A livestream "vibe check" hosted by Every (Dan Shipper, Kieran Klaassen, and Mike Taylor) evaluating and reviewing a newly released AI model. # Channel Presentation Observations ## Observable Channel Identity: **Every** ### Self-Description Evidence (Explicit Claims) - **Organizational name and hierarchy**: The host states, *"I am the boss. Every editor"* and guests are introduced as staff: *"Kieran Classen, extraordinary composite engineer who invented compound engineering"*, *"Mike Taylor, our new the head of Vivaldi"*, *"Natasha, senior engineer from applied artificial intelligence"*, *"Tyler — designer at Every"*, and *"Dan Schipper"* as the *"fearless leader"*. - **Publication model**: The channel packages model reviews into newsletters: *"Vibe Check will be released soon. He will arrive soon your postal code boxes... all our screenshots, all our tests... into one whole the package."* - **Subscription positioning**: Closing statement: *"Everyday is the only one subscription that you needed to stay on the edge of AI."* - **Proprietary products**: A benchmarking platform is presented: *"checks.every.to/p/dance-editorial-checks"*, described as *"we begin transform our testing for benchmarks... tests based on our daily work."* ### Intended Audience (Demonstrated) - **Practitioners and AI power users**, not general consumers. The discussion assumes familiarity with model names, pricing tiers, effort levels, and tooling (Cursor, Codex, Claude Code, MCP). - Evidence: technical granularity like *"effort should be to consider as your willingness to spend... budget for the problem"* and workflows such as *"Fable 5.1 low... as an orchestrator and Grok 4.6 like all subagents."* - **Developers and designers specifically**: segments on Ruby coding benchmarks, 3D rendering comparisons, and design tools (*"I really want to learn... Origami Studio"*). ### Claimed Scope - **Timely, multi-vendor AI model evaluation**: format framed as *"our latest Vibe Check. We are doing this, when, uh, new models come from large laboratories"* — covering Anthropic, OpenAI, xAI, and Google products in one stream. - **Work-based benchmarking as differentiator**: *"when you see new model and you see the test result... that's not really tells you whether this model is suitable for my work? So we started creating tests based on our daily work."* - **Direct lab access**: an Anthropic employee (*Tariq*) joins as a guest and answers questions, demonstrating claimed proximity to model developers. ### Distinguishing Claims vs. Demonstration - **Claims**: "only one subscription... to stay on the edge of AI"; models *"will be remembered"* (guest's claim about the lab, not the channel). - **Demonstrated**: live, unscripted multi-staff format; a color-coded verdict system (*"green gold... star?"*); public check results at a branded URL; heavy emphasis on subjective preference (*"I felt to be more comfortable subjective"*). ### Notes - Transcript appears machine-translated; quoted wording reflects that. - The audience-facing tone is casual, insider-oriented, and community-interactive (*"we just received a question from the audience"*). # Delivery Observations — "Live: NEW MODEL VIBE CHECK" > **Caveat on evidence quality:** This transcript appears to be a rough (likely machine-generated/translated) rendering of a live stream. Some phrasing is garbled, so quotations are reproduced verbatim and inferences are kept conservative. All quoted text is exact from the transcript as supplied. --- ## 1. Overall Format and Teaching Habits - **Panel-plus-guests structure.** The host introduces team members by name and role early ("Kieran... extraordinary composite engineer who invented compound engineering"), then brings in successive guests (Tariq, Natasha, Tyler, Dan Schipper) in a rotating-segment pattern. - **Show-then-explain habit.** Speakers repeatedly promise demonstrations before delivering analysis: *"I'll show you something"*; *"Maybe it will be easier, if I show you that I mean."* - **Live-tech friction handled in stride.** Multiple screen-share and audio issues are narrated openly: *"Sometimes it's not works at Dia"*; *"Can you hear me? Wait, how's mine? microphone?"* - **Verdict-first framing.** The show's stated method is a color rating: *"how do you Do you evaluate? She is golden. star? Is it green? Is it yellow or red?"* — repeated by several speakers. ## 2. Recurring Vocabulary and Phrases - **"Daily driver"** — the most persistent recurring term, used by at least four speakers: > *"I use, uh, OpenAI models... as your main daily driver"*; *"Opus 5.5 now my daily driver that is simple madness."* - **Color/verdict shorthand** ("red," "green," "gold"): *"I rated it at red, and sometimes on green."* - **"Workhorse"** for reliable mid-tier models: *"Luna is a worker horse, Sol—working horse, Opus 5.5—working horse. There are so many workhorses."* - **Effort-level talk** dominates the second half: *"low," "medium," "high," "extremely high"* levels recur constantly, e.g. *"average mileage for coding"* and *"low more like sketch."* - **Enthusiasm markers:** *"It's just so funny"*, *"madness"*, *"Wow, that's enough incredible"*, *"jaw drop"* moments. ## 3. Tone and Register - **Casual, conversational, overlapping speech.** Frequent fillers, restarts, and interjections typical of live unscripted talk: *"Um, uh, yes, I"* patterns appear throughout. - **Humor through specifics.** Comedic reactions center on concrete, absurd details rather than general jokes — the sourdough sword, Karl Fog as a boss fight, the "$4200" rent stat: *"I am nothing. didn't do it. I swear By God. I just... sent a request from in one sentence."* - **Editorial honesty as a norm.** Speakers emphasize candor about disliked models: *"It was amazing, how much people like it not really I liked it... I hated him. He excellent."* ## 4. Communication Patterns Between Speakers - **Prompted handoffs.** The host consistently redirects with a named question: *"Mike, don't you want to to start with which one your verdict?"*; *"Kieran, you say that do you feel that this skill...?"* - **Guest-questioning loop.** Guest segments follow: welcome → personal experience question → demonstration → follow-up about workflow implications. This repeats for Tariq, Natasha, Tyler, and Dan. - **Audience integration.** At least one viewer question is read aloud: *"we just received a question from the audience where they said: 'Is improving most behavior on average or lower level thinking?'"* ## 5. Persuasion and Explanation Techniques - **Concrete before/after comparisons.** Speakers show paired examples across models: *"on the left Opus 5.5, on the right Soul 6"*; *"Opus 5 against 5.5... this difference of approximately 0.5 steps."* - **Analogies to human work.** Models are compared to colleagues and employees: *"it's similar to work with a colleague-creative"*; the boss-task analogy (*"your boss gives you a task that must be done by an hour, not 12 hours"*). - **Personal anecdote as evidence.** Claims are anchored in first-person usage: Tyler's freezer-cooled Mac in Hawaii; Mike's 2-hour blind writing test; Dan's editorial checks platform. - **Closing advice/recap.** The stream ends with actionable recommendations: *"Also check out Opus 5.5 on low efforts. I think this is It's worth a try."* ## 6. Stable Patterns vs. Isolated Moments **Stable (repeated across multiple speakers/segments):** - "Daily driver" framing; verdict/color ratings; effort-level taxonomy; show-your-screen demos; named handoffs; personal-workflow-based evaluation. **Isolated (single occurrence — low confidence as habits):** - The Karl Fog/sourdough-sword game riff (unique to Tyler's segment). - The security-classifier aside about OpenAI's recent tightening (mainly Tariq/Dan). - The basketball-player analogy for model sycophancy (one instance, from Tariq). - The Grok 4.7 DM mishap anecdote (*"he sent personal to me message... 'Marcus, I'm so sorry'"*) — told once. --- *Note: Because this transcript is garbled in places, fine-grained stylistic judgments (e.g., exact cadence of speech) should be treated as provisional; a cleaner transcript could sharpen observations about filler use and interruption patterns.* # Observation Record: "Live: NEW MODEL VIBE CHECK" — Claude Opus 5.5 (and related model) reactions ## Context A live "Vibe Check" stream hosted by **Cade Lee** (editor-in-chief at Every, substituting for Dan Schipper), reviewing newly released **Claude Opus 5.5** from Anthropic, with comparisons to **Fable 5/5.1**, **GPT-6 Soul/Astra**, **Grok 4.7**, and **Luna**. Participants: **Mike Taylor** (head of Every's evaluations/"Vivaldi"), **Kieran Classen** (composite engineer, creator of "compound engineering"), **Tariq** (Anthropic guest), **Natasha** (senior applied AI engineer), **Tyler** (designer at Every), and **Dan Schipper** (joining briefly). --- ## Verdict ratings of Opus 5.5 ### Mike Taylor - Rates Opus 5.5 **"green gold"** — nearly gold, but deducted because the model *"ruined a couple of tasks through bad time management."* Only one benchmark mark against it, in his view. - Describes it as **Fable 5.1-level in feel**, but cheaper and more token-efficient, and says it shed the off-putting quirks of Opus 5 that had driven him to OpenAI models as his daily driver. - Now splits usage **roughly 50/50** between Opus 5.5 and Codex, with the Codex half explained by inertia of existing threads, not preference. ### Cade Lee - **Formerly rated Opus 5 "red"** and disliked it; now calls Opus 5.5 his new **daily driver** — a shift he calls "madness," since Fable 5/5.1 had always been his daily driver regardless of cost. - Claims Opus 5.5 is **better than Fable 5.1 in certain areas** (effort-level tuning, 3D detail/spatial movement), and that it competes with top-tier models — something he says was never previously true for a Claude model overtaking Astra or Fable. - Highlights **effort levels** as the standout feature: low effort produces a "sketch" that leaves room for user input rather than over-filling; higher effort adds detail without overreaching. ### Kieran Classen - Ran coding tests using his **Compound Engineering** plugin; reports **Opus 5.5 beat GPT-6 Soul on complex Ruby coding problems** — notably at **medium effort**, not high. - Notes **effort levels across providers are not comparable**; Opus 5.5's low level is more usable than expected. - Finds Soul 6 faster but **"flaky"** in long autonomous runs — repeated restarts needed — so for his workflow Opus 5.5 is the more dependable "long-term workhorse," even acknowledging new OpenAI security classifiers as a possible cause. --- ## Effort levels (major discussion thread) - **Tariq (Anthropic)**: Framed effort as *willingness to spend budget on a problem*; Claude retains a **minimum baseline** — it always attempts to solve the problem and never abandons its approach. More effort adds early-stage deliberation and test-tuning. Suggested that in future, genuinely smart models could be **Pareto-dominant** (no verify-retry loop needed), making them cheaper overall. Notes effort switching no longer incurs cash penalties mid-session. - **Tariq's analogy**: a task given "12 hours" gets assumption-heavy output; a task given "1 hour" forces clarifying questions. - **Mike Taylor**: Confirms low-effort results are *"very impressive"* — previously he wouldn't have considered low effort viable. His only deduction: at **extremely high effort**, long runtimes may not add proportional value. - **Natasha**: Worked at **medium effort** throughout and never felt the need to raise it; didn't hit usage limits. - **Kieran**: Calling effort-level selection a **skill users still need to learn**; the team plans an effort-level guide. --- ## Tariq (Anthropic) — stated positions - Calls Opus 5.5 a model the team *"fell in love with internally"*; likens its significance to Sonnet 3.5 and Opus 4.5. Announces **Sonnet and Haiku 5.5 releases** are coming. - Personally finds workflows (previously token-prohibitive on Fable) practical on subscription plans with Opus 5.5; says Opus 5.5 is *"simply better than Fable 5.1 now"* on average, with possible exceptions in planning/safety. - On model personality: **"the way you communicate is an expression of your intelligence"** — more complex language does not mean more intelligence. Attributes improvements to accumulated user feedback over time. - Argues there are often **no trade-offs in intelligence work**: models can become *"cheaper, faster and smarter at the same time."* - On who benefits: **token-price-sensitive users** benefit most; cheaper proactive models change the cost-benefit math for tasks like code review, speculative work, and agentic tools (e.g., Claude Code-type products). Notes Opus 5.5 works well as a **research tool**. --- ## Natasha's evaluation - Tested the model's ability to **teach**, via a multi-part interconnected HTML "research explainer" skill. - Rankings from her testing: **Opus 5 did well → Fable improved clarity → Opus 5.5 was dramatically better**: polished, interactive, with embedded JavaScript games; she calls it the **best designer and best teacher** among tested models, attributing this to writing and frontend ability. - The task ran **~5 hours in one pass**, using **dozens of subagents** across **50+ files** with self-checking. ## Mike on time-to-completion - Opus 5.5 failed one of his benchmarks by **not finishing within a 10-minute timer** on an educational-materials task — he acknowledges this was an unfair constraint for his workflow, and the model was *"so persistently working on the educational material that it didn't finish the overall task."* He still rates the model very highly. --- ## Tyler (designer) — creative/ambitious tasks ### Genshin-Impact-style San Francisco game - Built from a **single one-sentence prompt** ("make a game like Genshin Impact set in San Francisco"); ran **~5.5 days** without finishing; music, voice, ~60 FPS. Satirical content (rent counter at ~$4,200, fighting Karl Fog with a sourdough-starter sword, a "Series Z" special move raising a $40M round). - Tyler insists he made **zero creative contributions** — *"I didn't do anything, I swear."* - Jaw-drop moments: the model **self-evaluated its own game** (1–10 rubric on visuals/content), caught issues Tyler never would, wrote its **own grass shader**, and declared *"I won't stop until I get at least 8 out of 10 overall."* - Comparative claim: **Astra never made his jaw drop** the way this did; *"I don't think it's the best model anymore."* ### Pokémon world simulator & other builds - A real-time Pokémon simulator (prompted once, ran **~1.5 days**, heated his Mac Studio so much he put freezer packs on it in Hawaii); generated a self-authored report on events in the simulation. - Built an **AI-friendly clone of Origami Studio** (node-based prototyping tool) over **~3 intermittent days with ~4 prompts**, intended as a learning tool to master the real Origami Studio; drives it via Claude Code terminal commands. Explicitly uncertain whether it becomes a lasting workflow tool — *"maybe it will be my design tool, maybe not"* — but hopes it will make him a better designer. - For **real design work**, he still uses **Grok 4.7** for cost-efficiency, but says *"if tokens were free, Opus 5.5, 100%."* --- ## Dan Schipper's segment - His new daily driver is **GPT-6 Soul** (cheaper than Astra, sufficient for writing/computer/general work, though below Fable 5.1's ceiling). Still uses Astra for highest-level knowledge tasks. - On Opus 5.5: he mostly tested at **high effort** and found responses slow (~15 min) and **long-winded, burying the lede**; the team told him **low effort is "the magic,"** which he hadn't tried. He advises checking the model *"with a certain proportion of skepticism."* - Presented **Every's "checks" platform** (personal benchmarks built from real daily work, with pass/fail checks). On his editorial-writing checks: **Astra first, GPT-6 Soul second, Opus 5.5 close behind**. - Critique: Opus 5.5 is *"very, very smart"* and a strong author, but gives **long, thick answers** that take too long to reach the point; he prefers minimalist, immediate prose where Soul and Astra are better. - Confirms recent **stricter safety classifiers in Codex/ChatGPT** degraded long autonomous runs (constant re-approval loops), giving Anthropic a temporary usability advantage — while noting Anthropic had extra time to refine its own classifier from the Fable rollout. ## Mike's PowerPoint/brand-style benchmark - Opus 5.5 **deviated from brand-guide instructions in ways he liked** (black instead of green, custom SVG style), which he compared to *"hiring a creative director"* — following instructions while knowing when to make exceptions. He initially thought the deviations were errors, then realized the style was preserved coherently. - Praised a **Hoboken restaurant map** as both accurate (matching places he knows) and visually superior to other models' outputs; noted old Opus 5's habit of building unnecessary custom SVG maps appears fixed. --- ## Grok 4.7 discussion - **Mike**: Found Grok 4.6 *"surprisingly good"*; **Grok 4.7 felt rushed** — in his benchmarks a **real regression**: many unfinished tasks, outright errors even on 3–4 retries, design quality a step back. Suggests retesting since behavior changed markedly within a week. - **Tyler**: Agreed the **first days were rough** (hallucinations, server downtime, one incident where the model sent a personal Slack message to "Marcus" unprompted after taking an instruction too literally), but it **improved over the week**; by the morning of the stream he found it more reliable than 4.6, though hard to evaluate because it's an incremental improvement. Still uses Grok for ~100% of real work and won't leave it. - **Kieran**: Grok 4.7's 3D work is unimpressive (more like Luna, which looks slightly better) and it struggled with the step-following Compound Engineering workflow; calls Grok 4.6 a **"workhorse,"** but notes the landscape now has *"so many workhorses that are better than the frontier models — it's madness."* --- ## Concluding positions - **Kieran**: The pace of releases makes models nearly **indistinguishable on capability**; selection is increasingly **subjective preference**. He uses all models and recommends trying Opus 5.5 at **low effort** rather than defaulting to high. Names **Opus 5.5 the winner of the week's launches**. - **Mike**: Echoes saturation of benchmarks; likes Allie's/Opus's sharper personality versus "smoother" OpenAI models, comparing it to working with a creative colleague rather than a corporate office. Reiterates 50/50 Claude/Codex split. - **Tyler**: Uses Grok 4.7 for real work, Astra-with-Blender for computer control, Cursor projects constantly — but calls Opus 5.5 *"adorable,"* jokes *"my wallet hurts,"* and says he can't express how good it is. - **Cade**: Announces the full written Vibe Check and checks results coming to subscribers; notes the editorial difficulty of Opus 5 (people could only say "I just didn't like it") versus the enthusiasm for 5.5. --- ## Tensions and qualifications preserved - **Mike's "green gold"** is explicitly conditional: the model lost gold due to time-management failures in his benchmarks. - **Dan remains skeptical** of Opus 5.5 based on high-effort usage; his positive view of low effort is secondhand (team advice), untested by him at recording time. - **Tyler's enthusiasm for Grok 4.7 coexists** with Mike's regression finding and Tyler's own admission that the early week was bad — resolution deferred ("maybe they'll fix it," "we need to retest"). - **Tariq's optimism** ("no compromises," Pareto-dominance) is forward-looking and framed as prediction, not settled fact. - **Mike's 50/50 split** is attributed partly to inertia (existing Codex threads), not purely model preference — he states this qualification himself. - **Tyler's Origami tool's long-term utility** is explicitly uncertain, by his own statement. # Video Attribution & Provenance Record **Video:** "Live: NEW MODEL VIBE CHECK" (id: `01a0f15a-4558-74d1-87d5-3679dfbfa59d`) ## Speaker & Role Inventory ### Every staff (hosts/panel) - **Kade Lee (transliterated; "Cade Lee," "Kate," "Kiana" also appear)** — Editor-in-chief ("boss of Every"), filling in for regular leader **Dan Schipper**. Hosts the stream, moderates, and asks guest questions. - **Kieran Classen (transliterated)** — "Extraordinary composite engineer" who "invented compound engineering," works on "Cora." Discusses coding/compound engineering benchmarking. - **Mike Taylor** — "New head of Vivaldi." Gives verdicts ("green gold"), runs writing blind tests, describes benchmark checks. - **Natasha ("Nietzsche"/"Nitesh" variants appear)** — Senior engineer on applied AI at Every, works with Mike on evaluations. Demonstrates "research explainer" HTML tests. - **Tyler** — Designer at Every. Demonstrates a Genshin-Impact-style San Francisco game, a Pokémon world simulator, and an AI-friendly Origami Studio clone built with Opus 5.5. - **Dan Schipper** — Every's usual "fearless leader" (absent at start, joins later). Shares GPT-6 Soul impressions and editorial checks platform. - **Katie** — Every tester; referenced only (absent from stream); ran writing tests; main diagnosis: "truly a great partner." - **Marcus** — Mentioned only, as recipient of an accidental message sent by Grok 4.7 during Tyler's testing. ### Guest - **Tariq** — External guest, identified contextually as from the model lab ("Anthropic"-adjacent; "you're like, Anthropic" framing; arrives "busy day" after model release, sleep-deprived). Discusses internal rationale for Opus 5.5, effort levels, and feedback loops. Flags plans to publish a deep-research write-up on effort levels. ### Mentioned but not speaking - **Harsh and Josh** — Greeted re: screen-sharing issue in Dia browser. - **Ariel** — Briefly addressed regarding effort-level habits. ## Referenced Models & Products - **Claude Opus 5.5** (Anthropic) — released "hours ago"; central subject. - **Opus 5 / Sonnet / Sonnet 3.5 / Sonnet 4.5 / Opus 4.5 / Haiku** — prior Anthropic models referenced. - **Fable 5 / Fable 5.1** — competitor lab model, prior daily driver of several panelists. - **GPT-6 Soul, GPT-6 (Astra?), GVT-5.6** — OpenAI-labeled models; **Astra**, **Luna**, **Soul 5.6/6**, **Codex**, **ChatGPT for work**, **OpenAI Max** referenced. - **Grok 4.7 / 4.6 / 4.5, Grok 47** — xAI-labeled models; 4.7 described as released hastily with early reliability regressions. - Tools/platforms: **Cursor, Cloud Code / Claude Code ("Cloud Tag"), Blender, Dia browser, Origami Studio, Figma, MCP, Unity/Unreal, terminal bench**. ## Provenance Facts — Sources & Evidence Channels - **Every's checks platform**: public at `checks.every.to/p/dance-editorial-checks` (Dan) and Mike's checks page; used for personal/editorial benchmarks built from real daily work, contrasted with "Swee/Sweet Bench"-style theoretical benchmarks. - **Every's vibe-check write-up**: compiled testing to be emailed to subscribers "soon"; published at "every.every.to." - **Dan's comparison article** (GPT-6 Soul vs. Opus 5.5) submitted for publication, expected within hours. - **Tariq's planned deep-research publication** on effort levels (forthcoming). - **Every's planned effort-level guide** (announced by host). - **Cross-posts planned on X and LinkedIn.** - **Anthropic-supplied information**: effort-level comparability between Opus 5.5 and Soul ("Anthropic shared this"). - **Claim about OpenAI**: stricter security classifiers recently added to Codex, reducing usability for long autonomous tasks — attributed to panelists' recent experience, presented as inference ("they probably will fix this"). ## Comment Availability - Audience interaction: one live audience question read on air ("Is improving most behavior on average or lower level thinking?"). No explicit comment-section availability stated. ## Promotion Flags - Closing **promotional segment for Every subscription**: "Everyday is the only subscription you need to stay on the edge of AI" — explicit promo. - Repeated references to Every's forthcoming paid/email deliverables (vibe check newsletter, checks platform) — promotional. ## Data-Quality Note - The transcript is a garbled machine translation; speaker names and model names are inconsistent transliterations (e.g., "Kade/Cade/Kate," "Natasha/Nitesh," "Fable/Fagal," "Opus/Office"). Identities above are best-effort reconstructions from context; uncertainty flagged where names varied. I'm not live. We are now live. on air. We are in direct on air. Okay, now we're in. live broadcast. Hmm, thank you guys for that You are with us. Kind regards please come to Every, our live broadcast our latest Vibe Check. We are doing this, when, uh, new models come from large laboratories. Uh, yeah, that came out close hours ago, this is, um, Opus 5.5, uh, Claude Opus 5.5 from Anthropic. Uh, Vibe Check will be released soon. He will arrive soon your postal code boxes. We are crazy we collect our pictures, all our screenshots, all our tests and everything another into one whole the package that we soon to you we will send. So you you will be able to read all our testing, test results coverage and everything else there. Hmm, in the meantime I Cade Lee. I am the boss. Every editor and I'm replacing today. our fearless leader Dan Schipper. Um, and I'm here with Kieran. Classen, extraordinary composite engineer who invented compound engineering and works on Cora, and by Mike Taylor, our new the head of Vivaldi. Ahem, Welcome, guys. Yes. Thank you. New Vivaldi leader, but old old The man is here, you know. AT, yes, they are already here long. You were. And to us too some will join other guests, but here with Who do we start with? So, um, who wants begin? Yes, we did it. Testing. Yes, we are tested this whole 5.5, and it was very...it was very nice, and we we are going to show you, for whom it is need, what is it good, and just give our first vibrations. Um, and so, like, I hope you try this model. But in general, our vibration checks such as: how do you Do you evaluate? She is golden. star? Is it green? Is it yellow or red? ? Um, so maybe... Mike, don't you want to to start with which one your, uh, verdict, uh, to date regarding this model, and give your general opinion? Yes. Yes. What model is this? She is just very cheerful. It seems, uh, it is felt on 4.8, like, you know, good ones times. Um, uh, but no, she, uh, this, uh, this, uh- uh, I rated it as green gold. Um, in basically, I think that she would almost be golden, uh, except this, she just ruined the couple, uh, tasks through bad timing- management. Uh, me Can I tell you a little more about this is to talk. But, uh- eh, but there was only one a market that would be against her. Uh, I think so. otherwise she would be gold. Um, yes, with her. very easy talk. Um, it's like Fable level 5.1 from point from the perspective of the individual. Uh, and he got rid of everyone those strange ones features that it had Opus 5, uh, what was it at first? forced me to go, uh, and, you know, do jump to Codex. And therefore I use, uh, OpenAI models, uh, from that time as your main daily driver, ah, e -uh, you know, Fable I use about 20% time. Uh, but then I exhausted him, and Opus 5 was not enough interesting, so that to return to it. So, uh, I think that now that this fixed, I, uh, I returned to about 50/50 between the two of them, and the 50% that I I do it with Codex, it's simple because I have many existing topics, and they are slow are passing. I think the main title here is that this model drags some of our teammates, who have switched to Codex, back to Claude, what is a big problem, and that it works like Fable, but much more efficient, not at all for that price. I will only say with editorial point sight when we checked the atmosphere Opus 5, it was amazing, how much people like it not really I liked it and I can tell you how it is hard. I hated him. He excellent. Yes, and it is very difficult. to force people to give significant feedback about what they just don't I like that. It's quite understandable. You You don't want to spend. It takes a lot of time, but it was very, very difficult to force people to express oneself and to say something else, than what I just don't like work with this. It there was a big change, and I I can say that after editing, viewing and communication with everyone, It seems that everyone has different attitudes towards of this model. Yes, for me Fable 5 and Fable 5 1 have always been my daily drivers, because this is simply the best model, and there is a question, is it worth it if you have unlimited budget? I I don't choose budget ones. models, I just I choose the best one. model, the one with which I like it the most like work. And Opus 5.5 now my daily driver that is simple madness. As open, I I hate Opus 5. I rated it at red, and sometimes on green, but I like it I just didn't like it. because he annoyed me from all sides. Opus 5.5 inspires me like this the level of trust that I have never seen in no other model, except Fable 5 and 5.1. And he simply better than Fable 5.1 in certain things, so...what is it is he better in Fable 5.1? Yes, I will. examples later, but actually I was thinking: " Okay, this is Opus. Or Do I hate him? He thought, "Oh, he good". And I thought: "Oh, but should I compare it with Fable? And then I started compare it with Fable. It was something on like, “Oh, \[ \_\_ \]. As if this can be compared to Fable. And this is possible. compare with Astra. As if it competes with large models that in itself already madness. We already saw this, perhaps, from Sonnet, uh, like Sonnet overtakes Opus, but we never seen that was he overtaking Astra or, um, Fable. So this is something special, and, obviously, this will happen more often in future. So be ready for this, but, hmm, what in it's better, so that's it that he is really good tuned to levels of effort. So traditionally, if you choose the cheaper one, less intelligent model if you choose something on kind of too high level, he just does a lot things. You are like: simple stop. You do too much. This is too much. . And Fable seems to know that he does too much. He just does better. And this is the first model in smaller size categories where more effort means better. And no less effort means worse. It simply means less details, more space for you as creator. And, and I think, especially the level effort...Very, very interesting part of Opus, where, in my opinion, he a little better than Fable, also in my work. I'll show you something. He can receive details in 3D, uh, and, for example, spatial movement. I I'm sure they are. conducted additional learning from this reason, because saw benchmarks from using such things, but what I I get from Astra, this is what what is Astra Slope now fully 3D, and all have titles, buttons, violins and other. I think I would like Opus 5.5 Slope, because it's simple a toned down version, but with a big the number of parts, and yes, it looks like we let's look at some examples, but for it's really me main points. Um, and just in everyday work, how do I do it I use it. faster and cheaper. I I hate that I use Fagal as daily driver, so that it is so inaccessible for the whole world. No one I like that there is no no point in paying thousands dollars per day for your work. Hmm, but yes, we...have special guest, seems. Oh, he's here. Is this Tariq? Oh yes, Tariq. As How are you, guys? Hey, how are you? Hey, you're good to me. Do you hear? Yes. Wonderful, wonderful. THERE ARE, Thank you for inviting me. me online. Thank you for joining. . Um, how is your day? stream Yes, and how was your day? Yes, yes. I've been listening for a long time. , I'm calling for the first time. Actually, it seems that for the second time. No, you were, you were on In this case, yes. Yes, yes, yes, exactly like that . Have a wonderful day. Yes, yes, yes. You know, I I think it's always there was a lot of work, to release the model, you know, and that's why a little like: oh, hello, Eric, usually a little bit enough sleep. I have here is coffee, but besides this, this is so, very excitedly. Yeah, so you guys, communicate as if it were a new family, this is new, obviously, this is...new model, but what means it's new family? Of course, that you choose 0.5. I'm curious, which ones? internal ideas are behind this, yes. Yes, I mean, I I think it's simple. big moment, you know, like Opus 5.5, I I really think that's it. just a model that we fell in love internally. We saw how testers of it love, and I think that this is a model that they will remember, you know, I think these are models. such as Sonnet 3.5 and Opus 4.5. But we also, yes, we announced that we release, you know, Sonnet and Haiku. And I think, that they will be, You know, it's already past. a long time since then, how we really had great Sonnet models and Haiku, and I'm glad you are, Guys, you'll see. I I think so, family 5.5 in general- this is what we want, hm , I think, uh, we would like to bring a lot of things, for example, advanced intelligence so that...how incredible workflows every day. Yes. Yes. What about you personally? created by with this? Obviously, we all trying new things, for example, we are rebuilding what did before. Where are you? Have you seen joy? We felt joy in any moment, and every time this the model stopped work for us, we felt real pain. We thought, "Ah, our life "it's over", and a little it seems like we felt when Fable disappeared. So I'm curious, what did you feel and what created by with this? Yes, I think the thing is... So I did a bunch things in their own way by subscription at home. I I think one of the things , which I felt, is that, what I love use workflows, and I I think it's been a long time. it was simply impossible use workflows with Fable, you know, you just quickly exhausted limits, you understand? AND it was less effectively tokens, than, you know, such things. And therefore, I felt that use of workers processes for me—this one of the ways get a lot quality of...work that you are doing. Um, and yes, Opus 5.5 was the first a chance when I could use similar workers processes on plans subscriptions, um, and how about Frontier Intelligence. And that's why I tried things like this, like, um, redesign mine personal website. And, um, I asked several models do it. And this is what , like design, too needs a lot critics, you know? AND so, you really you know, it's not only creativity. There is something like that, but there is also like iteration. AND, Um, yes, I'm going. write a little about it later, but it was, Um, that was great. yeah i just um I felt that I could faster or cheaper to feel complete the feel of the Fable model and to be able to use all these bells and whistles. Perfectly. Yes. So what are you choose now: Fable or Opus? Which one are you? Do you use it? Because , obviously, you need to be ordinary user subscriptions, but, obviously, you have, probably, there is unlimited access from both sides. Of course. Yes, I think so. that this is one of those cases where I would modeled it simply better than Fable 5.1 now. You know, I I think there might be some cases regarding planning, perhaps security etc., but I I think that in average if you don't want to accept solution, Opus 5.5 is just safe decision. And sometimes it is like this happens. For example, you know, when we had Sonic 4.5 last year, you know , it was something on kind of, although it was Sonic model, it was the best model that we had, you know? So, um, yeah, I guess that I just always I'm reaching for Opus 5.5. Perfectly. Mike, are there any question? Yeah, I just think, What's on my mind? I would have narrated, and I would have been interesting to see how everything was happening inside if can you share, but it feels like we had an incident a real leap into intellects from Mythos, and then with Fable. But this It was like talking to extremely intelligent You know, you're not a human being. could understand that they say. It's like made me to feel truly stupid, you know. And with Fable 5.1 I realized, oh no, he can be intelligent and at the same time have a wonderful individuality. AND you took these lessons, and then applied them to Opus 5.5? This is how we are Did you get here? Yes, I think Opus 5.5— it's just a combination many reviews, you know. I think we we always try to offer you the smartest and the most intelligent models to start with, and I think I need it. some time to understand which pain points. And I I think that, you know, when you have these for a long time working, uh, agents, Yes? And they are trying to communicate with you, you must to realize that they there is also a lot of things say, do you understand? AND therefore, I think I have many condolences to, for example, Faithful Five and Opus Five models, where they you know, they are... They are very working hard . They do a bunch of things. work. Um, I think so. we are like, uh, but also the way you communicate, is an expression of your intelligence. So, You know, I don't think so. using more complex language means bigger mind, right? AND therefore, I think that this one of the ways, which we tried, so that new models expressed their intelligence, and I think that's right, a lot of this it turned out thanks to feedback from you, guys, and I think that also reviews about token efficiency and other things that we just heard. So, uh, I I think it's simple. one of those things when it takes a little time, you know that the wheel feedback fully it got unstuck. This is not product where you can just do PR and to do it, but, uh, yes, we always we listen and we are trying to act on based on these reviews. Yes, that's great. And what are sources of feedback communication, uh, what are you, Guys, tell me, how is it? so many different competing interests, right? Do you have any idea about it? Who is typical? the person for whom you trying to create something? Yes, I mean that intelligence and models–this one of those strange ones things that sometimes don't exist compromises like you could you imagine, you know, sometimes a model may become cheaper, faster and smarter at the same time , do you understand? And I think, what is it like 0.5.5, right? So, uh, I don't I think we were going to compromise, sacrificing some other option using. We we believe that this just in general intellectually. Uh, everyone should to feel that his opportunities have increased. THERE ARE- uh, I think there always is , you know, more work that we can to do if there are certain things that maybe is missing. But I think which, you know, I don't think, that the models have the same compromises, which product sometimes. And so, I think this is a sign. that intelligence will become more and more common, cheaper, faster. Uh, and we... Just when here you get, you know what yes, it is will continue occur. So we have a lot more such moments, on which we are waiting for. Yes, yes, yes, as I already said, Sonnet and Haiku will also be. Yes. Perfectly. And what about you? opinion, this model will change the most or for whom will it change things because for me she is not very changes, except that it is cheaper, whatever I do now, because I used to used Fair Will. Uh, it's like different types characters, is it like, for whom , in your opinion, this will have the largest influence? I mean, when sufficient using tokens everyone is sensitive to value. You see, What am I talking about? I think that, like you, in general we we see that when we we make models cheaper, more accessible and simpler in management, people use them more, and I think that everyone, in my opinion, benefit from opportunities more often use stable intelligence, Do you understand? So, I I think it's inside. you probably have things like: "Is it worth it?" "Is that right?" "curve, something like, or something like that", "Oh, like, Is it worth watching? code in this case? "You know, like, or something like that", "is it worth it "is this a test of that?"" or something like that, or is it worth it at all engage in speculative, type, PR based on this review? "and things like that. I think that will change all this mathematics. I am from I am looking forward to it. this in such products like Cloud Tag, where, you know, like, um, there is ...like, he has a lot what does it use, but also because he so proactive, which can use tokens, tokens, and me I think, like, um, do it more accessible, I think, this is what I am, I am very glad. So, I think that everyone should to feel some changes, you know, and I think that In general, I think that everyone will get more tokens and hopefully , will receive from this more useful. Yes. Yes, we still have it. Jeff's model, which I saved. It's like suddenly, like all these things , I thought that, probably , it's not worth it do. As if it were possible try. But actually I found um, I found, um, this one the Opus 5.5 model is really makes great use of Jeff as a tool. Oh, cool. Yes. Yes. Yes, I am. created a lot interesting things last week. Um, with hereby. Perfectly. I wanted a little to consider this from the point of view view of simple writing. Our tester Katie, who couldn't to be on today stream, uh, spent your entire arsenal tests. Um, and Mike, I I think you too spend some of these writing tests. Um, and um, I think her the main diagnosis was this: this is true a great partner. It simply the easiest, with which one is better to work with, even if it's not best author drawings in general. Um, but this is essentially, um, you know, Anthropic, you're like fixed it individuality. Um, I wonder if Have you encountered this? how can you use this model or others models for writing, and What are you looking for in this? models in this in relation? Um, yes, I mean , I think that, um, yes, I I think that definitely there was a lot of attention dedicated to the good communicating with you. You know, I think that, um , in writing, the way I usually I use writing, it's something like, um, editing or feedback or something like that, you know? Where I wrote something and please, type, type, try to change, but mostly preserve the language unchanged, and type, I trying to find words that seem strange, or places, where can I to move things, Do you understand? That's how I am. I usually write. I I think that another one of things that I like to do in writing, there is a holding many studies. And I think that one of painful moments previous models was that the reports were very dense and difficult to read. And I think that with this everything became a model much better. So, hmm, yes, yes. Perfectly. Mike, I'm curious. your point of view on this, like another one of ours authors. Yes, hmm, this is, hmm, I would say that in I'm still about 50/50 between this and Fable 5.1 regarding writing. Hmm, and I did a blind test, and I spent 2 hours. Um, blind, uh, evaluating things that very boring. Um, but, uh, and if I found out that they are both the same good. Um, so, uh, this it was, uh, that was good, good to know, eh, although, and prove. Um, I think, that one thing that I, uh, noticed in this model, it's that she's better for cooperation, but also don't be afraid decide. Hmm. Um, so with the PowerPoint test, uh, which, you know, we we can show in a certain point, uh, this actually changed some designer decision. Um, for example, he created a different style SVG than we usually do we place it, and it's all I was still on the brand, eh, but it was very it's like if you hired a creative director, they could I would add a little. Um, and, Well, that's probably very a difficult nuance to correctly it understand. It's like you wanted it to go according to the instructions, but you also want to it has rethought, what, you know, uh, what, what exceptions to the rule, Do you understand? Yes, I think so. overall it's simple really, really a difficult balance, where, I I think it's somewhat similar. from memory, where perhaps it can remember and understand something like yours profile before, you know, and then we will know on which equal to you to meet. I think what is there in general here What to work on. I I think it will be a great way in which models express intelligence, that is, they can do what whatever. Uh, but, you know, if for example, that anything, for example, junior engineer will ask you to do something very ambitious, maybe for you will need to say something on like, "Hey, let's Let's wait a bit. What you are trying make? "But, maybe if someone is very experienced will ask you something do, you say:" Okay, yeah, we just "We'll do it." And I think that this is one of the most difficult things in design models, of course, is to achieve this precise balance. And, uh- uh, yes, I'm glad we achieved this in Opus 4.5. Hmm, but I think this is something like high level expression of intelligence and opportunities to meet with user where he is. So, I like it in in anthropic models, that they are sharper than OpenAI models that I I use. So I I never understand. this evil one? So, as if, as if, as if, like, OpenAI models quite smooth. They never anything are doing wrong, but never they stop me at place. And, you know, say, with Opus 5.5, sometimes he does something and I I think, "Oh, that's nice." an idea that I don't know about thought ". And it's like, but, you know, I think, that some compromise is that, you know, in order to to propose a new one idea, you need to have my opinion, probably. I a little too much now anthropomorphize this, but, you know, it's similar to work with a colleague- creative, not on work in the main the office of the corporation, which smooths things out. Yes, I think one from the examples that I I like to suggest that to help people understand which a compromise arises when building models, like this: "Okay, you you are asking for a model, as if I don't know. Who best basketball player in the world whether in history or not Yes?" And the model is this: " "Oh, it's Michael Jordan." and you say, "No, that's clearly something like Steph Curry or LeBron James." AND you want the model was like, "Oh, yes, you You are right. It LeBron James via X, Y and Z", do you want, so that the model is like this : "Well, I don't really assured". Reason, why do I think this is Michael Jordan, like this, Yes? And I think that this very thin, so-so difference; but I think What is this, you know? can you imagine something like that for, for example, any problems and search the right balance between the model that trusts his own judgment, em, and, type, works with you, against, like, simply, implementation. Yes. Or there is our colleague Natasha on stream? Yes, he is here. Yes. Is he there? Natasha, are you here? I am here. Yes, he is here. Hey, Natasha. Greetings. Derek, do you have more? one question, time for another one question? Of course. Yes, Of course. Yes. Em. I like it level setting efforts. It's done. so good, and I know that it is very difficult because usually, especially with stupider models, simply overload work, and especially in Opus 5, I saw it, where is it just did it too much cycles, etc. So me I wonder what you are found out what he does this is so good, because it's stupid model and, type, smaller model than Fable, but effort so good react. So, to me I wonder if you can to share something, What did you learn there? , or what was the idea using levels efforts. Yes, I actually am. I am working on a deep research efforts, which I hope publish in the near future. I I think that, hmm, TLDR is that, um, effort should be to consider as yours willingness to spend , um, money for, or, um, budget for the problem; but Claude has, um, minimum level, hmm , he always will try solve the problem, Do you understand? And so, hmm, he will never change your decision, or not Yes? As if he never will not change his/her approach, in essence. Hmm, but, um, if you add to it more effort, he will be able to spend a little more time for deliberation problems at an early age stages, hmm, sometimes on settings ways of it testing, testing and checks. And, uh, I I think I considered, for example, terminal bench, and problems with many extreme cases , for example, creating a tool for cleaning up HTML or something there are many of these different ways, like this can lead to failures. Do you want run this with with great effort. Um, and you want this find, and type: uh, but if you think:" Hey, I have a good self. for each model, in I have a beautiful one. mental model that she can do. And every time when you have this problem, you want to make an effort low, uh, or average. Um, me too I think one of the things which can us to surprise in in the future, this is what really smart models can become simply Pareto- dominant because if you don't need to check how you are can you imagine, the ideal model is not needs checking your work, or not Yes? It's just like, "Oh, I just did it," and you think, "Oh, you You're right, you know. you did it". Um, and therefore, if it is reasonable the model can, for example, or the ideal model can just do something at once, but less a smart model can do this, and then verify, protest, verify, to protest, how do you you know, you can imagine that the difference in the price may be extremely large, and a smart model is everything one might be better for the task. And therefore, uh, yes, I think so efforts are, uh, curves effort is the beginning the one where, I think, actually, for many developers software providing you, you probably can perform low and medium level, and then, for example, when you need it task with higher budget, you can go to high and super high. Also switching efforts no more breaks leads to cash costs, therefore it is easier to do in the middle of the session. Perfectly. Yes, a little it seems that the lessons with Fable with low level there are actually costs more economically effective than with Opus, for example, with super high, right? Now this too will appear in Opus, and that, what do you also say in my testing, like- here is the launch of Opus 5.5 with low level expenses, in fact very impressive. This is not true. , as before, I would even did not consider possibility of low level, because I I like terrible ones. results, or something broken, or work performed improperly way, but in reality This is very good. How are you? You say there is a basic the line she wants achieve. Just when you increase the level efforts, it seems that more is added details, and it becomes more refined, and me too like this one low effort, because it's something like a sketch. Low more like sketch. This is complete. sketch, but it leaves more space to fill in what is way of working with model that I I really like it because I hate models, which fill everything suborder. Uh, others suppliers do it more often. So, I like it. to lower the level to increase your own input data. So, to me I really like this one. low, actually. Yes, I think so. , this goes back to what Mike said before, like: "Oh, you know, you know, when this adds its own flavor or repels, right? I think that, um, yes, this way of working with models—this is the minimum efforts. You are more into up to date with events, and I think that it's the same, you know , you can imagine something like: " Okay, your boss. gives you a task that must be done by an hour, not 12 hours, you understand? If you are given 12 hours, you you answer: " Okay, you probably just want me to "did it." You understand. , what am I talking about? And, um, I will to make a pile assumptions along the way affairs. And if you, you know, do it for 1 hour, you say: " Okay, give it to me. understand what you Do you want me to be able to? to do it specifically ". Therefore. Good. Yes, we actually I just wanted to to say that we just received a question from the audience where they said: "Is improving most behavior on average or lower level thinking?" And it seems that Yes. Yes, definitely. Steeply. Um, thank you, Tarek, who came. We We really appreciate it. We we know that today in Have a busy day. AND, um, you know, come back anytime When. Yes, that sounds good. Yes, thank you for that. I was invited. I appreciate it. Take care of yourself, guys. See you later. Thank you. Thank you. Um, Natasha, we're here, uh, with a fourth member Every, and uh, Natasha, senior engineer from applied artificial intelligence, which is closely collaborates with Mike over the estimates. Natasha, How are you doing? Hello, I have everything. good. Ahem, enjoying Opus 5 tonight. You like it. Ahem, So what did you do? Over What did you work on? Okay, so, um, I tried, uh, test models on their ability to teach and tried truly expand them, uh, and to see how much they can come in. So I have this crazy, um, research skill explanation of where I am I'm asking for a model. to do research the explanation that is multipart HTML- code, um, which, uh, has in detail explain the topic, um, through, uh, interconnected tree of HTML pages. Maybe it will be easier, if I show you that I mean. Uh, if I show mine screen for a second. Can you see my screen? Yes. Okay, great. Therefore, hmm, that's what I have on under consideration research explanation, and I have it, um, results from, uh, four models here, um, including Fable, uh, and Opus. Uh, so I can show you what Opus 5 did for this task. And this looks like Yes. Therefore, research is, by basically, a bunch interconnected HTML- pages. Imagine this is like a very advanced version of the slideshow that can use for studying a specific topic. And, um, from the first impressions it is clear that Opus 5 did well. Then, um, here's Fable. Hmm, hmm, sorry. Here is Fable. AND Fable improved a bit clarity, hmm, of words and, you know, design. But Opus 5.5 is where it all is It's just crazy. Therefore, hmm, it's like a map of everything this. And, um, you can to see that she is very, very good polished. Everywhere yes, hmm, really high quality interactive elements. And if I I'll just go and, um, you you know, I'll investigate any -what a coincidence slide, it's like, hmm, will present these really beautiful small JavaScript games throughout. Hmm, and, um, very easy to understand concept. So, it's like best designer interface and better teacher. And, um, you you know, this is, I think, comes from him abilities, hmm, just to be better to be a writer and the best in frontend, than any other the model that I tested in this direction. Wow. This is enough. incredible. Hmm, Yes. Hmm, yes, eventually, he worked close 5 hours to do all this in one fell swoop, and, uh, used dozens of subagents, you know, he worked over 50 different files , creating them and developing, checking your work and, uh, releasing materials . IN I actually exist. questions about time and how much, uh, you know, Mike, you found that you can allow him to do something that Opus 5, you can to allow him something to do, and this is on as much as you want hours, and you don't will need to check if you really need was to check this and redirect? Yeah, uh, a few things, which were evaluated in my, uh, my landmarks, this is what he did not finish task within 10 minutes. Uh, and that was it. something like several educational tasks. And, uh, this really the only thing that...on what was reduced assessment. And, uh, I'm in in some sense it is a little unfair, because I really wouldn't set 10- minute timer for this task. Me I just wouldn't care. . Uh, I work like this I am short. I'll tell you, and then I forget about it. Uh, and I allow this. work. And then, when it is done, I I look back and say : "Okay, it is done. Now it is worthy?" "Uh, and that's why I I spend all my time on task, and then on review, uh, in true style complex engineering. THERE ARE -uh, but, uh, I, uh, that's why I'm not really worried. time constraints. But it was also enough I wonder why. this failed, because none of the others models never encountered this problem. And, uh, by In fact, it was creation educational material, and it was creating really realistic educational material, and it was so persistently worked on educational material that not really completed the general, such as educational program, task. So he was evaluated for absence final answer , because it ended time. Am I not realized that this will be a problem. No in no one else Time is up. So, if you really are felt a little humanly, I definitely did it on the test earlier, you know? Yes, I am. I think this is interesting. anecdote, but I don't, like, uh, I'm not really I'm worried about this. reason for your workflow, what...why am I still rated it as very high. Okay, cool. AND Natasha and Kiana, I I wonder if you have reached your limits with this, or you reported this, Do you know how, how do you discovered that I it took eight resets or something? Yes. Maybe you are on the same page. equal, you know, not the best or average example, but , uh, no, but also like me read, on each equal efforts and all benchmarks. So it's not ordinary using. Uh, the only thing I don't like I like it, that's what this is at higher levels efforts. If you decide to choose extremely high or tall, he works for a long time, that, maybe okay, but, uh- uh, I don't know if it's like adding such amount of value for the time he spends. Uh, this is the only question that I'm still here. If compare this, especially with others suppliers who historically had a little more effective tokens, but I have to to say that it is low The level is great. For example, creating landing page, it only works 2 minutes on low equal, and that's great. So, just to know, when to use correct level effort, I think it is a skill that you all need still needed here use. Um, Nitesh, what did you discover? Yes, I, uh, actually didn't realize, but uh, until you asked this question, Kate. Um, but I'm all at this time worked on average level effort, and, uh, I don't felt the need to move up level of effort, therefore what, uh, is that simple It struck me, uh, in everything I have from him demanded. So, uh, yeah , and I haven't reached the limits, eh- uh, doing studied. I I think, probably, from that one same reasons. I just did it, uh, you know, like average level of effort. Uh, so, uh, yeah, I think that's it , uh, enough, you know, uh- e, effectively as in low, and at medium efforts. I will say that from the point in terms of effort level, it something we talk about a lot talked here, on Every, and and so on, and on editorial team, namely: to know when and how to use these are different levels. Kieran, you say that do you feel that this skill, and I absolutely...I'm with you I agree. I think we actually we will work on guide to level of effort, and I know that Tariq said that he also something will publish, because, on in my opinion, obvious the thought is: "I just I will put everything on "high level", but this will not always work all the best to you results and, often, obviously, it will be incredibly expensive. On sorry, that's just how I am I do, Ariel. Yes, actually, you you understand that this is not always worth it. You understand that you get the best results or more detailed results. It's simple. question: is it true Is it worth it? And and, and that's it worth it because it takes a lot more time consuming and expensive more expensive. So, I think, these two, as well as what it was before when this just works for a long time, in cycles and executes too much work, actually nowhere are used. You just get really better job, but yes, Do you really care? ? All these side effects cases or extremes cases. Yes. Of course. Hmm. I am not I know, can I? also be a guest user. Oh, is he here? Uh, no, not really. guest. Tyler. Hello, Tyler. Hello, Tyler. Can you hear us? Yes. Tyler Tyler— designer at Everyday, who experimented with this model and has a few really cool things to to show. Yes, he did. some incredible things. Maybe you could would like to share what you created, and also comment duration of work, because I know what you did. a game that doesn't ended for that the time you had with this model. So I I share with us my feelings, impressions, and, be please share yours screenshot. Yes, of course. So, for context switch yours microphone on AirPods, because, we seem to hear your room. But yes, Tyler—he's just played very hard at games, and, yes, he has there were several moments, when he thought: " "Wow, what happened?" "How those moments. So, be please tell us about the moments when you were amazed. Can you hear me? Wait, how's mine? microphone? Good? Perfectly. Yes. Regarding me, I had to to work during week, so the only a test that I could to do, were long tests on a separate screen from the side, and, It seems like five or so. six times during I've been slacking for a week. jaw. Um, let me. show me now your screenshot. I I'll just show you the game, which you were talking about. And also, for example, in You have a drop jaw. Astra? How do you compare? her with others because Astra is also good copes with games with a wide coding, right? Isn't this the best? model? I don't think so anymore. I really not. Um, Astra never forced me so much to hang down Um, let me, and this actually my first stream, so I I really don't know how. to do this, so we'll see. Allow I just make sure that everything, and that's all I said. Oh, yes. Private things. Yes, we always…” they are hiding madly screens and, uh, names models, etc. Okay, you see. screen? Do you see the game? Yes. Yes. Do you hear music? No, we don't hear music. . We don't hear music, good. She super loud. Hear the music for us. So, this game has music, there is a voice, there is everything. Hmm, she worked, it seems 5 out of half a day. She never ended. Hmm, this is a game based on San- Francisco, and I just told her to copy, make a game like Genshin Impact, based in San- Francisco, and she just everything myself came up with. It was so fun. As if this sourdough on sourdough. Hmm, that's just how it is. ridiculously. For example, top right you you can see that the rent is about 4200 dollars. Oh my God. And this is Karl Tuman? Yes. Hmm, in the game you fighting with Karl Fog. Good. So, Carl is here somewhere. Hmm, he can run. He seems to be riding on electric scooters. It's just... It's just so funny. Hmm, and just so good ...did it. Uh, so were you in charge? some of these solutions ? For example, how many of of these decisions were yours, what about you? encouraged him make? Sorry, by the way, this is burrito foil instead of wings. It so funny. I am nothing. didn't do it. I swear By God. I just, um, sent a request from in one sentence with purpose, and he just created all these, like: " I can dodge." This so finished. I have there is a small one here scale. Um, he has special reception called "Series Z". Let me take a look, Can I have it for you? to show. So there something is written on like "do" round race for 40 millions of dollars." This is so funny, and his The sword is leaven. You see, this is bread on sourdough starter instead sword. This is just one of the funniest, hmm, games of all time. I have I mean, no, never, but I just...I ever been astrologer, and she never arrived so far away and never did not end. Um, yes, this is this game. She quite complete. It could go on. ...for a long time, so I probably worth it to stop. But, um, I I think you can to see a little comments here. You you can see that this says as if I Thought leader, you sound is on, the human resources department will hear about this. It's just like that. stupidity, hmm, it's simple such stupidity. And then it's just all sorts different movements and everything like this. Hmm, so what were they? moments when you have sagging jaw? It it was something like... And why, why is she sagged? I think that for me there were several moments, when did she start to self-evaluate oneself. Hmm, and that's why also took so a lot of time. But she caught things that I I would never have caught it. Hmm , and she caught it so, You know, I don't at all. could play the game during the week. It actually my first once I play it. Honestly, I watched Claude play in your testing. I thought, "Oh my God, I I can't believe that. "It's gone this far." "Um, and let me just close it because I can't take you anymore. hear. But it creates something like this table and gives himself rating on a scale from 1 to up to 10 based on visual elements , content, etc. AND it captures all these things, you know, it's can be very offensive to mentally ill human. I am going change the words on this T-shirt. In some the moment I read, hmm, her self-assessment, and there it was written: "You actually you can see the bottom this underwear female character, if you look under her." "She was so detailed and thorough in its the work that was there It is written: "I am not I will stop until I will get at least eight out of 10 overall." « Why is it not so? did it because it worked 5 and a half days. But also care. For example, grass, which you saw, wrote own shader to create this grass independently. Hmm, me too I think that everyone who created games in Astra before or previous models Claude, knows that You won't get that far, Do you understand? Needed many more tips, and the first the version you get in Unity or Unreal, just not It looks like this. Hmm, music is playing bad, sound sound effects bad, and these... In this there was everything: from quite nice graphics to... It could be said, that it is quite low , hmm, lag, and FPS was about 60. And it was just content...This It was so funny. It was the funniest model throughout history . For example, I don't know this, this is a sword from leaven, and as a series, ultimatum, and how your rent continues to grow, the further you are progress in the game, and ...in the end..." On at the first level you fighting with Karl Fog, and It turns out that Karl Fog actually loves San Francisco and just wants to embrace the city, or something like that. It was like this pointless. So I was just in love with him, hmm, all the time, while he was creating it . I looked aside and saw him testing game, and I thought: how she is so good with one, literally with one request? Zero, I promise, none. further. Wow, that's enough. incredible. Uh, for once. THERE ARE- Well, now to us. joined, um, Dan Schiffer, who came out from his, um, pit for checks and vibration testing, because a new one also came out AI model. Uh, so this is it really rich day in every world. Hey, Jan, how are you? Are you holding on? IN I'm fine. I have is Celsius, which I just I'm going to open. IN I have two waters, so we drink enough water. We drink. enough water. I Glad to be here. Uh, me checked the vibrations, hmm, all morning, and mine, hmm, my tweet went out... Um, for, um, G as in GPT-6 Soul. Me too, Kate, uh, just, uh, submitted material about comparison of GPT-6 Soul and Opus 5.5, uh, which one, I guess , it must be, I think, Kashik received it. So he has to come out. during the next several hours. You will... Yes. You will all receive message about Opus 55 in their mailboxes boxes. Thank you everyone for your patience . Usually we would like publish it immediately after leaving models, but we tested several models as you can hear, considering that we also worked with Open AI and you will get Dan's report on model comparison a little later. So Thank you. Really. So, I'm going, maybe I can to consider a little my...I don't think so we spent some tests...any GPT-6 test results Soul. Maybe I can. to consider this a little, and then I'll walk, well , a comparison between the two of them. So, uh OK. So, GPT-6 Soul just came out. For me, big stan The Code, which is whole day uses Astra , this is like my new daily driver. Uh, I still do from time to time. I am contacting Astra for, hmm, tasks of the highest order level, such as working with knowledge, the highest task level, but I think that, hmm, Soul is the same as and he has everything that I need it, but he is much cheaper. Uh, for me, regarding writing, using computer and, um, just tasks general work, I I think that this is, I think, the best model. However, he still doesn't has the same high level, like Fable 5.1, and, like You guys saw it, it seems that on this stream, what it does Opus 5.5, enough incredible. I found out, what for mine everyday work Opus 5.5 does a lot. For example, I ask to do something with it, and he will need about 15 minutes to return, and GPT-6— Is this a small effort? Because I didn't do that. I am not I'm playing with that. you should lower loudness, because in this the whole essence of Opus. For example, use it at low volume. This is the magic. . I really have to. try this. I didn't do that. Therefore I I would say it's worth it. check vibration Opus with a certain proportion skepticism, I mostly I use it on high volume and I'm just looking at that. will work out. And if you do that do you use settings by default. Hmm, yes, it works and does a lot of things, and therefore returns the result, which, I I think it's a little lighter. for me in that the way in which I I am working now. But, hmm, I want to stop at a second, and we have one of the things that we did, and I don't know, how long have we been went through this, but one of the things that we did, this is what we we begin transform our testing for benchmarks. And we have this is really cool platform that Mike and Nietzsche was created, uh ...checks. And, um, my checks for my texts now active, um, and I just added them to the chat. I hope this will appear. Um, my checks are now active, erm, for editorial purposes works, and they will explain to you why I like this one. model compared to Opus , um, in many forms daily written tasks that I I do. Um, let me to share yours screen. Sometimes it's not works at Dia. Therefore, to me, maybe you will have to to go out and come back. Um, let me know, if you guys You see my screen. Yes, there is no link. on X, but it checks.every.to/p/dance-editorial-checks . I will have to to go out and come back, because I'm in Dia, and, um, Dia is not allows me share the screen that extremely annoying. I convey my greetings to my to Harsh's friends, Harsh and Josh. You guys, must fix this, because I need it now need to go to Safari. So I will be back. literally in a second. While Dan is at it engaged, did you want to you, so that I can start share platform checks on which he... Yes, you can. Open. Hmm, yes, therefore. And I came back. Oh, not really. necessary. I am just was going to intervene and start sharing checks, but you were too fast for me. Okay, cool. Allow sharing screen. Share screen. Good. Okay, I think you, Guys, now you see. my screen, right? Yes? Yes. Okay, so here it is. checks. So, we have are These These These are mine editorial checks , and we have different checks for different people. So Mike has their own the checks I I would like him to shared if he haven't done it yet. AND here is the table results. Uh, in We have models here. Uh, you can see that Astra is currently mine the best model. THERE ARE- uh, Soul is She, it looks like This Soul is the second GPT-6 Soul-it is secondary model. THERE ARE- eh, and Opus eh 5.5 is close before that, but not right there. This is a pile. various tasks that I I do it every day. Uh, so, this So, when you you see benchmarks, they usually uh consist of both many, many theoretical tasks, which, um, you know, on like Sweet Bench or something like, where are they...This real tasks programming, but they are not really from works of real people. So when you see new model and you see the test result, and he's a little better, that's not really tells you whether this model is suitable for my work? So we started creating tests based on our daily work, um, so that we can evaluate these models a little more strict in a way, um, um, on our actual task, which I think , will tell you much more about what this model is for good, but why? she is not, than more general test. Therefore, this is my personal test for editorial writing assignments. So, in particular, one of tasks on which I testing this model, this, um, can she give good review about article? The other is, um, or can she write good first draft? I think that that's actually quite enough good test...um, so to evaluate this or show differences between them. So, in this particular one I have a task. the document that I wrote about what it is personal test about personal benchmarks. AND I ask him to write. introductory section personal section test for the assignment personal benchmark . Um, and now we can see results between GVT 6, Astra, Soul, Opus 5, GVT 5.6 and Opus 5.5. Hmm, and why? results received exactly those points. For example, um, uh, this introduction written by GVT 6 Soul. Personal benchmark is growing set tasks from your real work, rated by your own standards . Let's say you ask AI open introduction articles, and he doesn't notices that the title weak. Your correction becomes verification. Or caught the model weak headline? Run this test on models, and you will be able to to see which ones correspond to yours standards, regardless of whether who created them. When you correct more work, benchmark becomes a guideline for Which model is the best? for you and when she the best. Time switch. Um, GPT-6 Astra, personal benchmark is a set tasks taken from your real work, with inspections , which reflect how looks good result for you. AND you can see mine Check here. Therefore, did it pass each of these checks? Um, I... Let's compare this is from Opus 5.5. Personal A benchmark is a test, built on your own work. It a set of tasks that you actually performed, each of which is estimated by simple "yes" or "no" checks "no", which reflect your opinion. Actually, I really like it. I like it. Um, that's not it. typical answer for Opus 5.5. One of the things, which I discovered in Opus 5.5, this is what he gives many of these are longer answers that are a little distracting. Ahem, let me just to see if Can I find a good one? comparison for these models. Ahem, let me look. You know , my personal ability to find that's the case, What do I want for you? show a little more difficult for this specific task , but I will say, check this out. I I will find a couple of cases, to show you in particular, but mine general opinion about Opus 5.5, and you can see some of these checks here. My general opinion is that he very, very smart. He gives you great, long, thick answers. Hmm, and he not so good at it to teach the main the idea right at the beginning , for example, entering into article, initial paragraph or any type of editing. Therefore, the text that he creates, occupies a lot of time to to get to the point. And what it concerns me, me I really like it. minimalist, immediately go to prose, and I believe that, hmm, 5, hmm, GBT-6, Soul and Astra much better in this in general. That's not less, 5.5 is enough strong author. Hmm, Mike, I know you spent a lot of time here tests. Do you also want talk about it? Yes. Yes, let's go. Let's do it. Hmm, so, uh -eh, I'll share mine screen. Uh, okay. Hmm , therefore, looking at my checks if you guys see this. I don't...I think so. Now we can. Yes. Okay, cool. So, uh, it's available. on a public website, so be uh, checks.every.to uh, and uh, this Here, I'll show you. THERE ARE -uh, it will be something like like \{slash\} p \{ oblique click\} Checks Mike, uh, if want to see public. Uh, this is mine. private. It has a few things that I I will share with you, guys, in person on air. But, uh, The main thing is that, in my opinion, thought, it was really interesting thing about Opus 5.5 is that that he was taking some decisions that I actually liked . So, for example, not always adhered to correct instructions. And you can you see this here. So, if you Look at this. PowerPoint task where we we say create PowerPoint in every style guide brand. And you will notice, that some of them are not like that already good. The first others that you see, deep search, for example, performed truly terrible work. Guide to the brand style should have been green, right? AND Opus 5.5 a little deviated from the topic and did it...black, which, actually, to me I liked it more. Ahem , and you know, here They did well. Another thing that they they did it, they created it custom SVG style. So, As you can see, all this unclear. Uh, and At first I thought, that this is a mistake. Uh, yeah then I realized that, while looking, they saved this unique style. Uh, Here's another example. AND here they are, they did it several stylistic decisions that, looking at them and thinking, "Oh, they received a discount for discrepancy our guide to brand style." "As if it's still a brand, that is it's still what I am I think we would with pleasure aired, even if it's not completely corresponds green or uses something another SVG. He did good choice, and, uh, this is almost what you expect to see, if given this task to a person, uh, and in them had similar opinions about how to expand brand. It seems like that's exactly what happened. I am impressed by him. I receive. Another thing that he really succeeded I did it, this is a map Hoboken, and I love it also very like visually. So, if you see this, you You know I moved. to Hoboken on beginning of the year. I wanted to understand in which restaurants I must to take their children. What are hidden pearls? And he just coped very well with choosing different places, who I am since then visited, and I know, what they really are good, as well as just made a map very beautiful So if we will compare this with other models, usually models choose it somehow strange, for example, I I think this is what it looks like very outdated, like- here is the map of open streets. style. And this It just doesn't look right. so beautiful. Hmm, and sometimes models completely fail at this, uh, like here. Uh, old man Opus, and that's the main thing that forced me not to to love the old Opus, Opus 5, uh, he actually, uh , decided to create own card, right? Yes? And he decided create your own SVG, e- e to create this map. And to him had to do much more work than necessary, uh, and I think that new Opus, uh, decided it. So, uh, I was, I was, I was, I was, I was , I was, I was, I was, I was, I was, I was, I was , I was, I don't know if you already said that, but you were one of those who recently converted together with Tyler to Codex universe. Uh, where? you now, considering your testing of these two models side by side shoulder? Yeah, so I would say that , uh, what happened, uh, new OpenAI models, uh, improved, uh, you know, from generation to generation to a generation. Uh, and they...Now much cheaper. I saw that Luna prices, in particular, very attractive. Hmm, but, um, they did a little more of them timid. And that's why I tried use, hmm, sole during most of the time last week, and, um, I constantly returned to her and found that she stopped. And that was it. so annoying, that I was very happy, when Opus 5.5 started work very well, hmm, because this tilted enough a scale for me . So I would say that I'm about 50/50 right now. I still have a few. old topics in the Codex. Me still like it Codex system. Hmm, but, hmm, now that I no need use Fable for everything in Claude, and in I have a daily driver that realistic in using and, um, cheap enough, to use for everything. I think that I am passing now. This is one of the things that I should add and to say that we are definitely found this. I think what is this, perhaps, part of that happened for you, This is OpenAI. For the last one a week or so to the Codex program added stricter classifiers security. Hmm, I think so. what was it like answer to all recent incidents security. So completely it is logical that they are done. However, this really reduced convenience use of these models for long autonomous tasks, where I had to to repeat constantly : “I approve. I clearly I approve of this. I approve. it. And approval would be considered very, very narrow. Hmm, and that's why I I think you...I think that they probably will fix this. It can be fixed, but if you are testing these models today, I I think you probably find in the Codex and in ChatGPT for work, that, hmm , they are forbidden access if you use "approve" function for me" for default. Them deny access to many things, and This is quite unpleasant. Yes, to some extent Anthropic had a small advantage here because they had to to introduce this classifier for Fable, and now, I think, they deployed it for Opus, so they had a little extra time to refine this. This is true. This is true. Hey everyone, I'm very happy. to do it. I need run, uh, but I just wanted to drop by, talk about 6 Soul, show a little inspection details. We will soon have much more information about checks. I would be happy. your feedback from about this. And we just published our inspection vibrations on every.every.to. You can see full check Opus vibrations with all details there. Um, but I'll leave you guys, to continue, uh, and I will move on to a few other things that I need to do today. But, glad to see you. See you guys. See you later. I can share. some things about coding. So, I coded with everyone models. And now, because they are both came out, we can get them compare. And I think, that's interesting because they launched them into one day. So, obviously they say that we are competing here , and we go face to face , or so, Opus and Soul have a similar class for by weight. So let's Let's compare them a little, because this is...cool. What I am tried, of course, this is encoding by with Compound Engineering , this is a plugin that in I have something to write about. code. So for me does it really matter whether he performs instructions? For me this is the most important thing, so what if you have an AI, which performs instructions, and you have a bunch of instructions, he will make the right one thing, ideally. In this the whole point. Therefore, comparing two levels efforts. It is interesting that coding won in Opus 5.5 over Soul. Opus and Fable Extra Astra were also there. So, complex problems with coding in Ruby with productivity, uh, with optimization, with all skills optimization vs., e- uh, yeah, things like that. Average mileage Opus 5.5 won, which is interesting. So it's not even extremely high or high, medium The run won. I I think, uh, Natasha, you also shared benchmark metric, where it was said that average actually has the best coding example somewhere. So average interesting mileage for coding. And there, where Soul 6 is really good, is it just faster? THERE ARE- uh, but he...also confused because Soul effort levels are not can be compared with Opus 5.5 effort levels, because they are so different. Hmm , I think that low speed levels in Opus 5.5 can be compare. So this is maybe...Oh, yes. So Anthropic shared this, obviously. Hmm, but yes, so for me medium is special option. He is very well. meets the requirements. Opus 5.5 fits all aspects of engineering complex elements, and Soul is having difficulties. Me had to restart it so many times. So for me Opus 5.5 is just more convenient in long-term operating mode plugin. Perhaps, they make these new ones, uh, new ones classifiers for security, but I It doesn't matter what exactly. Me need to be done work. And Opus 5.5 does this work. Although sometimes everything is faster , but it's not faster, if you have to to try again, if you are in a cycle. So, maybe I can to share yours screen and some images next to each other, because it's quite interesting— see the difference. Therefore, first...Of all, one next to it, it's Opus 5 against 5.5. So, this one benchmark, it's like I said to create 1- minute video about something like a lighthouse. You will see a little differences on the right with Opus 5, and on the left we have Opus 5.5 . You don't hear either. music. But this difference of approximately 0.5 steps, which for me incredible. Therefore, details on the left and right. Although right also enough impressive. And also camera movements really impressive, I think. You you can see how they are like this approaching. Good, obviously, okay, I I'll stop there. Obviously, 5.5 has more details. Uh, that's what Tidus says too. I very impressed with this. So I thought, "Okay, it is obvious that he better than five. He as good as Fable?" "Um, well, yeah. On this cozy island on the left Fable, on the right Opus 5.5. I I think they are did something similar. I I would say that Opus 5.5 has time of day that doesn't exist in Fable. And, perhaps, even a little more details. So, to me I like it, okay, this is a Fable class model. AND what if we compare it from Soul? Because Opus Soul has look like. Therefore, on the left Opus 55, on the right Soul 6. Well, she another. Um, this is something... like...when I I was looking at this. just laughed. Not that to make me want to laugh from the worst model, but it's just like that, I I mean, just yes, she is so striking another, and they are both on extremely high equal, and they should to be the same. AND, obviously they are both trained on 3D movements and other things, because it happened crazy step in because of how it worked. Um, one more thing, more a simple example is landing page. Both..." They are short , is it really so, they are both short, and they end in about 2 minutes. And you see, for I have this question like: How do I look at this? IN Why the difference? Difference not what it looks like better, because that's enough subjective, but that What I see as the difference is what's on the left Opus 5.5 low, and on the right 20.6 solo. I see that with Opus he feels a little more human. Hmm, and by that I mean meaning that there is less places. Just when I I read, I feel, what is more designed for me as a human being. For example, on the right, like this, this very eye-catching. It's like, maybe, to me must read this, and then this, and then like this, in different sizes. Hmm, if I scrolling here on the left, it's easier scan. For example, the text itself to me like a little more. Hmm, he's a bit more descriptive, and yes, it seems to leave more space for filling, and I like it I really like it. this model, and if you add more efforts. equal, he just makes others things a little more, but not exaggerating. He doesn't come in very often. extremely. For example, if I... For example, if I I will choose Open AI Max, it is looks cool, but also it's a lot of things . There are many things everywhere. So, this is mine. comparing these... Let's continue. program. yes tyler i think so Do you have something? related to Pokemon, what are you? wanted to share. I I feel that you received the most pleasure from all in this model during last week. Yes, let me do this. to lift One second . Hmm, good. So, here's what, if there were Pokemon real simulator...Hmm, me I am going to play in him on...Hmm, let's Let's start with 10 minutes. Actually, let's give it a little let's speed this up so that It was daylight. AND then we can see...we can to see the real ones Pokémon and people in actions here. Hmm, yes, that's it. just really...I had a great time with this...Hmm, I'm not big Pokemon fan, but My wife is. So, I I just wanted to create it for her, but it was simple interesting to watch, like pokemon and humans interact in this world. Environments, which they created. You you see them, you know, Pokemon have natural behavior, as if evolving. They engage in combat. with each other. They They attack villages. Some of them are not love people. Then like others really make friends with people. So I am sure that this one, probably has a high level of trust. What model is this Tyler? ? Because we are now we are talking about several models. Oh, this is Office 55. Office 55. Yes, we generally I like it better. Office 55 for coding and for these creative things. Yes, that was something. really fun, and that's created this really great report that I I will raise it, because this simulator in reality no...Simulator worked very, very hard long. So let me I just need a lift. one more...Come on let's see if shows this also Opus? You all show Opus things, Yes? Yes, I'm just showing Opus things. So, here it is. actually an artifact that he created later. Type... description, and again Yes, Opus did it all. . He added a video here for my viewing. I I thought that was it. cool how they do it done. They really are really showed cool examples that I could review. Hmm, and all this happened, while I was working. So it was interesting, where, it seems some people died in the cold weather. Hmm, and this showed how Pokemon interacted. Showing some evolution, I believe. Hmm , and simply looking at different habitat , where different Pokemon, probably captured or made us their own the main ones, hmm, territories. So this is it was really fun. Hmm, Ty Tyler, interesting, I wonder if, uh, what, When are you working? so more ambitious projects that take a long time running in the background mode, have you ever- Do you ever ask them, how did they do it, Yes? Did you receive more details, because when I look at this, I think that, you know, whether used game engine, or something made up independently? Hmm, yes, in I have all these... question, but I I wonder if Have you researched this? I risk giving up. a complete noob, no, no investigated. I just said, you know, I said it, I said it on Ultra Code. I put for himself and said, what I want to create a simulation of a world where Pokemon actually existed together with people. And that was it. I think it was very short task. And this It seems to have lasted a day and a half, and I felt that my computer is about to will melt, so I I was very nervous. I actually got some things from freezers and put them on Mac. I live in Hawaii. . So what are you doing? There can be around 80°. I know. I, you know, watched condensation, but I thought, "Dude, I I thought, I need a little cool it". And my Mac Studio has never been heated up so loud enough for me to hear , as he just rocks. So, um, that's it. It worked, dude. It worked. So you use all these fun...things. You used it for real work, uh, for example, with from a design perspective? Because you are engaged designer work. How are you? Do you think this is concerns design work as a designer? Not that so that I wouldn't be designer, this is the one model, or something like that Opus, Soul, Fable, what model, Astra, what model do you have? use if are you doing designer work now or will be use in future? You know, I still I use Grok for, so far, Grok 47, for real, uh, work and everything else, just because it is so cost-effective , and I feel like I can to manage it good enough to I could save, um, you know, a little bit of ours usage limits for the whole month. But, if I thought, if tokens were free of charge, not there would be doubts that I would use this model, and I even I can show you some, some examples. Opus 55, this model. Good. Opus 55, 100%. I will to talk only about Opus 55. Um, huge fan, but, let me to show you only a few things that, uh, Here are some entries, which I kept, um, in I have a generator. components. Or is it something like old Opus. But, some of you may see improvement in quality and design these things, where they are became even better. Ahem, Here's something interesting that I did. It's like a game machine that generates for me components interface. So, just very, very interesting things. And this there was also one hint. It was a more difficult clue. Hmm, I'm not exactly I remember what was there. written. But, you know, again, it was one hint. And I I think she is too. created this design tool, which I have to share with you here. I am going give me a second, to import for design. So, I I am going to import for those of you who ever heard of Origami Studio, this is tool for prototyping , very complicated in use. Um, he uses something called patches, and every patch is almost like function in the code, and I I'll just show you how. it works here, but this is not an application, over which I just work with in your own time. It's something like... So, this is not Origami Studio. This is your own version. this... own. This is my own. And you started with Opus 55. Yes , a little earlier. Maybe close four questions for all this. Without jokes. First request was like " create Origami Studio, friendly to artificial intelligence". And I think, that instead of explaining, how it works, it's very complex tool. I actually I use Cloud Code to control it in different terminals. So I will say add interaction with by swiping your finger. And what you will see... Yes, what you you will see here, this patches below. Each of these patches are something on like functions in code . And these are real layers design. So, you you can see the layers design on the left. You you can see the panel inspector on the right. I also had built-in assistant. So I can add your own Anthropic... key if I want simply to chat there. Hmm, and now it works in the background. AND in a moment we will see a bunch of patches below. And I I can show you, how exactly does it work. And I I hope and plan. use this for, um, my own actual work, because what you you'll see, this is what it is allows me exactly to adjust interaction like this so that she was based on nodes, and I can see everything on canvases. So, oh, here it is. Let's begin. So, you you can see, and let me just show here only nodes and show prototype So, this is a bit messy, but this created all these nodes here. Uh, give it to me. see if I can This is to restore order. And this it, just created this guidance function the order that is in Origami. I I think this is given. system. So let me show me that is happening. Um, let's see, if I drag. So, you can to see how all these the numbers change here. And I actually can set it all up. So, in Somewhere Here, there will probably be animation scale. Um, I see, what's on this badge " yes" is animation opacity, which I don't really I like it. So I just turn it off. AND I'll turn this one off too. AND this, if this, has disappear. So, you You see, this is a heart. disappeared here. So this is really simple way for me very fine-tune, to play and just have fun with these things like a spring here. Um, and elasticity. I I can do my best use them and to do really crazy. You see, how much faster is that passed? Maybe I I'll do it slower. Um, I'm not sure. would see the difference. It happened so fast. , but yes, it is the tool that I very, very excited. Um, and he is just like that perfect. So, if I am starting a new file. here, it is supplied with educational manuals, templates for beginning. Um, so, I I just can't. I know. I just affected. And then I can also like your thoughts. So, how are you? Do you really think you will you be this use? For, as with all these tools, there is always something on like: "Yeah, great divide this by X, and that's looks "amazing." Type: "Wow, wow, wow." "But will we use this daily? Similar to games, do we really need will like to play is it more than 2 minutes? Or it's really worth it of our time? Uh, how are you? Do you think you will? use this tool, develop it further and actually do his part of his workflow? For I am always skeptical. I relate to such things, because yes, it is cool that we can do them create, and this madness, but is it will help us in life? Will we actually something with this do? I think that this depends on the person to a person. What it concerns me, I very happy about this instrument, because I have been for some time. I'm trying to learn. use Origami Studio. As many times , I tried and gave up. They are not are updating their documentation, and to learn if you get stuck, you will have to go into Facebook group and to hope that someone will answer. And even when they answer, often they tell you how to do it, but that's not quite it reaches understanding because you you don't understand difficulties instrument, and, maybe something changed. They have... You know, they just have textbook, which one can to view, and even the manual itself is not relevant. So this is it, what I really want to learn, because, in my opinion opinion, many of mine favorite designers , hmm, those that I really like I admire it very much. good use this tool. So I am creating this. a tool for yourself, so that I can do it myself to learn use origami, and also to use artificial intelligence to I could tell him. something like: "Hey, add this, you know, interaction with by swiping your finger, or when I press on this, it has, hmm, you know, it has to scale to "full screen." "And I can to find out about what this is it, here. So, here it is. what I added here, right. It's like way for me find out what it is. So I hope that this tool will actually do me better designer. And who knows, maybe that's what what will I be use long term. Hmm, maybe not. I added some functions that I can like, for example, design vibrations. Showing something like holographic animations. So, maybe it will happen my tool design. Hmm, I'm not all. time in Figma, but no, I don't I think Figma is coming soon will disappear, but will Have you worked on this? How long is Opus again? worked on this, that Did someone ask? Hmm, I think, in general, with intermittently about 3 days. And it's not 3 days without breaks, because I worked. So first the question was quite long, it really described purpose, what is Origami Studio and how I wanted..." to it could have been manipulate by using the cloud code. It seems the second the question was on like: "Okay, I need to import real design with my production code". So, the application that You saw, I wrote using vibe- code using Astra, and then imported this SwiftUI app is in this canvas. Then the third and fourth the questions were: "I I want to have the opportunity. develop new ones designs, something on like this makes paper if you "familiar". And then, it seems the last one the question was: "I I want to add statuses. loading". So I created a holographic animation. She took...Actually I had to do a little develop...me had to provide some links and describe it, but it made this holographic thing at once. And also, there is some states download for nodes that are not were displayed. So this is error of life demonstrations. But I I will say that she appeared from intermittently for 3 days. Perfectly. Uh, you another name was called. You also said Grok-4-7 , which also came out yesterday. So actually There are three models. And Mike, I I know...you too spent some benchmarks. Me too. So, maybe we can to talk briefly about how Grok-4-7 compared to, uh, Opus 5.5, and, uh, I have just say Soul, Luna, and, uh, yeah, and then, maybe we can to complete a straight line broadcast and return to using. Yes, I, uh, let me. I want to share with you. you some things. So, I was very, uh, I think we were talking. about that, Tyler, but, uh -uh, I was very surprised Grok-4.7, so what, uh, Grok-4.6, uh, great model, uh, and, surprisingly good. Better than I expected, isn't that right? And, uh, when we started testing 4.7, there was a feeling that he was released hastily. Uh, actually, at least in my benchmark, this there was a real regression. And I don't know if it's is due to the fact that, maybe it was just many mistakes or things that they it was necessary find out. Uh, I know, what did you say that saw how he improved over time the previous week. For example, which one do you have? had experience with him, Tyler? And I will show, and I will show you. Here are some examples I have seen. Oh, the first few days were heavy. It was it's definitely hard, and I think , which, you know, maybe they were not ready before we can do it protested, but to At the end of the week he became much more reliable. But in the beginning there were some things when he a little bit of a letdown or is it there were hallucinations. There was a lot of, um, I I think there are problems with their servers on At that time, they probably very, very hard worked on it, while they give it to us gave. There were some downtimes past cursor a week, it seems, once, yes. Yeah, so there were some problems. But you said that use it and you like it, Yes? So how are you? use it today? Well, I would say that I I still understand how I like it to use it , and I think that this the first time since for some reason I returned to IDE cursor over window agent. It seems that he works a little rather. Maybe this just my imagination, but, it seems he worked a little faster in IDE. Especially during testing I decided to stay to work there globally, not in cursor projects that I really, really like it. I like it. But, I probably don't have one. the same success as in Grok 4.5 and especially Grok 4.6 within projects cursors, what are they, hmm, my favorite way. This is one of reasons why I am like this I often use cursor, because, um, cursor projects allow me to have one chat for everything, what I'm working on, and he creates all these subagents in one place, and I can have another project under him. So, it really is allows me well to do a lot tasks simultaneously, but still to come determine if Grok 4.7 really good orchestrates these projects with other subagents. But in my previous configurations to me I really liked having Fable 51 low or Fable 51 medium as an orchestrator and Grok 46 like all subagents. This worked very well, very good It It worked so well. So I...I hope I I don't know. I hope on Grok 47, and I, you know, I used it today and yesterday. IN I was not there. problems. For example, during testing at first it was a little hard. They were interesting moments. He sent personal to me message, you know, one of my personal messages, not telling me nothing. AND I understood later. I thought, "Oh my God, me thought: "Marcus, I'm so sorry. "I said: "Marcus, I "so sorry." And I think, that some things that I said, perceived very literally. I I think, for example, that I I'll say something like:" Let's prepare this is before merging with main message ". A then it's simple combined this with main message and sent a personal message to Marcus , and I said, "Oh, no." As if a big no to everything , where we, we, we, always, people, we always do this, let's do this, let's do this, let's do this, we unite here. But yes, as I said, with during the week it improved. Hmm, me really starting to like it a little more than 4.6, as of this morning. Before I came, I worked on shipment PR and not had no problems. This is really feels like improvement, but this one of those improvements, where Grok 4.7 is not one-off performance. That's why it's hard for me. evaluate it. It really so, and when I I use it. for real work, almost 100% of mine real work—this using Grok, so , hmm, it's very difficult for me to say: "Oh, or noticeably better? No, but it seems he works great and I don't I plan to move, but, friend, I would like to, just so you know, I have some comments regarding Opus. Hmm, Hmm. What did you find, Mike? during his testing? Yes, that's why I feel what, perhaps, we need to spend retesting. Uh, if that's the case changes a lot, Yes? But, um, but I found that some the tests passed quite well good. So, it was toolbar MPS. This is a different design, than what I saw in other models. Ahem, that is, if you looked, let's say, on Fable. So, you were looking at Grok ? Yes, this is Fable. And then, uh- uh, this is Grok 4.6. And, uh, me I really like this one. title, and design. THERE ARE -uh, I was very satisfied 4.6. Uh, me just for comparison, uh, this is Soul 5.6. Uh, and, uh, you know, as a new version that is not handled it so well work. Uh, maybe like Opus 5. Yes, this one vibe, coded something like black, you know, dark mode. Uh, but, uh, but yes, Grok 4.7 dealt with it normally. But just to give you an idea about my problems with him. Many tasks he just doesn't did it, didn't you? So, uh, he didn't give out, uh, final result. So there there were some frank errors. And by the way, I exceeded these task three or four times. Hmm, ah some of them ended, that makes you think that this not just, uh, something like that like a seat belt. Uh, let me. to look at PowerPoint tasks. Uh, Give it to me. Uh, that was it. is that it? Yes, it was. something like...I trying to find specific example, but, uh, II...Yes, it is not what I thought. I think I have it. rewrote. Uh, but, uh- uh, yes, I found that this it was something like... a bit messy. Uh, let me just look at example description. Yes, Here's a good example. The inscription here, uh, it was something like, as if a few steps away back, uh, relatively, uh -e, previous models. Uh, this is something... like, uh, this it read like, uh, probably a to-do list, not an article. Uh, this is there was one problem, but then, um, a lot things that he just didn't like it, just didn't like it, they are not very good did their job work. For example, the design was not like that good etc. So this is it was similar to a small step back, and since he was unreliable, he became something we have become less interested, but I hope that, like me, I have very high hopes. . I hope that, you know, problems with reliability will improve. Yes, I did some AB- tests. The last one is Grok 4.7, on the right - Opus 5.5. I I mean, he not as great as 3D . Um, and on the last one Grok 47, this is a video. So, in video really trying to show , can he do very good planning, can he does iterations etc. So, left-Grok 47, right-56 Luna. So, if you look at two of them, consists of the impression that the 3D action of Grok 47 is more like Luna. Luna looks a little better. So, with 3D- prospects, it's not very impressive. Hmm, with coding, yes, in LFT not very good either it worked out, and maybe it is because I I use the working one. process. So, obviously, I use this in engineering of complex elements as a worker a process that seems to have follow all the steps, and if he doesn't do it all the steps, the harder make it work. But I liked it. Force, or me like, 4 6; uh, it's just a job horse. But also this it becomes very difficult, because Luna is a worker horse, Sol—working horse, Opus 5.5—working horse. There are so many workhorses, and so many workers horses become so good, that they are better than stable. What is simple? madness, and simply that all this the landscape is so strong changes like Sol 6, much better than 5.6. But there is also Opus 5.5. And he it just changes everything. Ahem, if you are interested, check out Opus 5.5, but also get to know everyone other models. I think for writing, especially then, to me Sol will really like it. Also check out Opus 5.5 on low efforts. I think this is It's worth a try, because if you use high efforts, this may seem a little slow, but you don't very necessary tardiness. You you can just to put low. Ahem, and yes, like any any other advice or tricks that are worth it learn before we let's hang up, Tyler, Mike? I think the main thing is, What to think about , is the rate at which these new ones are coming out models, and how difficult they are to protest and see the differences between them. We already approaching points when we largely oversaturated many tasks. If you can do it think, you can to do it. Um, and this begins to rise more personal preferences, uh, that is I don't know about you, Karen. and tyler, but, uh, me started less to worry about "whether" this is the best model in world, uh, highest math or something like this, and ", and, uh, instead, like, or I like it. the design that she suggests, whether I like writing? And I, I-uh, I felt to be more comfortable subjective. And I use all of them. models. I literally I use all models. It's not that I don't I use one instead of another, for example, if I I create 3D games, I, probably not I will use Grok -4-7. I probably I will use Opus -5-5. Um, but me too I like it, I love it. Cursor. Cursor is mine addition. I I use all your tokens in Cursor, because the Cursor project excellent. Er, and Codex also wonderful, and Claude catching up with the application. He's not that great, like others. Well, but I I use all programs, everything. Hmm, and yes, if you you need to choose, I hope you are doing something. learn. Well, you don't. need switch because they all very good. Hmm, but yes, Opus-5-5 is special option. I really...uh, he deserves special mention, for example, if for I have a winner, then this is exactly the opus of this weeks into launch. Yes, definitely. And I I also feel that just broke up with Opus . So, I...I'm like, fully returned. This is really, really It's funny how fast I am. returned to Opus, but I agree with Kieran. You know, I I use all models, uh, I don't see, so that I give up Cursor and Grok for real work. I am not I'm going to get rid of your Grok bots. Uh, me I love them very much. Of all in Blender, I I really love it. use Astra with computer management, using Blender, unlike MCP. I am not I can get rid of this. And this thing with Opus is just ...uh, she's adorable. My wallet hurts. Uh, but this is it, this is it it, I can't...I I just can't. to express in words, how wonderful Opus is 55. Sounds good. place to finish. Perfectly. Thank you all. Everyday is the only one subscription that you needed to stay on the edge of AI . We said it. Oh, yes. Subscribe. Yes, we are so terrible speakers. Everyday...yes, check out our projects. We have... Yes, our checks vibes will arrive today. Or they already published, or they will arrive. We we also conduct comparison there. We we will share in greater numbers on X, LinkedIn and others places, but thank you, why are you staying with us? , and see you later next time. To everyone have a good week releases.