what I did is I asked OPU to actually implement the plan and M \>\> it's not great. \>\> Oo, \>\> at least with the iOS and uh Mac app side. So then I asked Codeex to implement the plan. It one-shotted it. So no build errors, nothing. It's working really well. So right right now I'm just working on that. So \>\> clearly clearly a fan here and I love that. \>\> Codex. Monologue is built all of Codex. You should check it out. monologue.to. Um yeah that's that's fascinating \>\> somehow I'm not sure maybe it's maybe opus guys don't use swift and codex people use swift but my understanding is codex like the charg Mac app is actually built by swift uh like using swift but not the cloud code cloud code is a electron based app so \>\> that's interesting yeah I I pretty much use o uh codeex for proof the markdown editor I've been building I've not really use 46 for proof yet. Um, I've done a couple little like little things here and there, but I haven't built any big features, so I don't I don't really know. \>\> Yeah. So, I just missed the main part. \>\> You Yeah, you missed you. \>\> Real work. Naveen, what what question did you have for Sam? Navine. \>\> Yeah. What would you ask him? \>\> No questions. Just just want to \>\> wave. Just wave. Hi Sam, that's it. Yeah, I I I like obviously he doesn't know all the details, but I would love someone from OpenAI on at some point like to talk about like why they're like a little bit behind on like the agent orchestration side and why they're so like there's clearly a reason they're not jumping into it because uh like Anthropic had this like a year ago almost like they they they had it so they could have just copied it but they didn't. And I'm just very curious to hear what the reasoning was there and what their like like what what do they think or maybe they maybe they they say like that didn't work. So I would love anyone from OpenAI listening I would love to like dive into that because I think that would be an insane unlock for Codeex. \>\> I I agree. If you're from OpenAI DM me. We are actually going to be I'll be interviewing the OpenAI team, the Codeex team next week, I believe, on Tuesday and there will be a uh episode of our podcast AI and I uh released probably or the week after in the coming weeks. Uh so you should stay tuned for that if if you're interested in that. I will say like having talked to them a lot over the like years, I just my sense was internally if you're and if you're at Open AI and you're like this is not true, please like jump in in the chat and tell me. So, I'm just \>\> I'm going off right now, but I so I could be wrong, but \>\> I think that there was I felt this sense that the agent needs to be like really locked down and really like it's a tool where it does exactly what you tell it to do. And um the and it just has this very the the fact that Codex is a senior engineer has that vibe is I think a reflection of it was built for people that wanted to use it that way and that think that way themselves and therefore the idea of like swarms of agents felt like it it they were not getting there yet. It was like we just want to make this like power tool that does exactly what we say and is always right. And I just think that that's a bit of a different vibe than Boris from Anthropic just being like, I don't know, it'd be kind of cool to just like run this thing on my computer and see what happens, you know? Um, and but I and so I think it took them a while to catch up because my perception is they were looking at cloud code and being like, yeah, this isn't for serious engineering like and this is not secure. And then over the last like four or five months, they've been like, oh, this is actually like \>\> maybe there is something to it. \>\> Deal. Yeah. \>\> And it's the story also of the models converging a little bit. So they're like learning from each other because clearly like there is a point to like Naven loves codeex and he's like the the like the the the most brilliant IT engineering whatever like I'm not that I'm a musician so like it makes sense I vibe more with one versus the other and and them learning from each other and pushing is great. Yeah. Naveen went to the MIT of India which um millions of people take a test I think like two million people right Naveen \>\> yeah two million \>\> two million people Karen's internet probably just went out again two million people take a test to get into this school and only like a couple thousand people make it and Naveen is one of those couple thousand people and he loves codeex and Kieran's just he's just a musician um so Kieran is incredibly talented he's a musician he's a baker He's got that kind of like multi-dimensional brain that that we love at every and it's just like it's two different sorts of vibes and two different sorts of use cases. I think Kieran also uses codeex but only for like one of like those times where he's like I just need a lot of precision and I think Naveen you're similar like you've used quad code but only in those times where you're like I need a little bit more creativity and like yoloess right? Yeah. So I think I I remember using actually Opus when by coding the initial version of iOS right because you you don't know you're prototyping you're experimenting what the app can be what the product can be then Opus 4.5 or now 4.6 is like really good at vibing and like me just giving one line and then it will just run with it. uh but codex is like then yeah don't know it you need to be much more precise with codeex but now I'm not sure like I need like because with 4.6 six opus it's also like thinking really hard it's also like trying to solve problems so like I think that might not be true uh being you know the senior engineer versus uh like the previous analogist might be now becoming you know slowly not being true \>\> but still opus needs to like get better at swift. Yeah, \>\> Katie, \>\> I was just gonna say like the difference between senior engineer and everyone else is something that Sam that Sam nodded at. So Sam would approve. I will go out on a limb and say \>\> Sam approved. Um, all right. So I have a flight to catch. This stream has been going on for three hours. It has been incredible. We have had so much fun. We've explored Opus 46. We've explored GBT 5.3. We've explored writing. We've explored coding. We had Sam Alman on. It's been the best. Like truly awesome. Um, if you want more of this, we are going to be back for paid subscribers tomorrow for OpenClaw camp where we're going to talk about OpenClaw, demo what we do, demo our best use cases, demo what not to do, and we'll help you get it set up. So, if you've been looking at wanting to try OpenClaw, we will take you through how to set it up. Uh, you should also subscribe to Every. Every is the only subscription that you need to stay at the edge of AI. Uh we we do ideas, apps, and training. If you like this, go to every we have a ton of more stuff there. Every.TO. Uh we have a full breakdown of both of these models from our our extensive hands-on testing. We get all the models early. We try them ourselves for the stuff that we build. We build software ourselves. We also write about it. Um and we also do training. So that's that's all I got. Uh Austin, close us out. Anything that I missed? Anything that you want the world to know? We've got 26,000 people on this stream. You've got 20 seconds. Go. \>\> Uh if you want to get an open clause set up in 90 minutes, subscribe to every and tune in to the session tomorrow. We never stop. \>\> \[ \_\_ \] yeah. Uh Katie, uh final final thoughts for the crew here. \>\> Um if you're curious about writing an AI, there is a writing camp coming up later this month. So that is another reason to subscribe to every \>\> sweet. Um, Naveen, what was your favorite thing about going to the MIT of India? \>\> Favorite thing? Just people. It's a ton of smart people. Uh, but uh if you want to subscribe to every like you can get all the apps for free like you you can try Monog to talk with cloud codecs. I think uh talking with it more than typing is you get better results. So use monologue. You can get it from every.TO. Awesome. Thank you. Thank you everybody on this stream for joining. We'd love to see you in our Discord. We'd love to see you again soon. And uh remember, stay hydrated.