Livestream coverage of OpenAI DevDay 2026, presented by Every. # Video Identity Observations: "LIVE: OpenAI DevDay 2026" ## Observable Presentation - **Live event coverage format**: The video is on-the-ground commentary from an AI industry conference, with conversations in hallways and impromptu interviews ("Simon, would you like to do a short interview in a vlog?", "I think there's a Q&A session with Sam if you want to go in and check it out"). - **Creator holds event access**: References to special credentials suggest press or creator access — *"Let's see if they'll let me in with my creative access... I don't know if this is for creators or for the press. It could be both."* - **Access to notable figures**: The creator mentions a planned interview with a prominent executive — *"So, I'm interviewing Sam and I'm going to ask them"* — indicating insider-level access to AI industry leaders. - **Technical, evaluative analysis**: The commentary benchmarks and compares AI models hands-on, e.g., testing a pelican render across model versions ("Sol 6.1," "Soul 5.6") with a comparison grid, and discussing pricing tiers (*"a new plan for $500 a month"*, *"GPT-6 had a 50% discount"*). - **Deep familiarity with the AI ecosystem**: Casually references labs, products, and pricing history — DeepSeek optimization papers, Cerebras, Gemini diffusion, Anthropic, App Store rankings — and personal spending (*"right now I pay $200 a month to OpenAI and $200 a month to Anthropic"*). ## Intended Audience - **AI practitioners and heavy tool users**: The discussion assumes familiarity with developer tools, APIs, model benchmarks, and subscription tiers. The creator critiques announcements from a practitioner's lens (*"everything is a development tool for developers... isn't he the general agent?"*). - **Viewers tracking the AI news cycle**: References assume the audience knows recent releases (*"Didn't they announce it last week or something?"*) and past events (*"last year's DevDay, in retrospect, was a disaster"*). ## Claimed Scope / Self-Description Evidence - The creator self-positions as an AI model evaluator via hands-on testing, describing his own usage patterns (*"it's actually become the main way I use ChatGPT"*) and running live model comparisons on camera. - The recurring **"Pelican Bench"** motif — described as *"one of my favorite ways to understand the progress of AI"* and *"the stupidest of all reference points. And yet, somehow, it offers value"* — serves as a distinctive, personalized evaluation lens tied to the channel's identity. ## Distinguishing Claims vs. Demonstrated Patterns - **Claimed by creator**: interview access to Sam; creator status of guest Simon ("creator of Pelican Bench"). - **Demonstrated by material**: live model benchmarking, insider-adjacent access ("creative access"), sophisticated industry commentary, and personal paid usage of multiple AI subscriptions. # Delivery Observations — OpenAI DevDay 2026 (Vlog/Interview Format) ## Overall Communication Pattern - The delivery is **conversational and interview-driven**: much of the video is the host speaking with Simon, not presenting solo. The structure alternates between casual hallway chatter, a walking segment about pelicans, and a seated debrief on announcements. - The host frequently uses **short verbal acknowledgments** to keep conversation flowing: *"Yes, yes."*, *"Mhm."*, *"I see."*, *"Interesting."* These appear dozens of times and function as turn-taking and encouragement devices. ## Recurring Vocabulary and Phrases - **"Interesting"** is the most repeated evaluative word — used to accept information without committing to judgment: *"That's interesting."*, *"Very interesting."*, *"it's interesting, but it's not good."* - **"Yes, yes, yes"** appears repeatedly as an enthusiastic agreement cluster: *"Yes, yes, yes, yes."* - **"I don't know" / "We'll see"** is a recurring hedging response when asked for predictions: *"Do you think they're succeeding? Don't know."*, *"We'll see."*, *"I don't know how much I trust Meta with all this."* - Superlatives and exclamations appear in bursts, mostly around pelicans: *"Oh my God, I love DevDay."*, *"Brilliant."*, *"Incredible!"*, *"What a day."* ## Teaching and Explaining Habits - The host explains products through **concrete dollar figures and feature comparisons**, e.g. *"six to eight times faster, or eight times faster, six times more expensive"*, *"I pay $200 a month to OpenAI and $200 a month to Anthropic"*, *"a new plan for $500 a month"*. - He frames analysis with **historical context**: *"last year's DevDay, in retrospect, was a disaster"*, and references to past models (*"Google had its Gemini diffusion model about a year and a half ago"*). - **The Pelican Bench demonstration** is a signature teaching device — using an absurd visual benchmark to evaluate image models: *"It's the strangest thing. It's the stupidest of all reference points. And yet, somehow, it offers value."* He walks through a comparison grid methodically (*"Here is the comparison grid... that's the Soul 5.6, it has the right all-around bike shape"*). - He introduces guests by **foregrounding their credential**: *"Simon, who is a man with many talents... he is the creator of Pelican Bench."* ## Interaction Habits - The host **asks short, direct interview questions**: *"What do you use daily now?"*, *"Are you going to buy the $500 plan?"*, *"Any other... is there anything else that excites you?"* - He **manages logistics verbally and collaboratively**: *"Simon, would you like to do a short interview in a vlog?"*, *"Give me about 30 seconds to get to my things."* ## Running Theme - **Pelicans** recur constantly — as a shirt (*"I'm wearing a pelican shirt"*), a hobby (*"I have a camera to take pictures of pelicans"*), a benchmark name, and a planned task (*"I have to take five photos of Calvin"*). This is a deliberate throughline, not an isolated mention. ## Uncertainty Notes - The transcript appears to be an auto-translated recording; some quotes may be distorted by translation, so exact phrasing should be treated cautiously. - The *"bureaucrat"* utterance at [227] lacks surrounding context, so its meaning is unclear. - Speaker attribution is ambiguous in the early hallway segments — it's not always clear which interlocutor says which line before the formal interview begins. - The host's identity is never stated in the transcript; he is identifiable only by role (interviewer, pelican photographer). # Observation Record: "LIVE: OpenAI DevDay 2026" ## Overview A vlogger-style walkthrough of OpenAI DevDay 2026, featuring informal conversations with attendees including a person addressed as **Simon** (creator of the "Pelican Bench" AI benchmark) and a brief exchange with someone addressed as **Dan**. The transcript is fragmentary and conversational; speakers are identified only through direct address and contextual cues. --- ## Positions on the "Dots" Personal Agent Launch ### Early conversation (unnamed speakers) - One speaker reports their personal experience with **Dots** has been *"quite tough so far,"* while liking new things generally. - Another expresses distrust of **Meta** regarding its offerings: *"I don't know how much I trust Meta with all this."* They add the qualifier: *"I think they'll smooth things out and stuff, but it's kind of... like beta stuff? I guess. I think they're moving very fast."* - Both speakers criticize the format of **20 announcements**: *"The 20 was a bad idea"* and *"fewer updates per week would be fine."* - A speaker argues Dots is framed as a developer tool when it may be a general agent: *"everything is a development tool for developers. And I ask myself, isn't he the general agent?"* They note non-developers might use it, but qualify: *"Perhaps not entirely, since almost all developers still are."* ### Exchange with Dan - Dan complains about a cloud release: *"This cloud code is garbage."* - Speakers debate the claimed "20 seconds" launch pace; one claims to have counted them. ### Later assessment (Simon / interviewee) - **Dots is interesting but not good**: *"If you compare it to Tag, for example, it simply doesn't have as many features. It's simply too soon."* - He notes the development team is *"progressing very quickly."* - He raises an unanswered question about a **decision API** — possibly announced the prior week — which the interviewer says he will ask Sam about. --- ## Positioning Analysis (Simon, in the vlog segment) - What he most hoped for happened: personal agents (e.g., **Meta Muse**) *"taking off like wildfire"* — at the top of the App Store for two weeks. - He found it notable that OpenAI still presented **Dots** as a developer tool: *"His great example was doing a complicated database migration for your team... it was still funny that they didn't really focus on the fact that it's your life partner."* - He contrasts this with **Muse and Instinct**, which *"are targeting the average consumer,"* and notes Dots benefits from **ChatGPT distribution**. - Personal usage claim (with qualification): Dots has become *"the main way I use ChatGPT"* despite missing features and bugs — for roughly *"two or three or four days."* --- ## Model/Rendering Commentary (Sol 6.1 pelican render) - Simon critiques a pelican render: *"The Sol 6.1 has a slightly thick neck"* and *"that bicycle frame is garbage."* - He notes variability: *"you roll the dice every time and every now and then you get a bad bike,"* while observing the **Soul 5.6** version *"has the right all-around bike shape."* - An unnamed speaker notes the legs were positioned correctly. --- ## Ultra Fast Model & Pricing Discussion - **Ultra fast** is speculated to be Cerebras-based (*"they didn't mention it, but it's six to eight times faster... six times more expensive"*). - A new **$500/month plan** was announced. Simon is hesitant: *"it's dangerous, right?"* He currently pays **$200/month to OpenAI and $200/month to Anthropic**. - Simon on ultra-fast usage: demos (voice-controlled screen editing) are *"fun, but I didn't feel like it was something I would use regularly in my day-to-day life,"* though others *"currently rely on the quick mode."* - Cost caution: *"unless you have a hundred million dollars... you'd better wait a couple of months until they decide to make it cheaper."* - Recent price cuts noted: **GPT-6 at 50% discount**, **Luna half price**, **Opus 5.5 down ~20%** — *"the first time in a long time."* - Simon relays a theory (attributed, not asserted as fact): *"There is a theory that this is because some Chinese AI labs have been publishing new optimization papers"* — specifically that **DeepSeek** published optimization work *"and suddenly Open and GPT and the GPTs get faster like a month later."* --- ## Retrospective on DevDay & Distribution (Simon's closing remarks) - **Last year's DevDay "was a disaster"** in retrospect: launched items like *"the agent, drag and drop... things from Sora... They no longer exist."* - This year feels like *"safer and more obvious bets."* - On the personal agent (called **"Taclon"** per the transcript): *"The big question is always distribution. Now they have a market. They keep trying. I'm very curious to know if it works."* - He cites Sam's acknowledgment: *"I know we said we were going to try to solve the distribution problem."* - Unresolved: *"it will be interesting to see if that works or not."* --- ## Miscellaneous / Non-Substantive Notes - Recurring **pelican** theme: Simon wears a pelican shirt, carries a pelican camera, created the "Pelican Bench" benchmark — described by the interviewer as *"the stupidest of all reference points. And yet, somehow, it offers value."* - Venue complaint: fencing preventing access to the water's edge where pelicans appear. - Speaker daily-driver claim: still uses **Chrome OS** and bought another Chromebook. - An unnamed model ("based on Luna") reportedly *"performed better than Jev in some tests"* and is multimodal — presented as preliminary initial tests only. --- ## Unresolved Tensions Preserved - Dots is simultaneously called promising (already the speaker's main ChatGPT usage) and *"not good"* / *"too soon"* — both claims come from Simon across different moments. - The positioning question (developer tool vs. general/consumer agent) is raised but never resolved. - The Chinese-labs price-cut theory is explicitly relayed secondhand, not endorsed. - Whether OpenAI's distribution bet succeeds is left explicitly open. # Speaker & Provenance Record — "LIVE: OpenAI DevDay 2026" ## Identified Speakers & Roles | Speaker | Role | Basis | |---|---|---| | **Unnamed host/vlogger** | Creator / interviewer (appears to be the video's creator, doing on-site vlogging) | Conducts interviews, plans a vlog with Simon, states "I'm interviewing Sam," mentions "creative access" badge | | **Simon** | Guest — identified as **creator of Pelican Bench**, described as "a man with many talents" | Explicit in-transcript introduction | | **Dan** | Brief second guest / acquaintance | Greeted by name ("Hi Dan"), short exchange | | **Sam** | Referenced third party — OpenAI figure (implied); host plans to interview him | Transcript mentions "Q&A session with Sam" and "I'm interviewing Sam" | | **Randy** | Referenced by name by the host ("Randy, what other interesting things?") | No further identification available | > **Note:** No curated metadata was supplied; all identifications come solely from the transcript, which is partially garbled/transcribed with errors (e.g., "Dots," "Sol 6.1," "Taclon" may be mis-transcriptions of product names). ## Referenced Sources & Third-Party Claims - **Products/companies discussed:** "Dots" (announced personal agent at DevDay), Meta ("Muse," "Instinct"), ChatGPT, "Sol 6.1" / "Soul 5.6," "Luna," "Opus 5.5," GPT-6, "Astra," Cerebras, Google Gemini diffusion, DeepSeek, Anthropic, App Store rankings. - **Claims made in conversation (unverified):** - Muse "at the top of the App Store for 2 weeks." - "Ultra fast" mode is "six to eight times faster, six times more expensive" — attributed to what "they said," Cerebras guessed but unconfirmed. - New $500/month plan mentioned. - Price cuts: GPT-6 50% discount, Luna half price, Opus 5.5 ~20% — presented as recall, not cited. - Theory that Chinese AI labs' optimization papers (DeepSeek cited) prompted recent price drops — explicitly framed as "a theory." - Last year's DevDay launches "no longer exist" — speaker's retrospective assessment. - **Host's stated plan:** to question Sam about the "decision API" ("Didn't they announce it last week?") — answer not present in transcript. ## Demonstrations Shown In-Video - Simon generates a pelican render with "Sol 6.1"; comparison grid shown against "Soul 5.6" and "Astra" using Simon's own credits. ## Promotional Segments - **None identified.** Content is conversational/vlog-style commentary; no sponsor reads or product promotions by the speakers detected. ## Availability & Access Limits - No paywalls or comment availability stated in transcript. - Host notes uncertainty about press vs. creator access: *"I don't know if this is for creators or for the press."* - Event-floor restrictions mentioned (fencing near waterfront) — situational, not content access. ## Uncertainty Flags - Transcript appears machine-generated with degraded names ("Net to Muse," "Cloud Tag," "Open Floor," "Pelican Bench" partly inferable); product-name accuracy cannot be confirmed from transcript alone. - No speaker full names, channel identity, or affiliations available beyond what is quoted above. Look at what's really happening. But, oh , incredible! Brilliant. What did you think? I liked it . I mean, I think... I'm curious to know if you tried Dots. My personal experience with Dots has been quite tough so far. Really? Yes. I like new things so far. That? I like the new things so far, but I don't know how much I trust Meta with all this. Yes, that's a problem. I mean, I think they'll smooth things out and stuff , but it's kind of... like beta stuff? I guess. I think they're moving very fast. Yes. I did n't get all 20. I don't know what the 20 things were. I got like 10 of those. Yes. The 20 was a bad idea. Yes. And even like I was saying to my friends yesterday, fewer updates per week would be fine. I agree. I agree. The thing about points is that, basically, everything is a development tool for developers. And I ask myself, isn't he the general agent? That's the... Yes. I mean, this is a development tool for developers. Yeah . They've barely mentioned everything. Perhaps non-developers could use it. Perhaps not entirely, since almost all developers still are. More than 90 percent of each company are developers. Simon, would you like to do a short interview in a vlog? Clear. Let's do it. We can go out. Yes, yes. Do you want to do it? Give this to whoever you're talking to. It's your turn. Okay. I don't feel it. Hi Dan. Hello how are you? I'm glad to see you . I'm going to do a vlog with Simon, but I want to catch up. Okay. I wish they would do the 20th release, which is in the cloud, 2 hours ago. I had to create my life record system in code in the cloud. Okay. This cloud code is garbage. Have you tried dot? No, yes. I've tried. Who was that in my thing? Yes. That part of the ads is pretty crazy. Uh-huh. Difficult to... Well, they said they were going to launch in 20 seconds. I know. I do n't think they had 20 seconds. I wasn't counting them. I counted them. In other words, it depends on how you tell it. For example, what gets grouped and what doesn't, but there are about 20. That's fine. We have the number one spot on the App Store, it doesn't matter. Oh, do you think they're succeeding? Do n't know. You were number one on the App Store for the last week and a half. The same. What are your...? I might ask you, but what are your ...? What do you use daily now? What do you use most often? Most of them are Chrome OS. Chrome OS, yes. Chrome OS Remote. Um, I'm still using Chrome OS. I just bought another Chromebook because I like Chrome OS ... It's great. It's great. So I've gone back to Chrome OS. Oh. Can I give you this? The clip? Oh. In fact, I'm wearing a pelican shirt . Oh, I love it. I'm going to need... And I have this camera. And I have a camera to take pictures of pelicans. They won't let us near the edge on either side. Really? Yes, as if they had a big fence preventing you from reaching that part. That's annoying. It's really annoying. Are there pelicans here? Oh, yes, of course, yes. Piles. Yes, if you're lucky, you might see one around the corner. You know, I think that thing over here... What? Can we go this way? OK . Alright. Clear. Bureaucrat. Well, at least it's a good day, isn't it? We have that answer for Net to Muse and we have that answer for Cloud Tag, right? What's the other one? Yes, yes, yes, yes. I've been trying out Dots, it 's interesting, but it's not good. If you compare it to Tag, for example, it simply doesn't have as many features. Mhm. It's simply too soon. They are progressing very quickly. What I would like to know is that decision API. Didn't they announce it last week or something? That's what I'm going to do. So, I'm interviewing Sam and I'm going to ask them. Yes. Yes. Because it's a really interesting question. Because I could talk to people who did it. Yes. It's not actually made. Elaborated. Yes. It's based on Luna and on us. Have you tried it yet? No not at all. In our initial tests, he actually performed better than Jev in some tests. I mean , given the strength of the Moon, that's not a... Yes. And it's multimodal. Falcons. \[ \_\_ \], Hawks. Very well, we have... We have pelicans. Look at that. Look at this camera. Lovely. \[ \_\_ \]. For fast. Oh my God, I love DevDay. Do you see the pelicans? They're right there. I think so , right? That's it...That's it...Those are...Those over there, the ones that fly? So, the question is... Here we go. Yes. Great, can we show some shots of pelicans? Can we see them? I can...Give me about 30 seconds to get to my things. He's not my best friend. Actually, they're a bit far away. Let me see. Well, we're here with Simon, who is a man with many talents, but one of the most important is that he is the creator of Pelican Bench, which is one of my favorite ways to understand the progress of AI. It's the strangest thing. It's the stupidest of all reference points. And yet, somehow, it offers value. Yes, indeed . Yes, indeed. And we just saw, we just saw some live pelicans flying. A pelican...Oh, that's not a pelican flying. Oh my God. Yes, so you're not just a fan of pelicans in theory. You're a fan of real pelicans in the wild. In the Bay Area. The Bay Area is a great place to see pelicans. Yes, I'll have a picture of a pelican... What are they? I'm curious. There it is. Our photo. There it is. The photo of the pelican right there. It's not a great photo, but it's them flying. That's pretty good there. There you have it. Oh. Look at that. Absolutely easy. I love it. I love it. So, what do you think? Well, that's interesting. Well, what I most hoped for has happened. I thought there would be some kind of personal agents, right? Like Meta Muse, who are taking off like wildfire. It's been at the top of the App Store for 2 weeks, I think. We've known that since January , right? Open Floor in January. One second. Okay. This category of AI personal assistants, the opening version of AI is this thing called Dots that they just announced. I found it interesting that they were still presenting it as a tool for developers. His great example was doing a complicated database migration for your team. I suppose it's partly because this is a developer conference, but it was still funny that they didn't really focus on the fact that it's your life partner. Yes, yes. They talked about it a little, but all the examples were still very technical. I was watching, I think during the presentation, well, specifically in the rehearsals, and they're still trying to figure out how they want to position it. But it seems to me that, with Muse and Instinct, they are targeting the average consumer. The advantage of using such a tool is that it is already integrated into ChatGPT, a platform that people use to work and develop projects; Therefore, they already have that distribution that ChatGPT offers, yes. But it's actually become, I mean , I think it doesn't have every feature, it has some bugs, but it's actually become the main way I use ChatGPT, which is really thoughtful. Probably like two or three or four days. Okay. Yes. So that's it, okay, that's interesting, right? What else? What's up? Randy, what other interesting things? More pelicans? I just want elegant pelicans. I'm sure that... I mean, we could go out if you want when we're done. Okay, that was pretty good. We are seeing that Simon created a render of a pelican using the new version of the Sol 6.1 software. The Sol 6.1 has a slightly thick neck. A slightly thick neck, yes, yes. In fact, I think the bicycle is rubbish. That bicycle frame is garbage. Oh, interesting. Now, this is the default level of thinking, and I haven't checked what your default level of thinking is, but it's a surprisingly deformed bike. How does it compare to other OpenAI models like Astra and the previous Soul? Of course . Let's see. Where is my credit? Here we go. Here is the comparison grid. If you look at these, uh... Yeah, that's the Soul 5.6, it has the right all-around bike shape... Oh, once we went down to the Soul 5.6, that's when we got a bad quality bike. I see. So, it could have been... I mean, you know, you roll the dice every time and every now and then you get a bad bike. The legs are fine. The legs are on both sides of the frame. But yes, it is interesting. So, um...Simon, any other...? Aside from that, is there anything else that excites or particularly interests you ? Things, the ultra... Ultra fast, yes. Fast, they say it is... I guess it's Cerebras, so they didn't mention it, but it's six to eight times faster, or eight times faster, six times more expensive. Yes, that's what they said. Yes. And they have this new plan for $500 a month, I know. If you use Ultra a lot, are you going to buy the $500 plan? I'm from...Yes, I could... Yes, try it. Yes. If I try it, I'll get on, it's dangerous, right? Yes. Yes. Yes, yes. Yes, I mean, right now I pay $200 a month to OpenAI and $200 a month to Anthropic. And that's it...Yes, yes. I'll think about it . Yes. But I want to experiment with models because there's been a whole issue with model performance in general. There's Cerebras, then Walk, and then Google had its Gemini diffusion model about a year and a half ago, which was incredibly fast, and they didn't do anything with it, which was always puzzling. Interesting. So, what do you think you'd do with Ultra Fast when you finally get to play with him? So I created demonstrations in the past directly against Cerebras, where the really fun part is exactly what they show there. When you connect a voice model, you can literally tell your screen, "Show me a graphic like this. Now add this to the house. Now add four wheels to the bicycle." Whatever. That's super fun. Mhm. Again, I played around with those demos and they were fun, but I didn't feel like it was something I would use regularly in my day-to-day life . Yes. Well, I don't know. But at the same time I know that there are people who currently rely on the quick mode because it's only for programming. I always use the quick mode, yes. But I don't use the ultra-fast one. I've noticed you spend a lot of tokens very quickly. I think it's the kind of thing that's fun to play with, but unless you have a hundred million dollars or something like that, you'd better wait a couple of months until they decide to make it cheaper, that would be my guess. Yes. We'll see. I mean, the price we had was so low just a week ago. I know. GPT-6 had a 50% discount. Luna was half price. Yes, yes, yes. Opus 5.5 went down, how much? About 20% ? I know, it's the first time in a long time. I feel like they didn't lower the price for a while. Like OpenAI and Anthropic, they are now doing it. Oh. There is a theory that this is because some Chinese AI labs have been publishing new optimization papers. And I didn't know that I hadn't heard that. Yes. I believe DeepSeek, DeepSeek definitely published an article describing some of their optimization. And suddenly Open and GPT and the GPTs get faster like a month later. That's interesting. Very interesting. Okay. Is there anything you'd like to leave us? Well, last year's DevDay, in retrospect, was a disaster. If you look at the list of things they launched last year, it was stuff like the agent, drag and drop, all that stuff. They no longer exist. And things from Sora. They no longer exist. This year I get the feeling they're going to make safer and more obvious bets. Yes. After what we just saw, who wouldn't want something a little faster? It looks like the personal agent is going to work. They called him Taclon. The big question is always distribution. Now they have a market. They keep trying. I'm very curious to know if it works. What happened to that? Well. Yes, this is another attempt. And Sam nodded. He said something like, "And I know we said we were going to try to solve the distribution problem." Well, it will be interesting to see if that works or not. Okay , it's always a pleasure talking to you. Yes. Good conversation. Brilliant. Thank you. I think there's a Q&A session with Sam if you want to go in and check it out . Yes, okay . You don't have to do it. OK. The truth is that I have no plans for today. I have to take five photos of Calvin. Perfect. What a day. Let's see if they'll let me in with my creative access. What a day. Oh, I don't know if this is for creators or for the press. It could be both. Don't know. We'll see. I had to look myself up in the database. Believe. Really? Painful. Hello. Thank you. What are you doing?