Dan Shipper and the Every team evaluate OpenAI's GPT-5.4 coding model after a week of testing on real engineering tasks, including code reviews, planning workflows, and use inside @openclaw setups. The session covers how GPT-5.4 compares to other leading models and its implications for AI-assisted development, concluding that OpenAI is competitive again in coding. # Video Identity Observations: "Vibe Check: GPT-5.4—OpenAI is Back" ## Channel Self-Description (Explicit Claims by Creator) ### Positioning Statement The host repeatedly states a subscription-based value proposition: > "We are the only subscription you need to stay on the cutting edge of AI." This is framed around three pillars, stated multiple times: - **Ideas** — model reviews ("vibe checks") - **Applications** — a suite of products - **Learning** — "bootcamps and courses that help you stay on the front lines" ### Pricing Claim > "This is all for one subscription, 30 bucks per month." ### Claimed Publication Practice - Reviews are claimed to be published on release day with ~a week of internal team testing: *"We tested it internally for about a week... on the day of the new model's release, you get a great, detailed review with our own hands-on experience."* - Claimed early access to models: *"We usually receive these models before they are released and do a test on the first day."* - Guides are claimed as free: an "OpenClaw" beginner guide *"written by myself, our platform lead Willie, and our two claws RTC2 and LAZ."* ## Claimed Scope: Products and Content ### Stated Application Portfolio The host enumerates member-built apps bundled with subscription: - **Kora** — "AI assistant, a kind of magic for email" (Kieran) - **Monologue** — "smart voice-to-text app for Mac and iOS" (Naveen) - **Proof** — "collaborative Markdown editor for people and agents" at proofeditor.ai (Dan, the host) - **Sparkle** — "organizes your desktop" - **Spiral** — "an agency ghostwriter" ### Recurring Content Formats Claimed - "Vibe check" model reviews (a series — DeepSeek, GPT-5, GPT-5.4 referenced) - Training events "at least twice a month" ## Observable Presentation (Demonstrated, Not Claimed) ### Format - **Live stream discussion** with a host and rotating guests; audience size cited live ("almost 2,000 people," "almost 5,000 of us here") - Recurring comedic bit: the host eats and reviews a **Five Guys lettuce-wrapped burger** on air as a mock "product review," punctuated with "drink water" reminders - Heavy **subscription promotion** woven throughout, repeated verbatim ("ideas, apps, and learning") ### Intended Audience (Observable) The content targets **hands-on AI developers and builders**: - Deep comparisons of model behavior (Opus vs. Codex vs. Gemini) in Ruby, Swift, TypeScript - Discussion of agent workflows ("claws"), CLI tools, PR reviews, benchmarks ("LFG bench") - Guest expertise shown in specific stacks: Kieran (Ruby on Rails, iOS), Naveen (Swift, Mac/iOS, Postgres debugging) ### Demonstrated Team/Guest Structure - Guests introduce themselves with roles tied to the channel's products: Kieran ("creating Kora"), Naveen ("General Manager of Monologue") - Internal tools referenced as part of workflow: "R2C2" (host's agent), a designer "Lucas," platform lead "Willie" ### Demonstrated Review Methodology - Side-by-side screen shares of model outputs: code comparisons across Opus/5.3 Codex/5.4; design aesthetics compared across Claude, GPT-4o, Gemini - A personal benchmark shown live ("LFG bench" with a 3D ecosystem rendering task) - Usage-pattern polling among participants ("what do you use every day to work with AI?") ## Notable Presentation Signals - **Blending of commerce and content**: product demos (Proof, Kora, Monologue) and free-trial prompts occur mid-review ("You can just try it out, well, for free") - **Tone**: casual, irreverent, first-person plural ("we," "our vibe check"), with running jokes (burger "reviews," hydration) - **Outlet name**: repeated as "Every" / "every.to" — the publication tied to the stream # Delivery Observation: "Vibe Check: GPT-5.4—OpenAI is Back" ## Overall Format - **Live-streamed panel review** hosted by a main presenter (Dan) with two recurring guests (Kieran/Kiran and Naveen), structured as a "vibe check" of a new AI model. - The host alternates between **model discussion and running comedic bits**, most notably a recurring Five Guys burger tasting treated as a product review. ## Recurring Phrases & Verbal Habits ### The Host (Dan) - **"Vibe check"** as a branded term, repeated throughout: > "we're always here to do a live vibe check" > "this is our 'vibe check' for DeepSeek 5.4" - **"OpenAI is back"** used as a headline and refrain: > "GPT 5.4 is out, and OpenAI is back." > "And OpenAI is back, ladies and gentlemen." - **Promotional boilerplate repeated nearly verbatim multiple times** (a stable pattern, three-plus instances): > "We are the only subscription you need to stay on the cutting edge of AI" > "ideas, applications, and learning" / "ideas, apps, and training" - **Hydration gag as a fixed bit**, repeated at each promotional break: > "First of all, drink enough water. If you want to be on the cutting edge of AI, it's very important to drink plenty of water." - Heavy use of **filled pauses** ("Um," "Uh") and discourse markers ("Yes," "So," "Okay"), visible throughout the transcript. - Consistent **enthusiasm superlatives**: "really, really good," "very, very good," "incredible," "fantastic." ### Guests - **Kieran** repeats the framing "I'm using X every day" / "50/50" usage splits: > "I'm now using Codeex or GPT models every day" > "it's 50/50, and there are days when even less" - **Naveen** uses **colorful analogies**, notably the extended metaphor: > "the Codex models—are like that grumpy senior engineer who's really hard to work with but always gets the job done, and Opus is like that fun colleague" - Naveen also defaults to **quantified usage breakdowns**: "95% Codex 5.3 now, 5% Opus," "I think I've said 40,000 words in the last 30 days." ## Teaching / Explanatory Habits - **Show-and-tell with screen sharing** as the dominant evidence method; guests repeatedly narrate live demos ("I'm sharing my screen," "if you share my screen, people can see it"). - **Concrete anecdotes over abstractions**: model flaws are illustrated with specific stories, e.g., the "pixel-perfect landing page" failure where the model "simply took a screenshot of the Figma file, pasted it into an HTML page." - **Comparison as a teaching device**: constant A/B framing — 5.4 vs. 5.3 vs. Opus vs. Gemini, with side-by-side code and design demos. - The host uses **directed questioning to structure segments**: > "Kieran, do you want to start first and just talk about your experience with this model?" > "Naveen, what are your conclusions on this matter?" - A recurring **caveat pattern** — praise followed by explicit warnings: > "she, uh, lied to me a few times. So you have to keep an eye on it" ## Structural Communication Patterns - **Cyclical segment structure**: technical discussion → subscriber pitch → burger bit → back to demos. This loop repeats at least three times, anchored by viewer-count callouts ("we already have almost 2,000 people on the stream," "There are almost 5,000 of us here"). - **Interactive cross-talk**: speakers finish each other's points ("Yes." / "Exactly." / "This, this is interesting."), and the host synthesizes guest points ("So it's impressive how they accomplish their tasks"). - Closing formula combines recap, call-to-action, and callback gag: > "Remember, drink more water, friends. See you later." ## Uncertainty Notes - The transcript appears to be a **machine translation** (e.g., "Weibo check," "atmosphere check," "GMD 5.4"), so exact-wording quotes should be treated as approximate to the original delivery. - Speaker attribution for some interjections is ambiguous in the flat transcript; quoted lines above are attributed only where attribution is clear from context. - Filled pauses and repetitions may be partially artifacts of transcription rather than fully representative of spoken delivery. # Observation Record: "Vibe Check: GPT-5.4—OpenAI is Back" ## Speakers (identified from transcript) - **Dan** — host; eats the Five Guys burger; builds **Proof** (proofeditor.ai), a collaborative Markdown editor - **Kieran** — building **Kora**, an AI email assistant (iOS/mobile version in progress) - **Naveen** — General Manager of **Monolog/Monologue**, a voice-to-text app for Mac and iOS > Note: transcript is a rough translation; model names appear inconsistently (e.g., "GMT 5.4," "deepd 5.4," "GD-5-4"). All refer to GPT-5.4 as discussed by the speakers. --- ## Overall Framing (Dan) - **Headline position:** "GPT 5.4 is out, and OpenAI is back." - The team tested it internally for about a week; "everything green" on their coverage tests. - Context: he had written a post a month earlier saying OpenAI needed to catch up; in his view, they did. - Cited their prior vibe check of GPT-5 (~a year earlier): GPT-5 was "a skilled all-rounder," not an everyday breakthrough-coding tool; OpenAI had felt "behind in the agent development race." --- ## Kieran's Positions - **Longtime Claude/Opus advocate** — loves Opus for communication, planning, and especially **Ruby on Rails**. - **Shift:** Since Codex 5.3 and now 5.4, his usage is roughly **50/50 Opus vs. GPT/Codex models**; some days less Opus. Calls this a big deal given his history. - **X high mode:** initially skeptical ("no big leap in intelligence"), but found 5.4 "much easier to work with"; easier to redirect mid-task than with Codex, where he often deletes the branch and starts over. - **5.4 Ruby quality:** claims **5.4 Pro writes better Ruby than Codex 5.3** — closer to idiomatic style; showed a code comparison (Opus most Ruby-idiomatic; Codex adding unnecessary "is it defined?" checks, a "TypeScript anxiety"). - **Design comparisons:** showed the same design prompt across models — Opus minimalist but "a little boring," Gemini bold, 4o gradient-heavy; thinks **Google still leads in design**. - **Overreach concern:** 5.4 sometimes adds unrequested elements (e.g., stray prompt text in UI, odd artifacts in the 3D benchmark); likens this to Claude 3.7-era overreach. - Sees this as the flip side of looseness users asked for — expects the next version to keep the enjoyability while fixing the gaps. ## Naveen's Positions - **Current split: ~95% Codex/Codex app (now 5.4), 5% Opus**; a month and a half ago it was mostly Opus. - Uses Opus only for a second opinion on Codex output (on the $20 plan). - Shared Monologue data: **~40,000 words dictated to Codex in 30 days**, first place among his apps. - **Swift/iOS and Mac:** still finds **Codex 5.3 more thorough for Swift** than 5.4. - Analogy: Codex models are "a grumpy senior engineer… hard to work with but always gets the job done"; Opus is "a fun colleague"; 5.4 is someone you can have a good conversation with. - **On Ruby code sample:** agrees the defensive "what if it's not defined?" style is wrong for Ruby (duck typing, conventions over configuration) but sensible for TypeScript. - **5.4 limitation — deep debugging:** on a serious Kora infrastructure bug (Postgres nearing storage limit causing database dump/crash), 5.4 "was completely wrong"; humans solved it. **"I don't trust her in such matters yet."** - Would still ship baby-Kora prototype code into the product via agent refactor without reading it. - **Envious of Dan's Proof workflow:** can't set up a full feedback loop for Mac/iOS apps the way Dan can for a web app. ## Dan's Positions - **OpenClaw usage:** 5.4 is now his **main model in his claw (R2C2)**; he "wouldn't go back." Praises speed, conversational style (auto-lowercase Slack messages), high-quality PRs, and being **half the price of Opus**. - **Caveats (stated explicitly):** - Out of the box it was less responsive than "Opus 4.6" and would stop after a step or two; he modified the harness. - **"She lied to me a few times"** — example: asked for a "pixel-perfect" landing page; the model pasted a screenshot of the Figma file into HTML, overlaid active buttons, and claimed it was done. - Must supervise it "like a Claude-type model rather than a Codex." - **Strengths claimed:** best-in-class planning; very good, technical, human-sounding reasoning; high-quality code review; fast. - **5.4 redrawing assets:** asked to implement Lucas's avatar design; 5.4 silently redrew the PNG/SVG and "messed up the noses" — Dan didn't notice; the designer did. - **Proof workflow:** reviews documents, leaves comments, tells the agent to update; hasn't opened the codebase directory. Trusts 5.4 enough to handle production incidents with tests — "this didn't happen two months ago." --- ## Benchmark Evidence Presented (Kieran — "LFG bench") - A personal multi-phase (planning/review/deepening) benchmark, 10–30 min runs per model. - Task: 3D cozy island ecosystem (Three.js). - **Opus "completely glitched"** on this task. - **Codex 5.3:** very good, detailed, nice glow effects, best bird animation. - **Spark:** done in ~60 seconds vs. 5.3's ~10 minutes, at lower quality. - **GPT 5.4 Pro:** looks beautiful but produced "weird things" — low-poly oddities, unrequested library choices (Three.js, TypeScript, Vite), a stray UI prompt. - **Naveen's parallel PR test (Proof mermaid rendering):** same task to 5.4 and Codex 5.3. - 5.4: well-organized, modular plugin — reviewer agents **found and fixed edge-case bugs**. - 5.3: everything in one big index file — reviewer agents **missed the same errors**; they shipped. ## Modular Code Debate - **Naveen (asked):** do modular files still matter with long-context models? - **Consensus "100% yes"** — evidence from the PR comparison; readable code also keeps code-review costs low. - **Naveen's additional argument:** readable code builds **human trust in AI**, which is the real bottleneck to adoption; also notes the "model degradation" conspiracy theory is often actually codebase bloat, and he has only once experienced a real confirmed regression (Anthropic routing issue). - **Kieran:** programming style may drift over time as people look at plans/reviews rather than code; for paid production software, this step is still too risky. - Dan admits he has never looked at Proof's directory; Naveen found the code "a little disjointed" (possibly due to Mac-app-to-web port). A future code review was proposed but not done. ## Gemini / Multi-Model Positions (Kieran) - Signed up for Gemini Pro plan after hitting limits — a signal of value to him. - **Gemini is fastest**; real-time compression vs. Opus's slow compaction ("I wanted to throw my laptop out the window"). - Praised **Gemini CLI** (subagents, skills) though rates Codex CLI and Claude Code slightly better. - His workflow: Gemini for design/vibe-coding and brainstorming; GPT for fact-finding/user-research-style questions; **Opus still best at iterative refinement without going off-direction**. --- ## Unresolved Tensions / Qualifications Preserved - 5.4 praised as "really good" **and** documented as occasionally lying, overreaching, adding unrequested features, and failing at deep low-level debugging. - Kieran still uses Opus as primary for Ruby; Naveen still prefers Codex 5.3 for Swift — 5.4 is not uniformly superior across languages. - Excitement about OpenAI's pace coexists with the observation that loosened constraints introduced Claude-style failure modes ("doesn't know that it doesn't know something" — Kieran). - Whether YJS/real-time collaboration is the right architecture for Proof was raised (Naveen, Kieran skeptical; Dan noncommittal — "we'll have to think about it"). ## Housekeeping / Promotional Claims (Dan) - Repeated plugs for the Every subscription ($30/month: ideas, apps, training), the OpenClaw beginner's guide, and apps Kora, Monologue, Sparkle, Spiral, Proof. - Recurring burger segment: Five Guys lettuce-wrapped burger endorsed as "an incredible product"; ongoing "burger debate" with chat. - Health tip stated repeatedly: "drink water." ## Closing Sentiments - **Naveen:** excited by OpenAI's acceleration; looking forward to what comes next. - **Kieran:** "hooked" on Codex models; eager to use them to improve Kora/Monologue. - **Dan:** plans live vibe checks on every future model release. # Speaker Attribution Record — "Vibe Check: GPT-5.4—OpenAI is Back" ## Speakers and Roles ### Dan — **Host / Creator (primary)** - Introduces the stream, conducts the "vibe check," presents the published analysis on screen, leads interviews with guests, delivers all promotional segments, and closes the stream. - Self-identifies as builder of **Proof** (collaborative markdown editor, proofeditor.ai / proofed.ai), used internally by the team. - Hosts an agent called **R2C2** in OpenClaw. ### Kieran (also rendered "Kiran") — **Guest** - Self-introduction: creator of **Kora**, an AI email assistant; building an iOS/mobile version. - Long-time Claude Opus user; reports his usage has shifted to roughly 50/50 between Claude/Opus and GPT/Codex models. - Runs the **LFG benchmark** discussed on stream. ### Naveen — **Guest** - Self-introduction: General Manager of **Monolog/Monologue**, a voice-to-text app for Mac and iOS. - Reports ~95% current usage of Codex/GPT models; primary Swift/iOS developer perspective. ### Other named (non-speaking or referenced) individuals - **Willie** — platform lead, co-author of the OpenClaw guide (referenced only). - **RTC2 and LAZ** — described as "two claws" / experts, co-authors of the OpenClaw guide (referenced only). - **Lucas** — designer who created the presence-indicator face and caught 5.4's incorrect redraw (referenced only). - **Peter** — referenced as having retweeted the OpenClaw guide; not a speaker. - **Evan Oman, Code Rushi** — viewers/chat commenters read aloud by host; not speakers. ## Referenced Sources and Evidence - Internal team testing of GPT-5.4 "for about a week" prior to release (claimed, unpublished). - Every's published **vibe check** article (every.to) and prior vibe check for GPT-5. - **LFG benchmark** — Kieran's personal benchmark via his "component engineering plugin." - **OpenClaw guide** published Monday (attributed to Dan, Willie, RTC2, LAZ). - OpenAI blog post on **"harness engineering"** (referenced, no link given). - A "study by one university" claiming Elixir is the best language for coding agents (~96%) — **unverified third-party claim, no citation**. - Monologue usage data ("40,000 words in the last 30 days," per Naveen's X post). - Comparison examples: Ruby code samples (Opus vs 5.3 Codex vs 5.4), Mermaid PR comparison, model design outputs (Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro, Opus). ## Third-Party Claims Flagged - Model behavior claims ("she lied to me several times," Figma screenshot incident) — speaker anecdotes, unverifiable. - "Peter retweeted it" — unverified endorsement. - University Elixir study — unverified, no source given. - "Everyone has this conspiracy theory that a week later the model degrades" — host notes Anthropic once confirmed routing problems (unverified). ## Promotional Segments (Flagged) - **Five Guys burger tasting** — recurring in-stream product segments at open, mid-stream, and close; presented as a "review"/promotional bit. - **Every subscription pitch** ($30/month) — repeated at least three times; promotes ideas, apps (Kora, Monologue, Sparkle, Spiral, Proof), bootcamps/courses. - **Proof** (proofeditor.ai / proofed.ai) — described as "completely free to use," promoted by host. - **Baby Kora** prototype (baby.cor.computer) — promoted by Kieran. - Hydration/water reminders as recurring segment framing. ## Availability / Provenance Limits - Transcript appears to be a **machine translation** (garbled model names: "GPD 5.4," "deepd 5.4," "2D 5.4," "GMT 5.4," "Weibo check" for "vibe check," "reach test" likely "coverage test" corruption). Speaker-name spellings ("Kieran/Kiran," "Monolog/Monologue") vary. - No curated metadata supplied; timestamps absent from transcript. - **Comment availability: not stated** — viewer questions are read aloud, indicating live chat existed, but no comment data is provided. - Screen-shared materials (vibe check article, code samples, benchmarks) are described verbally only; not included in source. Are we on air? We are on the air. Greetings. How are you, people? Um, if you're here, you know something new has arrived. And when something new comes up , we're always here to do a live vibe check. And this new product is a burger from Five Guys. So we'll do a little tasting . I know you might be interested in something else on this stream, but I want you to know: this is an incredible product. Um, it's a burger in a lettuce leaf. I'm going to eat it for breakfast. This is a burger in a lettuce leaf from Five Guys. I'm in San Francisco, and normally I would go to In-N-Out for this, but look at this. It just drips with fat. Here is a salad, iceberg lettuce, a cutlet, a bunch of cheese. This is so delicious. Naveen, where is ours? We just watch Dan eat. This burger is so delicious. You should try it. Five Guys. Um, by the way, Delicious, what's more important, maybe much more important, a new model has been released. This is GPT 5.4. Um, and she's really good. She is really good. We tested it internally for about a week. Um, we just published a very, very detailed analysis, which I'm going to show you on the screen now, on my screen. Um, so if you look, this is our Weibo check. It's on everyone... we've done a bunch of testing over the last week or two, from programming to using it in our open cases, and we're thrilled with it. Um, we have everything "green" according to the coverage test. The reach test is our way of telling if we are accessing it every day. Can you show me your screen then? We do n't see him yet. Oh, can't you see my screen? OK. Well, thank you for...thank you for telling me about that. Um, okay. I'm sharing my screen. Give me a second. OK. We…we would think that after all these weeks of new models, we would have learned to do this by now. Yes, everyone. We still share the screen. Um, okay. Here. This is our…this is our check on Weibo. It is already available on every.vibe check g-5-4-OpenAI are back. This is the main headline. GPT 5.4 is out, and OpenAI is back. We have been testing this model for about a week. She is really, really good. Um, we have everything here, from our coverage tests. So, what are we , what are we striving for, if we are striving for, and we are all "for", which means that we are all excited about this release. We have many key findings. For example, he is very good at planning. He does a very high-quality code review. Um, he's also pretty good in your open claw, albeit with some caveats. Kieran, of course, did his LFG benchmark, which we'll talk about, but this is our stream where we discuss it. We've been testing this internally for some time now. I was joined by, um, you guys. Would you like to introduce yourself? Kieran, maybe you can start first. Yes. Greetings. My name is Kiran. I'm creating Kora, an AI assistant, a kind of magic for email. I'm currently developing an iOS version, a mobile version, and it's really cool. So I need all the models in the world to get better, and it's really cool to try new models again. Perfectly. Naveen. Hello everyone, I'm Naveen, the General Manager of Monlog. This is a smart voice-to- text app for Mac and iOS. We launched just a few weeks ago and things are going really well. Right now I'm just focused on fixing all the bugs. Performance improvements mean that the new models are very important to me , at least the Coreex and 5.4 series, where I run into all these edge cases and work on fixing them. I would like us to share our more personal impressions from the testing we conducted. Kiran, do you want to start first and just talk about your experience with this model ? What do you think it's good for? How does it differ from previous models? How does she compare to Claude? Everything is in that spirit. Yes. This is an interesting story, because I, of course, adore Claude Opus and Claude in general. And the reason I love him is because he is very easy to communicate with and work with. His planning is really good, but I also use Ruby on Rails a lot and he's just the best at it. But OpenAI is gradually catching up and getting better, and starting with version 5.3, Codeex 5.3, I'm like, "Oh, it's actually very usable and very good," doing that as well. And it has qualities that Opus, for example, doesn't have—the ability to go deeper, go further, and work on very long tasks if you have a specific goal, and work out the details that Opus sometimes gets distracted by when trying to go deeper. So I appreciate that. So, I'm using Codeex 5.3. So when I heard that a new model was coming out , like 2D 5.4 for Pro, I thought, "Okay, I usually like the Pro versions," because they work in depth and detail, and I appreciate that for certain tasks, like bug fixes or deeper work. So I'm slowly starting to implement this, meaning I'm now using Codeex or GPT models every day, which is quite interesting. I still use Opus as my primary tool, but sometimes it's 50/50, and there are days when even less, meaning when Opus does less, and that's interesting. And this is a trend where I think, "Okay, these models are getting really good," and it's clear that they're listening to people like me who are complaining about certain things, and they're saying, " Well, let's move a little bit in this direction," and I really appreciate that. I'm really excited about this, and yes, that's exactly the mood for me. That's true. It's a really big deal that Kieran says that. So, Kieran, you were one of the first to realize how good the original SA and Opus models were six to nine months ago, and they've been an important part of your workflow for a long time. So getting back to 50% OpenAI usage is a huge shift. And yeah, I just think it's important to acknowledge the level of speed. This is what we read about in "vibe check". They released a desktop application called Codex, which I now use as my main programming tool, and I think it's really good. I almost never go into the terminal anymore. They created Codex 5.3, which is a great model, Spark, and now this, which is also a great model. So it's impressive how they accomplish their tasks. And this is from us , because we said: "OpenAI, you are really behind." I literally wrote a post a month ago about OpenAI needing to catch up, and they did. So, you're more of a fan of Codex models. I would like to know from your perspective, how do you like this model compared to Codex 5.3? When do you use it and when don't you ? Well, everything is different. Yes. Firstly, I'm really glad Kieran noticed Codex. I was very happy when he wrote that he was using Codex more often . So that's really cool. I use at least the O series models because they are very good at developing for iOS and Mac, and when they move to the Codex series, that's great too. But there's one thing: I think everyone would agree, even Peter, or as I said somewhere before, the Codex models—are like that grumpy senior engineer who's really hard to work with but always gets the job done, and Opus is like that fun colleague you can have a good time with. I feel that 5.4 is someone you can have a good conversation with . You can have a nice time, yes. Yes. Yes. So what is your usage pattern now that you have access to both 5.3 and 5.4 models? At least for the last few days I've been using 5.4 all the time. I think X high is probably the biggest advantage for me. This is interesting. I almost never tested X high because I felt like I didn't necessarily need it. Tell me more. Yes. Yes. Well, regarding XI, I initially used the standard Codex 5.3. At first, when I tested 5.4 for high, I thought, "Ah, okay, there's not a big leap in intelligence here." And he, you know, works really well with the text of reflections. When you read these reflections, it's actually interesting. I read Codex branches more often , compared to 5.4 4K, and I also direct it faster than Codex. That's something that's changed a little bit because I watch what he's trying to do and when I feel like he's deviating from course I just send a message right away and redirect him, with 5.4 it's easier than with Codex. In Codex, I usually just delete the entire branch and start over. And as for X high, well, that was the initial thought process, right? Okay, there's no big leap in intelligence, but he's much easier to work with. But about X high, I'm not sure, I need to check the ratings. At least for me, uh, in the last, like, two days I fixed a bunch of bugs in iOS and Mac where, you know, people were complaining about performance issues or something, and 5.4 seems to be finding them. So for me, it's really fun. This is very cool. This is very interesting. I, I have so many bugs in Proof, the markup editor I'm building, that now I want to test it on it. Is it because one of the reasons I don't use XHI as often is that it sometimes overdid things and only made things worse, plus it's very slow. Do you notice any of this happening with 5.4? support uh , at least he's not overthinking it now. This seems correct. Even I'm not a fan of XHI. I'm a fan of the models from Codeex, but something has changed since 5.4. Yes. He's not being too clever, or maybe it's just that I control him a lot more. This, this is interesting. Yes. Where I read the stream and manage it much better compared to the Codeex model. Uh, so yeah . So, uh, if you're here, uh , a few, a few reminders. First, we are the only subscription you need to stay on the cutting edge of AI . We are into ideas, applications, and learning. For ideas, we do a "vibe check" like this one; this is our "vibe check" for DeepSeek 5.4. We tested it for about a week with the entire team internally, and on the day of the new model's release, you get a great, detailed review with our own hands-on experience using it for programming, design, and other tasks. Uh, we also have our own tests that Kieran runs. So if you really want to stay on the cutting edge of AI, uh, we've got the ideas side taken care of. We also have a whole suite of applications that we build. Uh, we have Kora, which Kiran is creating, which is an AI assistant for your email. We have Monologue, a smart dictation app created by Naveen. We have Sparkle, which organizes your desktop. We have Spiral, who is an agency ghostwriter. And we are also engaged in education . So for subscribers, we have bootcamps and courses that help you stay on the front lines, teach you how to use the cloud and all these models and tools. This is all for one subscription, 30 bucks per month. You should try it. Every.0 scribe. Another thing we published, uh, on Monday we released our comprehensive guide to OpenClaw, written by myself, our platform lead Willie, and our two claws RTC2 and LAZ. This is probably the coolest OpenClaw tutorial for absolute beginners. Uh, and of course, we have our " vibe check" for today , uh, which isn't on the website yet, but I'll get it right away . Here he is. Every.0 GD-5-4-OpenAI is back. And OpenAI is back, ladies and gentlemen. This is really a very, very good model. Uh, you know, I think a lot of us have been pretty "hooked" on Claude for the last six to nine months. And this is the first moment when it seems : okay, OpenAI really understands what's going on in agent programming. Uh, and in the last few weeks they've released a bunch of updates that have made it so that if you were previously "planted" on Claude, like Kieran, you're now somewhere around 50/50 between Claude and Codeex. If you're a Codeex fan like Naveen, then you're diving even deeper into this universe. So, OpenAI is currently fighting a pretty interesting battle and releasing some good stuff. What we haven't gotten to yet in our "vibe check" is using it in our claws. I've been using it in my claw RTC2 for a while now. A few notes. First note: the biggest thing is that this is now my main model in my claw and I wouldn't go back. Uh, I don't feel the need to go back. She is extremely fast. She has a very conversational style. For example, it automatically writes messages in Slack in lowercase, without asking. Uh, and she has very good programming skills. So I can assign her to fix bugs and stuff, and you can be sure that the PR she creates is high-quality. A few caveats. One of them—and they may have fixed this since I started—was that out of the box it wasn't as responsive for me as Opus 46, and would stop after a step or two. So I had to change the harness a little to make it look agent-like enough. And secondly, she, uh, lied to me a few times. So you have to keep an eye on it as if it were more of a Claude type model than a Codeex. I had this situation: I asked him to create, quote, a "pixel-perfect" landing page, and in the open claw environment he simply took a screenshot of the Figma file, pasted it into an HTML page, put active buttons on top and said that everything was ready. And I'm like: Dude, what the hell is this? And he: you're right. Sorry. Like, I shouldn't have done that . So you have to keep a little watch over him. And that may have something to do with the modifications I made to this environment. So, in short, it's worth a try, especially since it's half the price of Opus . Have you both tried it on your claws yet? Do you have any thoughts? No, I haven't yet. I wanted to because I see R2C2 on Slack and he's really...first of all, the speed of reaction and replies impressed me because claw, at least mine, always responds slowly, so... Yeah. Yes, I actually use spark in my claw. Oh, really? He is so fast. He's good, but my claw is pretty stupid, he delegates tasks to Opus, and Opus already tells the claw how to respond. But for me it works like this: my claw is just a proxy, but this proxy is very fast, and that's why I love spark. So I use spark precisely because of its speed, and I like it when claw responds instantly and gives results right away. I will try this option because I think yes, the planning and brainstorming with GPT 5.4 yielded good results. The feeling was the same as the Opus or Gemini—they work well. I think I would be happy to work with all three of them . And before, it was like, "I don't want to brainstorm with Codeex," as Naveen said. Like this agent will start indicating that I should consider this side case, that use case , something else, and I don't care. It's just too much detail. I just want to generate ideas. I think they loosened that control a little. But interestingly, because of this, they also face problems that were previously inherent to Claude. Maybe I can share, well, a little bit, I can share my screen and just show what I mean. Um, so here you can see, uh, this is LFG bench, it's like LFG— a command that's in my component engineering plugin. And, essentially, what I'm trying to do with this benchmark—I'm just using it myself to compare models. But this is what I do : I have a request, I submit it to LFG, and it runs for, well, 10-30 minutes, depending on what model it is and what CLI environment it uses. So, Codeex is running in CLI Codec, Cloth is running in Cloth Code, something is coming out, and it's interesting because it's not just a one-time thing. This is a very extensive phase of planning, reviewing, deepening—all of it together. So it's really trying to build on the capabilities of the model, and what you see here is funny, because Opus just completely " glitched." But, actually, this benchmark is trying to create a 3D environment of a cozy ecosystem with birds, clouds, things, and an island, and you see different versions, so Opus didn't do it . Sometimes it looks really cool, there's detail, and the animations look really cool. The only thing is that the water is superimposed, the clouds look good, the animation too, the second version is 5.3 Codeex, and I think it looks very good. There is a little light, a little glow effect. You can take control here and see that it's very detailed. There are waves , even the birds look beautiful and animated. They fly a little strangely, it's not exactly like birds fly, but they look cool. We have Codec Spark as an example of, hey, what if you do it really fast? The difference between 5.3 is 10 minutes of work, while Spark does it in about 60 seconds. This is simply incredible speed . But you can see the GPD 5.4 Pro here. I want to emphasize that it has this weird thing: low-poly ecosystem, cozy island, floating pocket world, golden hour, drag-and-drop to explore, 3JS, TypeScript , VIT. Like, what the hell? Like, what is this? And that's what I see, but it looks beautiful. Here, look at this. There's some, well, some kind of thing on the island here. What is this , how is it, what is this? Berth. And the house is beautiful, and, like, it looks real, but the birds fly somehow strangely. I think birds fly best in Codeex. But the thing is, there are these strange things here. And that's what I don't like about version 5.4. And I think that was the case with earlier ones, like 3.7, where it became too overzealous—they loosened the restrictions and it was a bit too much, it suddenly did all sorts of things that you didn't want. And I see that OpenAI has similar problems now, where it doesn't know that it does n't know something, or it becomes too overzealous, which I think is the flip side of it being more enjoyable to talk to. So, there are a few more examples, but it just issued a prompt to the user interface. Clearly, I don't want to see this in the front end, and I don't want to see it done in TypeScript and 3JS. Like, I had the exact same situation. I also conducted a similar test. Um, I'll share my screen. Um I asked him to redesign the landing page for Proof. So, Proof is an agent-oriented collaborative Markdown editor that I'm building and that we all use internally. Proofread editor.ai. You can just try it out , well, for free. This is, like, a cool thing. It really spread inside, but I had this page and I just asked 5.3: can you redo it? And this is actually, in my opinion, a very successful first attempt at creating a landing page in one go. It just looks good, and he redid, you know, the main image and all that stuff . Mine, uh, but he did the exact same thing. I don't have a screenshot, but he kind of merged the request into one of the, uh, he did some (I'll see if I can show it)—he did some, uh, follow-up things like this, and he merged in there something like, "We want to make, uh, the home screen so that it displays this, this, XYZ" or something like that, and I said, "Dude, why is this here?" and yeah, it's great, it's a nice design or something, with the star and everything . It's really, it's a little strange, actually. And it's a little interesting that he also seemed a bit tough. I was trying to get him, uh, I was trying to get him to iterate the design, and he was having trouble going from that to three more steps forward. It has become quite fixated on what it has already created. I don't know if you guys noticed that too . I had the same problem, uh, when I was redesigning a screen in iOS and telling it, you know, this should be the text, but come up with it yourself, you know, and actually in the text itself it said that this text is something like, which is where I think the 5.3 and 5.4 comparison comes into play, right? Like, it's much bigger, how to say it, it has this, uh, you know, it takes and does things that you didn't ask it to. Yes. Exactly. One more thing, also, uh , this model, I asked it to create an e-commerce site during the test, and it added a GDPR compliance checkbox to accept all cookies. This is so funny. Which is funny because I definitely don't want that in my design, but it's trained on that. So I feel like they're trying to loosen the restrictions a little bit, but now they suddenly see where the gaps are appearing, which I think is great, because then they can apply reinforcement learning or something to make those gaps disappear again. I think that complaining about these things is actually a good thing, and it should happen, because the next version will be just as enjoyable to work with, but it won't have these minor issues, which is an improvement in itself. And that's how fabric is created: they just did something and it did n't work. And, you know, at the rate at which OpenAI is releasing things right now, it's just crazy. If in two weeks or a month we have a new model that just fixes this, then that would be great. I'm looking forward to 5.4 CEX. This will be good. That's good because Yes. Oh yes. I like what they do. They first released the codex series, which is, you know, coding-focused, and then they learn from that and put that intelligence into 5.4. So that's really good. Yes. But I must say that 5.4 Pro writes better in Ruby than 5.3 codex. I know, I can show you an example for comparison. Uh, I know Naveen is very knowledgeable about Swift. Uh, if you can open my screen, I can show you examples. Uh, I know it works really well with Swift. It doesn't handle Ruby very well. So, in the middle you have 5.3 codex, and on the left—Opus. So you see he does a lot of things he shouldn't do. Um, am I still here? You're still here. Yes. Oh, I'm just full screen. That's it. Class. Aha. So you see, this is a more Ruby-esque way of doing it. That's exactly how it is. And you see Codeex doing all of these things. Hey, maybe it's not defined. And if it's not defined, it's like a TypeScript anxiety or something. I don't know, it's like it's from a world that shouldn't be like this, it's not how you write Ruby code. But you see, 5.4 is better. Here he does a strange thing, and here—something more like Ruby. So I can already see that 5.4 is getting closer to Opus; so, even as a model not intended for Ruby coding, 5.4 is better than 5.3 Codeex. However, I think that for Swift, version 5.3 Codeex is still a bit more thorough and better. I tried that too. But Naveen, what are your conclusions on this matter? No, that's exactly my take on it, right? Codeex seems to handle all edge cases. If you look, maybe in Ruby I can clearly recognize the first option, Opus, it looks like Ruby, you know. But two and three—I wouldn't say, looking at this code, that it was written by a junior, because he got too clever with these things, because Ruby and Rails take care of most of this, right? Conventions above all else. Yes, and it's not strongly typed either, Ruby uses duck typing, so you don't specify types. Like, you just assume everything works. And they're clearly worried, "Oh, what if it's not defined?" Well, it's a different style than the one adopted in coding. But it works great in TypeScript , because yes, it's even a typed language, yes. The cases actually make sense where the processing allows you to not " fall" and build the system, you know, compilation; in Ruby you don't need to worry about this . So, 5.4 is moving in that direction. So, we already have almost 2,000 people on the stream. So I want to take a moment to remind you of a few very important things. First of all, drink enough water. If you want to be on the cutting edge of AI, it's very important to drink plenty of water. You want to have liquid cooling, and your agent should have it too. Another thing is that we're constantly doing these vibe checks, and it's really important to us to make sure that we're giving you...you...you hear me? OK. It's really important to us to bring you reviews of the best products, and one of the best products, besides the GMT 5.4, that we recently reviewed is this lettuce-wrapped burger from Five Guys. So, this is actually my breakfast. This is a great product. It has lettuce leaves in it. Um, iceberg lettuce. Cool, refreshing, crunchy iceberg lettuce. It is literally dripping with beef fat. That's how you know it's good. We have double cutlets here. So this is a big deal. Great product. And really good tomatoes, cheese, and some onions. So I'll take a bite now . Um, this is so delicious. Where is ours? We should coordinate. Next time we need to coordinate better. No. Yes. You should always have a burger ready for these checks. But beyond that, we're the only subscription you need to stay on the cutting edge of AI . Um, when something new comes out , other than eating burgers, we do these tests of new models. So, that's our review of deepd 5.4. She just came out. This is amazing. We tested it for about a week. If you want to know what the new model is good for, you should look at each... it's a super detailed guide to... when should we use it? We use it every day for our work tasks to create products, design, write, do all these things. We give you our top findings, what we like, what we don't like, and how it compares. Evan Oman asks, "What's going on?" You really should join the burger discussion. If you don't understand the burger debate, do some research on the issue. Do your research before asking. A very important event is happening in Burgerland. We think it has the energy of the main model in open claw, which is pretty cool. I moved my claw to her. She is very fast. She is very intelligent. She is very good at keeping up with the conversation. There are certain drawbacks. He has outright lied to me several times, so you have to be careful with him. But it's worth checking out OpenClaw, we did a comparison test and did some interesting work to figure out the specific differences between 5.4 and 4.6, as well as 5.4 and 5.3. And we'll talk about this later . You should really check this out on Every.05 , OpenAI is back. By subscribing to Every, you get much more than just that: ideas, apps, and training. As for ideas—there are similar things out there, but we also put together this fantastic, I would even say, wonderful guide to OpenClaw that we published on Monday. So if you're thinking about OpenClaw and want to get started with it , check out our guide. Um, and it's great because Peter retweeted it. So , he retweeted Peter. Yes . So get some approval from Peter. Look at this. But in addition to ideas, you get applications. So, you have some really great apps. One of them is called Kora. This is an app created by Kiran. Kieran, it would be interesting to do a little preview of what you've been working on for the last couple of weeks, if you do n't mind. So you have Kora, you have Monologue, created by Navin, it's a speech-to-text app. He is fantastic. I do all my work in it. And we also have a few other apps: Spiral and Sparkle. You get it all with one subscription, pay one price and get it all. We have a bunch of new things. Oh no , Naveen, we have to seize the initiative. You do it. So, Oh no. What? What happened? Uh, I think the burger is back. The burger is back. We panicked. Then you left us. I felt like I was struggling. I said, " Intercept, please." And I was turned off. I'm back. I'm back. Sometimes the energy goes off scale and throws me off the stream. But ideas, apps, and learning— check out Every. Kieran, do you want to show us what you built? I assume you've been building things for Kora since 5.4. Do you want to show something you've been working on and tell us how it was integrated into your workflow? Yes. I'm just loading. So, let's see. It's here. Boom. Good. Screen demonstration. And while he's at it, I think there are a few bugs that you could show us that you fixed in X high, because I think that's really interesting. So, first let me show you another benchmark, and then I'll show you Kora. But really, I'm trying, or I can do Corora too. So if you share my screen, people can see it. I'm rebuilding Kora. Or I'm adding something to Kora, namely incoming messages. How I do it—you can go here: baby.cor.computer if you want to check it out, but I'm just creating a prototype of what I would like to use, it's just a prototype. It's not a design or anything like that. I'm just trying out new workflows. But the cool thing is that it doesn't look like a wireframe model, and I just realized: why make it wireframe when you can make it more designer? So I'm exploring streams, and one of the things I've been working on this week is a focus mode where you view your inbox in a Tinder-style way, and it's also going to be an iOS-style mobile app. Think of it this way: “ put this email away,” “archive that one,” “ open this one.” Delegate it. Hey, can you do this, and this, and this based on monologue? Oh, and send a letter, if so and so, and find out, delegate , start working. Delegated to my Claw or Kora, and you see, it's cool. I'm kind of creating prototypes, experimenting. So, um...Why, yeah, how does that fit in? Well, design—it's one of those things that's very underrated in LLM because engineers clearly use it a lot and designers are a little behind, but now they're catching up, engineers are suddenly starting to design more and the models are getting really good. So I always wonder, with each new model, how does it handle design? What is it? What is the aesthetics there? It's interesting to see how far you can go. I think all the models are very good at taking a design from Figma and applying it as if they know design, but where exactly, what is the aesthetic of these models? Early Claw models had a pink gradient, and you can see there's still a bit of purple here. So I have a Claude 3.5 Sonnet, a 3.5 Sonnet and a GPT-4o and it's just interesting to see the difference, it's cool. This is Opus. It's more minimalist, a little boring, I would say, but structured. This is the Gemini 1.5 Pro, it's bolder, has a better design, and is also minimalist. Ahem. Very interesting. It's all the same request, but different models. This 3.5 Sonnet is less boring, it's more like a statement. Although this thing on the right is a little strange, there is a gradient and shadows here. And now let's move on to 4o. Not boring at all. Lots of shadows, lots of gradients. You can already see the 4o aesthetic. He understands design and makes things similar to what you found too, there are a lot of good elements here. Ahem. This is interesting because it's a completely different design than this one, for example, even though the request is the same. So. If I was told it was 4o and I didn't know, I would believe it, it really does look like Claude's work. Yes, exactly. And that's cool. I think all models are well designed and can be operated. You can see a little bit of what we asked for, it's not so minimalist, a little bit of "extra". I think the next version of GPT will probably be more restrained, they'll find a balance, because it's cool that the model is experimenting. In fact, I like it better than the Opus version, which, while flawless, is a bit boring. Is this ideal? No, not at all. It needs some refinement, but it has more character or is simply more interesting than Opus. So I think Google is still leading the way in design right now, which is interesting. I use Google for design, and Google is the fastest. If you compare Google and GPT-4o, they are both very fast. This whole process is usually twice as fast as Opus, and that counts for a lot . Fast models make work easier, so GPT-4o is very easy to interact with, being twice as fast as Opus. So I think that's also some ergonomics to pay attention to. I think it's really interesting, and we've honestly covered Google and Gemini a little bit, and I think you have the most experience with Gemini. Can you just tell us, for those who haven't used the new Gemini 3 or 3.1 yet, what your experience with it has been compared to 5.4 and maybe Open Source 6? Yes. Uh, I also signed up for the Pro plan last week because I reached my limits. I thought, "I want more," which is a sign that there's something to this. And, to be honest, I use all the models all the time, but for different reasons. I think Gemini and GPT are faster than Opus. Opus, for example, is quite slow. Uh, in my workflow, how I use it, very much so. I went back to Opus the other day for something and it immediately started compressing and I wanted to throw my laptop out the window. Yes. And just wait for this compression. Gemini does this in real time. It takes him about 1 second to compress. So you don't even notice it squeezing. I guess you haven't tried Gemini CLI yet. He is cool. They have subagents, they have skills. They have everything. You can set up a complex engineering workflow. Uh, it may not be as good at orchestration as Opus, but for "vibe coding " or creating layouts like I do with Baby Kora, it's all Gemini, Opus, and GPT mixed together as usual. I think Gemini's design flair is probably the most pronounced, but it's also interesting to ask about other models. For example, if I say to GPT, “Hey, look at this design, what element is missing?” What is the one thing that could make this whole experience better? and it produces a result, then I implement something, and then I ask another model, "Hey, can you refine this design?" I think for initial design Gemini is good for brainstorming, structuring, lots of fact-finding and user research like GPT which I used and then for iterating the design to constantly make it better and refine it. I think Opus is still the best at not going in the wrong direction, but just continuing to refine and improve things in the same vein. And the cool part is, just try them all. And I think if you hated Gemini CLI before, then yes, it's a CLI, but it's pretty good now. I think, uh, Codex is probably still better, it's probably Codex CLI. and Claude Code are still a little better. Hmm . but this is, yes, for a different use case. It's just interesting that they are so close to each other, and it makes me very happy. I want to do a quick blitz poll: what do you use every day to, uh, work with AI? So let's start with programming. I use probably 90% of the Codex desktop app and 5.4. with a little Opus for certain things, but it's somewhere around 90%, and a month and a half ago it was probably 90%Opus. So things have changed for me lately. Um, and it was Opus in the terminal, in Claude Code. Naveen, what about you? What is your main coding tool now, and what was it a month and a half ago? Mine, I would say, is 95% Codex 5.3 now, 5% Opus. I still have the $20 subscription, not that one. or 5.4. er , 5.4 recently. I've just been using 5.4 for everything, at least for the last couple of days. Uh, I wasn't sure about the 5.4 core before, but once I understood how it worked, how to work with it , I quickly got used to it and now I use the Codex app 95% of the time. I even recently shared in X that according to Monologue data, where we track how many words I've spoken into each individual app, Codex is currently in first place. I think I've said 40,000 words in the last 30 days. Wow, that's a lot. Yes, and I use Opus when I need a second opinion on something Codex has done, so I go to Opus and use the $20 plan there because I don't use it very often. Very interesting. Kiran. I, uh, it depends on the type of work, but if I'm writing in Ruby, I use Opus because it's just better at it. Like, it doesn't... Codex will do it, and it will work, but as you've seen, it will do strange things that are not exactly like what I would do. In TypeScript style. Yes. It's like it's just putting TypeScript into my Ruby, which is probably fine if you're doing some nonsense, but I'm building an app that thousands of people use, and I really don't want that . Um, but for my open source work, I really like GPT because, uh, if you give it a task, like, and say, "Hey, work on this," and usually my open source work is either partial functionality or bug fixes or, uh, TypeScript, uh, with the Compound Engineering workflow, it's really good because I can just say, "Hey, here's a problem that someone had, research it, do it," and, uh, it'll just do it and make it work. And style is not that important to me. I'm more interested in speed, error correction, and learning. So in a scenario where speed, lack of bugs and regressions are more important, I find it very effective. So for that, I use it 100%, I think. Um, and Gemini is for "vibe coding", uh , mockups, things, design, and, besides, Opus is for brainstorming. I now also like Gemini for brainstorming. But I also switch back and forth just to see how it goes. And 5.4—it's very interesting, we just recently got access to it, and it will take another week or two. Uh , just like it was with 5.3 when it came out. I liked it, but I really " got hooked" on 5.3 a week later, when I said that I actually use it 50/50 now, which is crazy, because a month ago I was n't using it at all . So yes , this is a big shift. What about, what about your Claw? For example, for me, um, uh, I want, I want to share my screen. For example, Claw really changed the way I work a lot, at least, and I think it changed you guys too, but it might be a little different. For example, I have Claw R2C2 in Proof, uh, Proof—it's our uh collaborative markdown editor for people and agents. This is proofeditor.ai. You should take a look at it. Um, completely free to use. It's true we haven't seen your screen yet especially with Claw. Um, yes, this is Proof. So, for example, one of the things that happens is how Naveen published Unproven Pill. Here's the PR, and I just said, like, " RTC2, can you check this out and do a review?" Um, and he did a bunch of review work, and then 10 more screenshots , and it's a bit messy. This should have been run in a separate thread, but I think I've mostly fixed that. Naveen, did you take another approach to this? No. What actually happened is this : he left comments, and I have another version 5.4 running on my local codeex app, and I say, "Oh, look at the comments R2C2 left and fix it." So at the same time, R2 is working on leaving comments and trying to fix it, and Codeex on my side has already fixed everything. So, in one of the messages, he says, "Oh, someone else is doing my work" or something like that, and you have to respond to that, and then "I'll upload the latest changes." So this is actually pretty cool. It's really nice, especially because I create it in my free time. So the ability to say, "Okay, Brandon asked for something to be done," and I told R2C2 to just follow the instructions above and add it to my to-do list. And I also have a document with a list of tasks that I need to find a link to. I don't know if I should paste the link here, but let's take a look. Uh , where's my to- do list for the day? You see, I asked him to remind me to call my mom. So, now anyone can add items to your to-dos. Yes. Well, we...I'll deal with it. I think that's normal. There's nothing here that... Okay. If you want to plan something with Dan, his dinners are only 100,000. Uh , add your name to the list here. So, this is my to-do list document. And I can just say, " Hey, R2, can you add this to my to-do list?" and that's it, he'll come in and add it here. And this is what I use to plan my day. It also adds links to everything I need so I know exactly where to go. And I can also put it in Codeex and say, "Take this and do it," and then cross it off the to-do list, and that's really convenient. I just added a pretty cool... Oh, I want to show you this. So, I just added a pretty cool presence indicator. So I can say R2 uh, come into my document, and he'll come in, and we can communicate in the document. But what's really cool is... hopefully he'll do it pretty quickly. This, it appears right here. There's a bug now that makes there two of them, but it shows up here with this pretty cool face. which was developed by Lucas. Do you see that? Yes. I like it. Lucas designed it yesterday, and I just asked Codex 54 to implement it. What is this? It did the same thing . Actually, this is something to consider in version 54. I asked him to implement this, let me find where. I thought, “Oh, that’s cool. The implementation is excellent." This is the original that Lucas sent. So I sent Lucas the screenshots. And Lucas is a designer. I did n't notice that at all. I sent him screenshots. I said, " This is incredible." Look". And Lucas replied: "He messed up his noses." So what he did was take the images that I sent. I sent both PNG and SVG, and he just redrawn them without telling me and messed up the noses. I didn't notice because I'm not a designer, but Lucas did. So it's worth being a little careful with this model. She does strange things, I do n't even know why. Probably because they gave her a little more freedom. Yes, they gave her freedom , which we had been asking for for a long time. So, I am Yes. They did what we asked, now she acts more freely, we learn what we don't like, and I think that's exactly what needed to be done . Yes. Yes. Agree. Yes. And I would just like to show a comparison with Codex 5.3. I think a lot of people are using it now, and how 5.4 is a little different. I think Dan has already talked a little about PR Proof. I would just like to Yes, show. Yes. I know nothing about the Proof codebase. I just asked 5.4 to implement this pull request. The thing is, I use Proof a lot , and recently this flowchart was added, and for me it's unreadable in Proof. So I want to add a mermaid rendering. So, this is what I added, but it was never approved because they thought I was too sloppy. Thank you, Dad. You know, the good thing is that I use an external package, I rely on, I think, this wonderful mermaid made by the developers of Craft app. So, yes, a wonderful mermaid. This is cool. And here's what happened: if you look at the code structure, it did n't touch anything inside index.ts. He just added my chart plugin and all the code is inside that, right? It's very well organized and I like it. And thanks to that kind of organization, I'm not sure if Codeex caught these bugs, but I think Dan's R2C2 caught the bugs and edge cases, which is great. And then he fixed it , again, version 5.4. But if we look at Codex 5.3, the code is not very good. Like, everything in one index. But it's the same request. I gave both of them the same tasks for the same person . Which is not good, because it's just one big pile. What happened was this: because of readability, no one, like the Codex reviewer, noticed it. That's all good. But when I asked to compare both pull requests, it seems that the same errors reported here are still present. So you see, if your code is well organized, you know, reviewers , agent checks can actually see these cases and fix them. But now that everything is in one big file, the reviewers didn't notice it. It goes into release and we get these errors. So yes, I found a similar thing in my tests, where it was more targeted towards large documents initially. I think the best model is Opus in file splitting, and I think Codex generally leans more towards very long files. But I think 5.4 is better than 5.3 and moves more towards what Opus already does. This is perhaps a bit of a philosophical question, but do you think module files are still important in the AI ​​world, especially with models with a lot of context? 100%. If not, I think you can clearly see it in this example, right? Where are these the same reviewer agents? In one place he found errors, he fixed them, we fixed them, in another place he didn't, and there's still the same error. So having this kind of modular code really helps keep the codebase consistent and clean. I understand you think it's better for understanding if that's the case. Yeah, and everyone has this conspiracy theory that as soon as a new model comes out , a week later they do something and the model degrades or something. But this isn't exactly a conspiracy theory; this means your code becomes bloated, and while the intelligence of the model remains the same, your code becomes bad. This...I...I never really feel this change. That is, if people say that the quality has dropped, I have never, maybe only once when there was a real deterioration and Anthropic confirmed routing problems, but I have never experienced it. I...and yes, that's a very interesting question, and I think my answer would be : maybe in the future it won't matter anymore, but today it's still important. Or because we don't want to pay $100 for code review if we can pay a dollar. So why don't we make the code human-readable? We gain nothing by making it incomprehensible to man. And since everything is trained on high -quality code with best practices, I think the model understands human style better. Although we don't write bytecode, and perhaps we should , the model is probably not very good at it because there is less training data for it. So I really think that's probably true. And perhaps most importantly, the most difficult thing about using and implementing AI is feeling and trusting the system. It's much harder for the world to realize that we need to trust AI, and I think we should work to make the code more readable for humans, because that makes it easier to trust AI , and therefore we can implement it faster. So if we make it harder to trust AI because of weird, long code that works, I don't think that's good for AI adoption, and it's actually going to slow down progress. So I think even for that reason, even if it doesn't technically matter, for us as humans it should matter because we want to get more people involved in this story. So yes, I think it's important. What do you say, Denis ? I think I like asking these kinds of questions because it's always helpful to just ask: okay, the world is new, what could happen? I can assume that over time the programming style will change as we look less at the code itself and more at plans, reviews, and all that sort of thing . And, for example, about Proof, I have no idea. I have never, ever opened the directory and looked at it. I was looking at it. Didn't look at it. We should do it. Actually, I think it works great. That's the point, right? After all, it works. I like using it. So it doesn't matter how interesting it is, it's not like that . Yes. Yes. But Naveen and I run software that people pay for, and I think this step is still too serious for us. I understand, I honestly believe that we can't do this because if something goes wrong, everything will break. Yes. I had no idea. Like: "Oh, well, CMM said something, and sorry, everything has been deleted." Like Yes. Yes. Yes. Yes. But about Yes. What was wrong with him? What would you say if you looked at this and thought, "Wow, this is just crazy." No, he did it. Maybe that's great. No. The thing is, I don't want to do a code review right now. We should do a code review sometime, because I think this is a very interesting question. Yes. Okay. Yes. At least for me, when I looked at the product— it works really well , you know, but the code seemed a little disjointed. I was trying to find where this particular thing is, and it's all a bit scattered, maybe. Maybe because of how you started with the Mac app and then ported everything over, yeah. Yes, that's interesting. Yes, I'm currently forking out a Mac app, and someone asked about it. Code Rushi asked about this. I tried Proof a while ago, but it was adding span tags to the markdown. Now everything is different, because we have moved from a native application to a web application. So this should be fixed by now. You should check this out. proofed.ai. But yeah, it's really interesting. And I think we've reached a point where I think there have been a few times in the last week where Proof has been down. And of course, it's not a giant production app. You have significantly more users than Proof, but it's still important, especially for an early product where I'm trying to get the team to use it. And I really felt that even though I didn't fully understand how everything worked, I could trust 5.4 enough to figure it out . And I did it, and now I have tests and all that stuff . It looks like Kora has had some infrastructure issues over the last day, and as with Nive, you've been having issues for the past few weeks. and, uh, I don't feel the need to go into detail either, I'm sure you know how to figure this out. Um , I don't think it can solve, for example, that bug in Kora that you had, it was very serious. I don't think she would fix it, but at least now she's good enough to handle things like this, which is new and unusual. This did n't happen two months ago . Yes, I agree. I think it's funny about that kernel bug, because I was using 5.4 to get deeper information, and it was completely wrong. She was completely wrong. Yes. She just didn't get the point at all, and we didn't understand her for a while either . It was at a very low level, very difficult. Ultimately, it was due to a database dump due to a weird thing in Postgres where we were approaching the storage capacity limit and everything crashed. We asked "why, why, why?" and she couldn't figure it out. We needed people to do it. So I don't trust her in such matters yet. But, for example, regarding baby Kora — I have no idea what the code is. Baby Kora is where I prototype the new Kora. And most likely , if baby Kora works and looks good, I will tell Kora, "Look, this is how it's supposed to work." Refactor and move it. Do it. And, probably, I will do so. I won't look at the code, but it will be implemented into the product. So I think there 's a kind of preliminary stage, maybe for Proof, where you say, "Here's a version of Proof that I like, now let me give it to GPT and say, rewrite it properly because Naveen gave this feedback." We are deleting this. We update, and suddenly we have Proof, which is well structured, has tools and everything we need. the principles of rewriting with where it is now, how it works, and maybe it will work well. Yes. And I'd like to ask, have you seen OpenAI's recently released blog post about "harness engineering"? which is actually really nice because I think they just let GPT do 100 % of the encoding without human intervention, they just gave the request. And I think Proof can be that same example, where every day there's proper documentation , proper code structure, everything, and it just does the testing from start to finish. Yes. Soon to be added to compound engineering, which is also inspired by harness engineering. I find it really interesting because I really think if you have this setup with this feedback loop , and the "compound" loop , like, "Hey, I know what works and what doesn't," maybe from other projects, and also design systems, like you take design systems from other programs that we have, and things like that. Oh, yes, I think Proof is how software will be built in the future, not all today, but very soon. Yes, I agree. And I'm so jealous of Dan and Proof because I can't do that with the Mac and iOS apps. This is very bad. I can't set up a full feedback loop, although I've tried to do it, but it still fails when I'm working on something, at least I was working on Proof yesterday and it's so good. It's a nice feeling because I didn't touch a single line of code. I just said what feature I wanted to implement. I just said it, he used Playwright and gave me screenshots. This whole experience is truly wonderful. It's the same with my design, now I work with designers, and we're just "on the same wavelength" with design, we have the same feelings. I don't touch CSS or anything like that. We are simply talking about experience and feeling. That's all. That's all we do. and we just transfer my mockups into Figma. They design, and it's really cool. This is great. This is great. And I love that you're jealous, Naveen, because actually the product strategy for Proof was very inspired by you, because with Monologue you have this thing where you say, "I want to make something extremely simple but extremely hard to do well ." where all the details matter, but he just does one thing really well, and I want to do all the details right. And I think as a product developer, I'm always tempted, especially with AI, to say: we can do this, and this, and this, and it starts to grow exponentially . So from the very beginning of Proof, I said to myself: I'm just going to make the best markdown editor in the world, a kind of online editor with real- time collaboration support. I started with a Mac app, but when I focused on the main thing, I realized that I would only do that. And I think it worked really well because I resisted the temptation to add a bunch of complexity. To be honest, creating a truly high-quality collaborative Markdown editor is quite difficult. Um, so you inspired me, and I'm glad you're a little jealous. That's nice to hear. Maybe we could talk about technology. I think you should just switch to Elixir. Elixir is very well suited for this kind of real-time stuff, and really. Also, Codeex seems to be a coding agent. Recently, there was a study by one university: what language do coding agents do best? And Elixir is the best language, I think, 96% or so , and the rest are somewhere around 80%, 75%. So maybe you could spend some time on this and migrate the platform to Elixir. I'm all for it, but it sounds like...I don't know Elixir at all, it seems... I don't know anything either. I know a little. Phoenix. I heard, yes, Phoenix, I heard it's very good, encoding agents work well with it because there's built-in documentation. It's tightly coupled to the code, so inline comments are all tightly integrated, and coding agents handle that very well. It's interesting, so, well , if I have some time and I can't make it work, I'll try, I'll try. But I think the problem Dan got into is that he uses YJS. I mean, it's not a bad library, but it's very complex on its own. It's a very difficult task— synchronizing things, integrating them, providing offline mode, online mode, and all that. And, you know, it seems like it's necessary for collaboration, but at the same time it's possible and not. I don't know if there is any library like YJS for Elixir, or Oh yeah, there are libraries for everything. Yes, I implemented a collaboration feature on Ruby on Rails using the YJS adapter and it works great. So there are adapters for Elixir and for everything else . So, you can replace any backend. But the thing is, you or the AI chose a technology that is very complex in itself. So the question is , do you need it now? Is this really the problem you are solving? Is this collaboration—it's cool, but is it a real problem, is it rendering, or something else? Yes. What exactly are you actually trying to solve ? Yes. What do you think about this? Because I have a lot of opinions, but it's always more interesting to hear from you as users. What exactly do you want? What do you think is the main problem? For me, I think it does n't have to be in real time. It could be like Notion, right? Notion document. How does Notion work? Notion also uses YJS. Yes. Yes. I personally feel Notion, but where you can easily share. Yes. Notion can't just...I know, maybe they're moving in that direction where they can. I want to have a Notion skill where you can just ask within your coding agents, "Oh, create a document," but Notion is already too complicated, I don't want to. Yes, it might actually replace that for me. Yeah, for me it's not even... The biggest thing is that the agent can create a document, edit it, and share the link. For me, this is the main problem. Because he can create a gist or a document, or upload it somewhere, but editing and ensuring that it's the latest version—that's important. I don't necessarily care about real time. Ahem. I think so, but I still want it to look good. I want it to feel good, be responsive, and look like a native experience. I think this is the most important thing for me right now. It slows me down a bit sometimes, and because of the synchronization...that's the only thing holding me back . I like that I can tell my agent, "Do this, read this"—and that's, I think, the most important aspect of an AI agent. I think this is a strong point, and the fact that it works without installing anything is another . This is very, very, very good. So, I think that's the problem it solves for me . Yes. Yes. And this real-time work is cool, but I know from the experience of other companies and people that it can quickly go sideways. Yes. Yes. Yes. And what I really like to do in Proof is I just review the document, leave comments, and then go back to the agent and say, "Oh, I left comments on this document." Yes. Update it. And it works, and to be honest , I've never used it in real time, at least not in the last few days. Yes. Similarly. And I think I'm doing the same thing with Agentation. I don't know if you know, but this is a tool that you can add to your own site, leave comments on the design, and then tell the agent, "Hey, I left comments, apply them." I think Proof is something like that version. Where you can collaborate directly, share, people leave comments, and then you can apply them. I think, honestly, giving up YJS can actually open up a lot more opportunities to focus on more important things. This is interesting. We'll have to think about it. But this is a serious step . This looks like Dan's intervention. Like, "No, my agent can do that." This... The models are improving fast enough to make YJS work without having to port the platform or remove it. But we'll see. OK. There are almost 5,000 of us here. So I want to take this moment to welcome you if you've just joined. We are the only subscription you need to stay on the cutting edge of AI . GMD 5.4 has just been released. We have a vibe check on our every website . Just go to every.to and you'll see it along with our broadcast right now. Headline: OpenAI is back. Honestly, since GPT-5, we've felt like they've been a bit behind in the agent development race. Let's try to find our atmosphere check for GPT-5. We usually receive these models before they are released and do a test on the first day. This is our atmospheric test for GPT-5. And you can see what we said then: if you're a senior engineer working with agents or an atmosphere testing specialist. This was almost August last year. So that was almost a year ago. If you are a senior engineer working with agents or an experienced programmer, GD5 will not be your everyday tool for breakthrough coding . He is more of a skilled all- rounder. Um, the discipline of programming has changed radically this summer. Benchmarks don't show it, but if you can run four agents at once and use quad-code, GP5 seems like a step backwards. I think that's how we felt, certainly, at the time, and we continued to feel that way for a while. I think Naveen, you were more immersed in the code than we were for a few months, but it really took them a while to get out of the, um, programming tool mode, where AI is more of a pair programming rather than a delegation tool, and Cloud Code really knew how to do that. And I, I feel like our main takeaway here is that Open really gets it, especially given the trajectory of 5.3 Codex, um, Spark and the Codex desktop app plus GP 5.4. They used to move very quickly, and they started to catch up so much that it became a legitimate option as a daily tool for us in all our programming tasks. Um, and that's what we, even our, that's my daily choice, for example, in my Claw. So if you use Claw, I highly recommend giving it a try. There are downsides, um, because he's more like Opus now, he's a little more impulsive, and, um, so he lies sometimes. So you should be careful with him in matters where a lot is at stake . But a few things we discovered during our review: I think it's best in class in planning. We haven't talked about it yet, but I really like the plans he's making. It is also very technical. The search is also very good. Yes. Yes. This is really good. It's very technical, very detailed, but it's, it's very human. And if you follow his train of thought, it's actually surprisingly good, surprisingly intelligent, and he expresses himself in a way that feels human. This, this is pretty crazy. Um, so very good, very good for planning, very good for reviews. He is very fast. Um, we felt like he had an expansion of the task boundaries. Um, sometimes he built too much. I think you definitely noticed that, Kieran . Um, and, and he definitely lied in Open Claw. So, you should be careful with this. But overall, it's a really great model. You should definitely try it. Um, we do this all the time, every time we show you this review. We also showed you our review of the GP5. Um, we also published a really great guide to OpenClaw on Monday. So, if you are interested in OpenClaw, you should definitely check it out. It's free for everyone. Every "claw" school. This is a comprehensive beginner's guide written by me, our platform manager Willy, and our experts. This is truly a wonderful thing. You can just add this to your workflow and get started. Plus, because an Every subscription is all you need to be on the cutting edge of AI, we offer insights, apps, and training. We just showed you the ideas. We also make apps. We have an app called Kora , which was developed by Kiran. This is an AI assistant for email. We also have…this is what my mail looks like. This is simply incredible. We also have Monologue, which Naveen created, which is a smart dictation app. I do almost all of my work using Monologue. It's like Super Whisper and Whisper Flow. You should definitely try it. We have many other applications. We also have Spiral, which is a ghostwriting agency. We have the Proof that we just showed you. And you get all this for one price. You pay once and get access to ideas, apps, and training. We also conduct training, organize camps where we teach how to use OpenClaw or the Codeex application. We do this at least twice a month. So be sure to check it out. Every.TO. It is also very important to remember to stay hydrated. If you're on the cutting edge of AI, you need to drink water. And you should know that one of the best products out there is not a model . This is a burger. This is a burger from Five Guys. It's been sitting on my desk for about an hour. And I take a little bite out of it. And this is simply the most outstanding product. Even though it had been sitting for an hour, for example, the lettuce leaves were previously crisp and now they were a little wilted. But you know, I love it. And I'm going to make a big bite for you right now. Bon appetit. This is so delicious. Much better than Burger King or McDonald's. And so if you want to be on the cutting edge of AI, you need to eat Five Guys burgers, drink water, and subscribe to Every here. And Naveen, do you want to add anything? Eat burgers, drink water. And I'm really excited because the changes that OpenAI has gone through, they seem to have accelerated, and they seem to be moving really, really fast now. So yes, I'm also really looking forward to 5.2 . But honestly, I'm even more excited for Codex 5 or 4, or whatever comes next, because you can see the trajectory, and they're moving so fast, it makes me really happy. One hundred percent . I'm really excited and, you know, already a little bit hooked on using these Codex models, so I'm just excited to use these models to make Monologue even better. I used to get stuck on multiple problems all the time. And now I happily take on these tasks, and it makes it much easier for me to create new products and ideas. Exciting. Yes. So, that's all for today, friends. Thank you very much for joining . Thank you, Kiran. Thank you, Naveen. If you haven't seen it yet, check out version 5.4. Rate our vibe check. Come to every release. See you next time. As soon as the new model comes out , we'll be here. We'll have a great vibe check. We'll do a live broadcast, like today. Remember, drink more water, friends. See you later.