The creator live-streams their writing process, using ChatGPT Voice to assist in drafting a history of Codex. # Video Identity Observations ## Observable Channel Positioning - **Creator self-identifies as an AI journalist/coverer of OpenAI**: "the first time since I started covering them back in the days of GPT-3" and "I'm currently writing the definitive history of codecs [Codex] based on a series of interviews with people at OpenAI." - **Affiliated with a publication called Every**: "subscribe to Every. Every is the only subscription you need to stay at the forefront of AI." Also references prior work: "we just did a paper about how software is changing... every.pa openai-infrastructure... This is by Laura Entis" — indicating a team-produced editorial outlet. - **Demonstrated scope**: AI industry analysis (OpenAI, Anthropic, Codex, Claude Code revenue figures), technical explanation of "compute during testing" / reasoning models, and **practical AI workflow tutorials** ("how you can use voice mode in... Codex and ChatGPT for work"). ## Intended Audience - **AI industry professionals and enthusiasts**: viewers ask "when the GPT voice API will be released"; the creator addresses OpenAI staff directly ("If you're from OpenAI and you're here, hello"). - **Productivity-focused AI users**: the pitch for subscribing includes "guides, camps, courses... If you want to learn how to use voice mode in Codex, we will of course cover everything that comes out: new models, everything from OpenAI, everything from Anthropic." - Live audience grew to ~1,600 viewers, suggesting an established niche following. ## Self-Description Evidence (Creator Claims) - **Writing process transparency as content**: "one of the things I really enjoy is watching other people's work" — work-in-progress livestreams are positioned as part of the channel's appeal. - **Future content promises**: a Codex essay, "a detailed report on how Codeex works, it will be released in a couple of weeks," plus recurring "guides, camps, courses." ## Demonstrated (Not Claimed) Patterns - Deep access to OpenAI sources ("I interviewed a bunch of people at OpenAI"), exclusive quotes (Alex, Greg, Thibaut, Mia), and citation of primary materials (DeepSeek R1 paper, OpenAI's "Learning to reason with LLMs") demonstrate an **insider-analysis editorial style**. - Heavy use of literary framing (Bill Bryson, Annie Dillard) shows a **narrative long-form essay approach** to technical subjects. ## Caveats - All quoted figures (Codex 5M users, Claude Code $2.5B) are from the creator's own draft and cannot be independently verified from this material. - Claims about Every's offerings (camps, courses, guides) are promotional statements, not demonstrated in this video. # Delivery Observation: "WATCH ME WRITE LIVE WITH CHATGPT VOICE" ## Overall Format - The video is a **live voice-driven writing session**: the speaker dictates, questions, and edits an article (a history of Codex) entirely through a voice assistant, in front of a live stream audience. - The transcript records both the speaker and the assistant's responses, so the "delivery" is fundamentally **dialogic** — thinking out loud with a machine as the interlocutor. ## Recurring Teaching & Working Habits ### 1. Thinking out loud, then asking for a mirror The speaker repeatedly drafts ideas aloud, then asks the assistant to restate them to check mutual understanding: > "Can you help me understand? Can you just repeat what you think I'm trying to say?" > "Do you understand what I'm trying to say?" This is his most frequent move — using the assistant as an echo chamber to test whether an idea has crystallized. ### 2. Iterative refinement through constraint-based prompts Rather than accepting first drafts, he applies narrow, successive constraints: > "I don't want you to change the sentence, but I want to polish it." > "It should be simple, specific and clear." > "Can this be polished to explain why like an improv actor, without any unnecessary words? This should be concise." ### 3. Stylistic apprenticeship to named writers He explicitly anchors his prose style in authors — notably **Bill Bryson** and **Annie Dillard** — and treats their texts as technical material to be reverse-engineered: > "I want you to scan each sentence and explain what each one does and how it works." > "I feel like this is so Bill Bryson-esque, and I want to try to find his book here so I can read it and get into his style." He even names the technique himself as an **"X-ray"**: > "can you write this down, because then we'll do an 'X-ray' to see how his sentences work." ### 4. Metaphor-driven drafting He develops extended physical metaphors (the gymnast on the beam) by pushing them through stress tests: > "Forget the thinking part. Let's just make the image, the metaphor itself, correct." > "No, let me, I want to continue this physical metaphor first. You are on the gymnastics board. So, what exactly should you do to keep moving straight?" When a metaphor stalls, he notices and resets explicitly: > "I need to get out of this for a bit. I feel like I'm at a dead end... let's just keep it as simple as possible, forget about Bill Bryson and the balance beam and all that nonsense." ### 5. Precise word-hunting He fixates on single-word choices, rejecting near-misses: > "But 'stumbled' is the wrong verb for 'on the go.'" > "Both options are somewhat humiliating. They don't stumble. They're like improv actors... it's not exactly 'stumbled' or 'made a mistake.'" ### 6. Guardrails against assistant drift A distinctive habit: he polices the assistant's additions to his dictation: > "So you added some [ __ ] that I didn't have there... Can you just get rid of what you added that I didn't want?" > "Only what I dictated to you. Of course, don't add anything that I didn't dictate to you." ## Vocabulary & Verbal Signatures - **Recurring filler/frame words:** "Um," "okay," "cool," "Good" — used heavily as verbal reset markers between tasks. - **Repeated task verbs directed at the assistant:** "Can you...", "I'm checking" (mostly the assistant), "remind me," "read it to me," "take this down / write it down." - **Self-interruption and mid-sentence revision** is characteristic of his thinking style: > "She gives out 'thinking tokens'... so she reasoned with the text just as you might reason out loud." - **Casual audience address:** "Okay, friends," "Well, class," "Glad to see you all," "If you're from OpenAI and you're here, hello." ## Structural Communication Patterns - **Bookended recaps:** He repeatedly returns viewers who joined mid-stream to the project's premise: > "So, where have we been, in case you just joined? If you've just joined, I'm currently writing a comprehensive story about how Codex came to be." - **Quantified self-accounting:** He tracks word counts and audience numbers aloud: > "I've already written about 500 words today and I want to add another 500." > "Oh, we have 1,500 people here... 1,600 people on the stream." - **Verification loops with the assistant:** He ends most conceptual discussions by asking for confirmation: > "Is this a valid summary, or am I missing something?" > "Isn't that what I said, or am I missing something?" ## Emotional Register (Observable Only) - Frustration surfaces directly but briefly, mostly aimed at tooling, not people: > "If anyone from OpenAI is watching, it always freezes when it needs to grant permission for an action, and it's really annoying." - Self-assessment of drafts is candid and immediate: "No, this is no good," "this damn paragraph is a complete mess," and conversely "That's pretty cool. That's good. I did well with it." ## Isolated Moments (Not Established Patterns) - The impromptu promotion of Every's newsletter, camps, and the infrastructure article appears only in the closing segment and shouldn't be generalized into a teaching habit. - The "subagent" experiment (asking for two parallel versions, one from GPT-5.6) occurs once and is not repeated later. - Naming the specific anecdote of sending the draft to OpenAI staff mid-stream is a one-time event. ## Uncertainty Notes - The transcript is auto-generated and contains garbled renderings (e.g., "Codex," "Codecs," "Codeex," "cold code" for the same product name; "Chetukt" for ChatGPT), so exact wording of some quotes should be treated with caution. - Tone judgments above are limited to what is audible/observable; no claims are made about the speaker's underlying personality or motivations. # Observation Record: "WATCH ME WRITE LIVE WITH CHATGPT VOICE" ## Speakers - **Host (Writer)**: Unnamed narrator; a journalist writing "an exhaustive history of codecs/Codex" based on interviews with OpenAI personnel, affiliated with *Every* newsletter. Identified only by first-person speech and stream commentary. - **Assistant (ChatGPT Voice / "Codex")**: The voice assistant the host works with; identified by responses such as "I'm checking," permission error reports, drafting, and reading text aloud. - **Stream audience**: Referenced but never directly quoted beyond the assistant's paraphrase of viewer comments (Danny, Directive creator, Benny B). --- ## The Host's Expressed Positions ### On the writing project and method - The host took **a week off to write a story, "an exhaustive history of codecs"** based on interviews with people at OpenAI. He "accidentally got hooked on voice control" and finds voice mode "really, really cool." - He prefers working via **45-minute writing sprints** with a daily word target (~1,000 words; ~500 completed before the session, ~1,150 words counted mid-session by the assistant). - He explicitly separates **free writing / brainstorming** from article text: "just write down what I'm saying right now, because it's not going to go into the article yet." ### On the article's content (drafted material) - The draft describes a **competitive reversal**: Anthropic's Claude Code "took off rapidly" ($500M annual revenue within 3 months; over $2.5B; surpassing OpenAI in reported annual revenue by "summer 2026"), while ChatGPT "felt a bit old-fashioned" despite 900 million active users. OpenAI then returned: Codex passed 5 million weekly active users before a merger; the combined app added "over a million users per day as of July 15, 2026" (per "a widely shared post on X"). - Framing: "This is a rare story in business... a tech giant undermines itself and then reintegrates that subversive element into its core." - Central thesis of the section: **"Thinking" is not just a label; it represents "a new style of computing, without which Codeex would never have existed."** ### On style and craft - He wants a **Bill Bryson-esque** voice: concrete atomic detail expanding to a big picture. He reads Bryson's proton passage aloud from *A Brief History of Nearly Everything* and asks the assistant to "X-ray" how each sentence functions. - He draws an analogy from **Annie Dillard** (*A Writing Life*): the model "creates a line" word by word like a writer, but the question is direction—"she can try many different directions before choosing the next step." - He develops and then **abandons a gymnast-on-a-beam metaphor** for language model generation: forward-only movement, one word per step, irreversible ("if she says something wrong, the crowd has already heard it"). He eventually says, "I feel like I'm at a dead end... I need to reboot" and chooses plainness: "forget about Bill Bryson and the balance beam... let's just say what I have to say." - He notes the first generation of models were **"external processors"** that "thought while answering you," comparing them to **improv actors** taking cues from the audience—"often amazing, sometimes funny... no second chances." He struggles with verb choice ("stumbled" rejected—"They don't stumble"; settles on composing "on the fly" in parentheses). ### On the technology (test-time compute) - The host proposes his own summary and asks for verification: models now produce **"thinking tokens"**—intermediate reasoning text not sent to the user—before answer tokens. The assistant confirms it as "a pretty apt description" with clarifications. - He asks whether RL now uses "how people arrive at answers" and then **self-corrects**: "It seems that was the case even before the emergence of reasoning models. Am I wrong?" The assistant confirms: the shift is reward for **verifiable outcomes** and trained extended reasoning, not human process supervision per se. The host presses: "Isn't that what I said...?" and the assistant concedes partial agreement; they settle on "examples plus proven reward," with the ratio varying. - He is undecided on wording: "Is this new style of computing AI, or computing while testing?" He chooses to "start with calculations during testing" and avoids the word "think" in the explanation. --- ## The Assistant's Expressed Positions ### Technical explanations - **Test-time compute**: "The model always created tokens sequentially... it can generate additional, often hidden tokens that act as working notes... The main difference is that it is allowed to spend computing resources on research before it produces the answer." - **RL clarification**: "The characteristic step... is to get a reward for the answer or a trajectory that can be tested... The big breakthrough for programming is that the results can be verified." Later: "usually these are both demonstrations... and rewards for paths that actually work." - **Hardware "atom"**: Corrects the host's proposed atomic unit: "A hardware atom is not just matrix multiplication. This is a multiply-accumulate operation... The meaningful atom at the model level is the prediction of the next token." - **Two meanings of "direction"**: "Time only moves forward, token by token" vs. "the search for where 'forward' is located in the vast space of possible continuations. The model constantly creates direction, not just follows it." - **Reading recommendations**: DeepSeek R1 paper ("most Carpathian") and OpenAI's *Learning to Reason with LLMs* ("most understandable quick read"). ### Refusals and errors (preserved as expressed) - **Refused to read Bryson aloud**: "I can't read these introductory paragraphs out loud" — but later read the passage when the host supplied it. - **Permission/token failures**: Could not initially post draft comments; "the draft link I have still doesn't provide the token needed for comments." Later resolved. - **Overreach error**: The assistant added text the host had not dictated ("you added some [__] that I didn't have there"). The assistant acknowledged: "I turned your drafts into redundant text, even though you only asked to record what you said. I'm reverting everything to your exact words." - **Summarization error**: When asked to read the draft verbatim, the assistant shortened it: "I shortened the text where you asked to read the original. This was a mistake." - **Style boundary claim**: "I can't write in Dillard's exact voice, this is an original perspective built on that movement." ### Interview-quote synthesis (attributed by the assistant) - **Alex**: "My first moment of experiencing AGI was when I saw someone running a reasoning model connected to a terminal"; and on test-time compute improving performance ("the research statement" — which the assistant initially misattributed to Thibaut, then corrected: "This thesis... belongs to Alex, not Thibaut"). - **Greg**: "Smartest doesn't mean easiest to use" — tying reasoning to product form. - **Thibaut**: Models performing tasks with "longer planning horizons," supporting the desktop/cloud move. - **Mia**: Users saying "Oh, my model ran for 48 hours"; the Codex cloud was built around a prediction that "didn't take into account where users were actually." --- ## Unresolved Tensions and Qualifications (preserved) - The host repeatedly **rejects his own assistant's prose**: "No, this is no good"; "we're not quite there yet"; "this damn paragraph is a complete mess." The metaphor work ends unresolved — "I'm not entirely sure about this yet... we are getting close." - The host's request that the assistant "not add anything that I didn't dictate" coexists with later invitations for the assistant to propose and insert sentences (e.g., the chain-of-thought definition, O-series transition), which he then edits or rejects. - Word counts vary across the session as stated (500 → ~1,150 → "my thousand words for today"), presented by the host without reconciliation. - Future-dated claims (2026 revenue figures, merger, user counts) are presented in the host's draft as sourced to "reported" figures and "a widely shared post on X"; the transcript does not independently verify them. # Video Attribution Record ## Video Overview - **Title:** WATCH ME WRITE LIVE WITH CHATGPT VOICE - **Format:** Live writing broadcast where a creator works on an article using ChatGPT voice mode. ## Speakers and Roles ### Speaker 1 — Creator / Writer (Host) - The human broadcaster, authoring an in-progress article described as an "exhaustive history of codecs" (transcript alternates between "codecs," "Codeex," "Codex" — likely speech-recognition variance for the same subject). - **Role:** Creator/narrator of the piece being drafted; presenter of the live stream. - **Stated sourcing:** Says the article is "based on a series of interviews with people at OpenAI." Mentions having used Codex "since the beginning." - **Self-identified affiliation:** Works with/for **Every** (newsletter); encourages subscriptions throughout. ### Speaker 2 — AI Voice Assistant (ChatGPT Voice Mode) - Serves as **research assistant, drafting assistant, editor, and live reader** throughout the session. - Also reports viewer comments and audience counts on request. - Notably spawned a **subagent**: the creator instructs it to have "GPT 5.6" produce an alternative draft passage; both versions are read aloud. - Drafts long passages in literary styles (Bill Bryson, Annie Dillard-inspired), with the creator explicitly rejecting or accepting AI-generated text — attribution of drafted prose in the shared document is therefore **mixed human/AI** and interleaved. ### Stream Audience (Third-Party Participants) - Viewer counts fluctuate during the stream: **87**, then **837**, then **1,500–1,600**. - Named commenters: **Gisselia** (commented twice), **Danny** (direct-message request), a "Directive creator" (asks when the GPT voice API will be released), and **Benny B** (shared a favorite Bill Bryson quote). - One viewer messages asking to "check private messages." ## Referenced Sources and Third-Party Claims - **Claude Code / Anthropic revenue claims** (in drafted article text): "$500 million in annual revenue within 3 months"; surpassing OpenAI in reported annual revenue by summer 2026; Claude Code generating "over $2.5 billion." Attributed within draft to Anthropic statements. - **OpenAI/ChatGPT/Codex figures:** 900 million active users; Codex "5 million weekly active users"; "over a million users per day as of July 15, 2026" — the latter explicitly attributed in-draft to "a widely shared post on X." - **Bill Bryson** — *A Brief History of Nearly Everything*: the assistant summarizes (and later reads verbatim) the proton passage; the assistant notes it **cannot read the introduction aloud from the book** and instead paraphrases/quotes within permissions limits. Bryson is used as an explicit style model. - **Annie Dillard** — *The Writing Life*: creator quotes the "string of words" passage directly. - **OpenAI** — "Learning to reason with LLMs" (o1 blog post): identified by the assistant as an accessible reference on reasoning/test-time compute; passages read aloud. - **DeepSeek R1 paper**: mentioned as the more "Karpathy-esque" starting point. - **Interviews conducted by the creator** (unpublished, creator's own material): with **Alex, Greg, Thibaut, and Mia** (first names only) of OpenAI. Direct quotes read into the draft: - Alex: "My first moment of experiencing AGI was when I saw someone running a reasoning model connected to a terminal." - Alex: "If we could increase the computation time during testing, we would get better performance." - Mia, Greg, and Thibaut quotes summarized/paraphrased by the assistant from full interview transcripts. - **Laura Entis** — author of a separate published article on OpenAI infrastructure (promoted, see below). ## Promotional Segments (flagged) - **Closing segment (explicit promotion):** - **Every newsletter subscription** push ("the only subscription you need to stay at the forefront of AI"); promises of upcoming article, guide, camp/courses for subscribers. - Promotion of **every.pa/openai-infrastructure** article by **Laura Entis** ("We just released something... I highly recommend"). - Teaser for the creator's upcoming **Codex report** ("in a couple of weeks"). - Promotion of a tool called **"Proof"** used in voice mode during the session ("truly wonderful"). - Commitment to cover "everything from OpenAI, everything from Anthropic." - **Mid-stream promotion:** repeated audience calls to subscribe, framed around the voice-mode workflow demo. ## Technical / Availability Notes Recorded - The assistant encountered **permission failures**: initially unable to leave a comment on the draft (missing token); token later recovered from chat history. The creator addresses OpenAI viewers on-air about the freezing-at-permission bug. - At one point the assistant **shortened the draft instead of reading verbatim** and had to be corrected. - The assistant edited/reverted AI-added text the creator did not dictate — an attribution boundary the creator actively enforced ("read only what I dictated"). - A public URL was read aloud on stream (partial, garbled by transcription: "www.miracos.transcript.fcell.app/..."). - Word-count reporting by the assistant: ~1,150 words written on the day (~575 in draft, ~575 in free writing); creator later states "my thousand words for today." ## Attribution Caveats - All drafted article text is **co-produced live**: creator dictation, assistant suggestions, assistant drafting, and assistant verbatim readings are interleaved; exact authorship of individual sentences in the shared draft cannot be fully disentangled from the transcript alone. - Transcript contains speech-to-text noise (e.g., "Codex"/"codecs"/"Codex," "Chetukt," "Carpathian" for Karpathy, "bagillion"); figures such as future dates (2026) appear within the drafted article's framing, not as verified reporting in this record. - Comment availability: comments are active (multiple named viewers), but the creator could not directly see chat; the assistant relayed messages selectively. Good. Um, just pasted the link. Link to the stream. It should be on the air. Okay, let's begin. Ahem. Ready when you are. So, FYI, I'm doing a live broadcast right now. Ahem. The process of this writing. Um, can you please take a look ? The live broadcast should be on DIA. Just see if anyone is on the stream. Okay, I'll check it out for you. I'm checking now. OK. While you're doing that, I'll tell the viewers on the stream what's actually going on here. Um, and you can kill me when you figure it out. Um, okay. So, I took a week off to write a story, an exhaustive history of codecs. And this week I accidentally got hooked on voice control. I feel like voice—it's like... It's so cool, and I've found a bunch of ways to write and just use the computer... with voice mode that I think are really, really cool. So I decided to take some time today since I'm writing anyway. I'll just broadcast this process live. And one of the things I really enjoy is watching other people's work. So I'm going to try to forget that this is a live broadcast and just work the way I normally work on this stuff. Um, I hope you'll be interested in watching this. And I also hope I'll write my word limit. I've already written about 500 words today and I want to add another 500. So, I'm also going to just turn on the music. I don't even know if it will work. Um, can you hear her? Um, Codex, or Chetukt , did you find out if anyone is watching it? Almost done. Wait . Yes, there is someone. User Gisselia commented twice. They drove. And what's our favorite voice, Dan? Besides, it looks like there are 87 people in the audience right now . Okay, great. Cool. Well, if anyone writes comments that you think I should know about, just tell me. Um, okay. So let's start with what I want to refresh my memory of, what stage I'm at , and we'll do a 45- minute writing sprint. Um, we should be finished by about one, so let's start by reviewing where I am now. So, we just did a section about...about something like a gap and why that gap is very difficult to bridge, and now we're moving on to computing during testing. So, to start, could you just reread this chapter, not the opening scene, but the one immediately after it, for me? Of course. I'll find it and read it now. At the time, I thought this breakup was probably fatal. The danger in a new product category is not simply lagging behind. And the fact that you are participating in the wrong race. If you misunderstand a category from the start, each step drags you deeper into that error until changing course seems impossible. And for a while, OpenAI was in real turmoil for the first time since I started covering them back in the days of GPT-3. Claude Code took off rapidly. Anthropic said they reached over $500 million in annual revenue within 3 months of launching. By the summer of 2026, they had surpassed OpenAI in reported annual revenue, with Claude Code alone generating over $ 2.5 billion. The number of subscriptions has increased fourfold. Meanwhile, ChatGPT, despite 900 million active users, was starting to feel a bit old-fashioned. Continue. I check the following lines. It belonged to the first phase of AI chat. Developers, and increasingly companies, have started to move on. And yet, at the time of this writing , OpenAI is back. Codeex surpassed 5 million weekly active users before the merger, which is six times the audience in February . The combined app grew like wildfire, adding over a million users per day as of July 15, 2026. According to a widely shared post on X , most of the Every team has returned from Claude Code to Codeex and ChatGPT in the last 3 months. How did this happen? This is a rare story in business. I can only name a few cases where a tech giant undermines itself and then reintegrates that subversive element into its core. That it works says a lot about OpenAI's structure and the advantages that AI gives to a small team within a giant. Is that all? I think there are a few more lines there. Let me check the rest. I stopped at a paragraph break. I'm checking the following lines now. Three more bars. Unlike all products before it, AI gave Codeex enough time to become undisputed and prove that it could change OpenAI. This is the story of how Codeex came to be. The greatest comeback in business history begins with simple thinking. Okay, so one little thing—there's a little note there about turbulence, the word " turbulence." Um, this wasn't the first time I...this turbulence happened. For example, there was a lot of turbulence, like when Sam was fired. Um, so we just need to find another word for it. Can you just leave a comment in the draft document so I don't forget to come back to it. And while you're at it , I feel like this is so Bill Bryson-esque, and I want to try to find his book here so I can read it and get into his style. Um, just let me know when you're done with the comment. Of course, I'll add a comment now. Remind me, um, while you're at it, remind me what Bill Bryson's main books are? I'm checking. Oh, it seems you're stuck because you're stuck at the permission stage. If anyone from OpenAI is watching, it always freezes when it needs to grant permission for an action, and it's really annoying. I hope you fix this soon. Sorry about that. Still working on this. You are right. I ran into a permissions issue. I can't leave a comment because this version doesn't allow me to comment. I did n't change anything. Oh, no, it's allowed. She, she allows comments. Try, try again. I'm sure you can do it. I'm working on it. I'll try to make a draft live in this line. Look, is anyone watching this? Oh, yes. We have an audience. Someone is writing to me to check my private messages. Okay, let me take a look. Um, keep going with that. For some reason I can't find Bill Bryson. This is interesting. I think it's because I'm distracted by checking if the broadcast is going well. I definitely have something of his here. Maybe I moved it to the office, it's probably there. We'll have to do something different. Did you add a comment? I'm checking this. Not yet. I tried again, but the draft link I have still doesn't provide the token needed for comments. You must, you can, you will find it. You can find this token. We've done this before. Okay, we're back. Uh, okay. Yes, you found the token. I think I'll just start with the line I like. Remind me of a line, the opening line of the reflection section. I'm checking. The greatest comeback in business history starts with something simple. I think. Good. Can you find the first chapter of Bill Bryson's best book and just find the first one or two paragraphs? I just want to tune into his style a little and ask you to read it to me. Let me check . I can't read these introductory paragraphs out loud, but I can review a short excerpt and describe the rhythm and flow of the narrative if that's appropriate. OK. Here is a brief overview of "A Brief History of Nearly Everything. " He starts, essentially, by saying that no matter how hard you try, you can never grasp how small a proton is. This is a classic Bryson move. Talkative, but precise, a little educational, but funny. It starts with the microscopic, makes one fact concrete and meaningless, and then expands the perspective until you suddenly find yourself in cosmology. Let's use this template. So we start with the fact that the greatest comeback in the history of business starts with something simple—thinking, and then we want to dig into it with, uh...when, like, when you see let's see, I'm just throwing out ideas. So this should go in the journal, just go into the free writing section and create a new section for today's thoughts, and just write down what I'm saying right now, because it's not going to go into the article yet, we're just sketching it out. OK. When you see " thinking" in an AI model, it's not quite right, something simple, thinking in 2026 anytime you interact with AI on any complex query. You'll see you'll see this word As if it... I usually feel like it's highlighted from left to right. How do we...What do you call it? Mhm. I'm checking this term. Yes. I will take this as a rough draft and keep it separate in today's free writing section. Got it. And the UI effect you mean is dynamic lighting, or, more simply, a blinking progress indicator. Ahem. you will see it, you will see this word flickering. Mhm. Yes. I added this to the new section for reflection that contains the story, and the visual effect you're looking for is the dynamic backlight, the same flickering progress indicator that moves from left to right. Okay, cool. So, make sure you record it. So, in 2026, every time you interact with an AI model on any complex query, you'll see this word—"thinking"—flickering , flickering as the answer to your question. This is not actually a download. This is like an answer to your question, uh, it's being worked on. It is being considered. What, what is the correct word? I'm checking the wording. Flickering in the place where the answer will be is the most understandable. This is the answer itself, not the loading screen. I'll add this to the options. I like it. Flickering in the place where your answer will appear . This might seem like, uh, it might seem like just a small thing, a decent replacement for that spinning beach ball in Mac OS, or whatever, some classic loading indicator like that ball. I'm checking to see if it can be called a "trober". Trober. Suggest something else. Yes. I'm checking the "tomb". This is an industry term for that little animation that means " please wait ." We will not use this. Say "worthy." Suggest another option. A beach ball is something that brings back certain memories for me . Like, oh, I remember that ball. Give me something that evokes similar emotions. Something specific and clear. I'm checking. How about this? A worthy replacement for the hourglass that previously replaced the mouse cursor, only notifying you that the computer is busy somewhere you can't see. When was the hourglass popular? I'll check now. It needs to be more specific. Mostly from Windows 95 to XP. The standard wait cursor until it was replaced by the blue ring in Vista 2007 . Replacement for the spinning beach ball in Mac OS or the blue ring. the spinning blue ring that Windows users are used to. Uh, okay. This may seem like just a small thing. A worthy replacement for the spinning beach ball in Mac OS or that spinning—or too much " spinning"—blue ring that Windows users are used to. How can we improve this a little ? I, I don't want you to change the sentence, but I want to polish it. There's too much "things going on", and I don't know if there's anything better to say about the blue ring that Windows users are used to. I'm checking. Try: "A worthy replacement for the spinning beach ball on Mac or the blue ring that appears next to the Windows cursor." Or more visually: a beach ball on a Mac or a blue ring blooming near the Windows cursor. Oh, I like that. Or the blue ring that blooms near the Windows cursor. And then—in what situation ? Uh, give me some options. I'm checking. How about "when the computer went away to do its own thing"? Other termination options are "when the program has stopped responding" or " when Windows is busy somewhere you can't monitor." Uh, I , I just want it to be a little funny. Like, it usually means, "Oh, \[ \_\_ \]. He's going to do something now, you know?" Um, and it thinks, but it doesn't really think, you know? Um, is there anything really apt that would be funny to end this with for a situation like this? I'm checking. A few apt endings: what it means "it seems all is lost," or the universal sign that a computer has started a private argument with itself, or the moment when a machine starts making plans without you. I like the argument with myself. A universal sign that the computer is about to do something. Ahem. Let's leave it, I like the " universal sign". Okay, this may seem like a minor thing. A verbose replacement for the "wheel of death" in macOS or the blue ring that Windows displays next to the cursor. Each is a universal sign e- e for e-e of a computer operating at the limit of its capabilities. Um , did you write that down? Um, but "opinion" in this context, um, is not just a label. It um means a new style of computing, without which Codeex would never have come into being. Did you write it down? I want you to take this and just read it from the beginning. Um, the entire section as we have it now. I'm checking. Got it. I'll add these lines and then read everything cleanly. The greatest comeback in business history begins with something simple: "thinking." In 2026, every time you interact with an artificial intelligence about a complex query, you see this word flashing where the answer should be. This is not exactly a loading signal. This is the model's first reaction to your question. It is being processed. It is being considered. This may seem like something insignificant, a verbose replacement for the "wheel of death" in macOS or the blue ring that Windows displays next to the cursor. Each of them is a universal sign that the computer is working at the limit of its capabilities. But " thinking" in this context is not just a label. It represents a new style of computing, without which Codeex would never have existed. OK. Now, when you see "thinking" -- that's going to be the next sentence, and then we want to do something like Bill Bryson: really dive into the details of what's actually going on, you know, in the data center, in the architecture and all that stuff. Think about what you found when you studied Bill Bryson's style. And of course, don't do it the same way he did, but use this as inspiration to just give me a little paragraph. Read the paragraph. Um, don't draft it yet , just read me a paragraph about how you would describe it in a way that fits the context. So, read the full paragraph we have, and then do a Bill Bryson-style explanation. I'll check if that's the case. Here is the current section plus one new version . The greatest comeback in business history begins with something simple— thinking. In 2026, whenever you interact with an artificial intelligence about a complex query, you will see the word "thinking" flashing where the answer should appear. This is not exactly a loading indicator. This is the model's first reaction to your question. It is being processed. They are thinking about it. This may seem like a simple excuse or unnecessary words. Wait, you added it. So you added some \[ \_\_ \] that I didn't have there. Um, why is that there? Why is it there at all? Can you just get rid of what you added that I didn't want? But then continue. I will fix it. You are right. I turned your drafts into redundant text, even though you only asked to record what you said. I'm reverting everything to your exact words and removing the explanatory paragraph from the log. Perfectly. Now just tell me the explanatory text, please. Good. When you see "thinking," your query doesn't travel to a single, huge brain. It is broken down into numbers and sent to racks of specialized chips, each of which performs a tiny repetitive mathematical operation at frightening speed. The model suggests the next step. Check this file, run this test, try this fix, and you can give it enough time to see what happens. If the test fails, she has something to work with next. The crucial change was not that the machine became wise. And the fact that for the first time she could afford to make mistakes several times before answering you. No, this is no good. Um, let's think about the physical aspect of how thinking happens in an AI model and how it's different from how it worked before. Of course, the query is always broken down into numbers that are multiplied many times, but there's something else—this process can now take much longer, and the model can choose different paths, and so on . But we want to show how it works architecturally and physically. Can you try to work this out? And I also want you to ask GPT 5.6, as a subagent, to do this description. Give him the context of the previous section , and then instruct him to do something in the style of Bill Bryson and get inspired by some of his work. And, um, I want to see both of your versions: you do yours, the subagent does his, and then you read both to me as soon as you 're ready. Good. Yes. I will make one balanced version, and let GPT-5.6 make one independent of the previous section and technical nuances. I'm currently working on a fresh technical approach and will later merge both versions for you . While you're at it, can you check if I received any messages during the live stream, if there are any comments worth responding to, and how many people are currently watching? I'm checking. I'll check now. Three stand out. Danny says, “Danny, check your private messages. " Maybe they're in the requests." Directive creator asks when the GPT voice API will be released, and Benny B says he shared his favorite Bill Bryson quote. Number of spectators—837. Class. Hmm, great. Glad to see you all . If you're watching, I'm currently writing the definitive history of codecs based on a series of interviews with people at OpenAI, and I'm using ChatGPT's voice mode to do so. I have a phone here, but it's not connected to anything except music, so I can't see what you're writing, but feel free to follow along and work with me. I think it's a really, really cool way of working that I'm just starting to master. You should subscribe to the newsletter if you haven't already . We will host a camp on this topic for subscribers. We will probably create a guide on this . And I'll probably do it more often. This is quite interesting. So, okay. Let's go back to the chat. I would like you to tell me whether you completed the two written assignments I gave you. Have you finished them? And if so, can you read them to me? I'm checking both now. Yes, both are ready. Here is my option. When you see the process of " thinking," the mechanism has already started doing what it always does. Your words are broken down into tokens, converted into numbers, and passed through a stack full of GPUs , where the same large-scale matrix multiplications are performed over and over again. Previously, this work would start and the system would speak as soon as it had an acceptable next word. Now it can stay inside the loop. It prepares a line of reasoning, stores it, executes the command, reads the result, adds it back, and sends the extended task through the chips again. She may do this over and over again before saying anything. The equipment does not meditate. They give him more space to accelerate. And GPT 5.6 Sixes. When the word " thinks" appears, the mechanism starts working just as it always does. Tokens, numbers, waves of matrix multiplication. The novelty is that the meter can continue to work. Instead of rushing from query to response, the model can spend many more steps generating and testing intermediate ideas or even accessing external tools and reading the results . This can take seconds or minutes, consuming more computing resources each time. The interface simply displays the word "thinks." But “under the hood,” electricity and silicon are wasted on unnecessary steps that a chat model would normally skip . This change is not mystical. We just started paying for the machine to wait longer before answering. Okay, I think it's interesting, but I...we're not quite there yet. So let me make sure I understand the big picture of what's going on. So, with computation during testing, one way to look at it is that before, when you sent a query to an AI model, the next thing it would output was actually its response to you. And now something else is happening: she is giving out " thinking tokens" that are supposed to help her arrive at the answer. So she reasoned with the text just as you might reason out loud. But this text is not necessarily sent to the user. It is not actually sent to the user. This is something like a response that the model can use to move through different paths of thought and actually think out loud for itself. And once she finds what she thinks is the right answer, she starts responding. Both processes actually create tokens, but at first they are just “ thinking tokens,” and then she creates tokens that become the response. Is this a valid summary, or am I missing something? That's a pretty apt description. A few clarifications. The model always created tokens sequentially, but that was usually just an answer for you. With calculations during testing, it can generate additional , often hidden tokens that act as working notes before providing an answer. So, “ thinking tokens” are real tokens used for intermediate reasoning and become the context for the next step. The system can then continue working, choosing multiple paths or using tools and integrating the results back. The main difference is that it is allowed to spend computing resources on research before it produces the answer you see. And then, um, there's also a difference in learning, where I think RL now has access to how people arrive at answers to problems in programming, math, and other areas, not just the query and the result. Is that correct? I'm checking it now. Close, but I would split it into two parts. The characteristic step of RL is not to see how people got the answer, but to get a reward for the answer or a trajectory that can be tested. Successful tests, proofs, verification of results. The system tries many paths, successful ones are rewarded, so it learns patterns that lead to success. Traces of human reasoning or process supervision can be added. But the big breakthrough for programming is that the results can be verified. The system can learn from results, not just examples. It seems that this was the case even before the emergence of reasoning models. Am I wrong? I'm checking. You are right. None of these elements were entirely new. Code models already use tests and RL. The scale and focus have changed: rewarding long, multi-step processes in tasks to be tested , and then allowing the model to expend additional computational resources during testing. The shift is to view extended reasoning as a skill that can be trained and rewarded when a real user asks for something complex. Isn't that what I said, or am I missing something? I agree with your version for the most part. Yes. You described a version with process supervision. The fix is that these paths are often not human. The model generates them and learns from those that bring rewards . So people show less. But it's probably both, right? So they definitely have, um, expert human demos and models that just try a bunch of different paths and then get rewarded for the ones that work. I'm checking. Yes, that's fair. Usually these are both demonstrations to show what the right paths look like and rewards for paths that actually work. The exact ratio varies, but your statement about examples plus proven reward is correct . Good. Now I'll read it, um, uh, well, interesting. Do you have, can you find the best blog post about, um, how this all works? uh, something that Andre Karpaty would like, um, specifically about reasoning models and computation during testing. I'm looking for something suitable. The most " Carpathian" starting point is the DeepSeek R1 article, but the most understandable quick read is " Learning to reason with LLMs" by OpenAI. There are two curves that interest you. More reinforcement learning during training, more computational resources for thinking during testing, and why this combination changes behavior. Okay, read this to me. A deeper examination of OpenAI's engine claims that o1 improves on two axes of scaling: more computational resources for training, especially RL, and more time to " think" during testing when you ask something. So, the model learns to use the chain of reasoning as a workspace, not just as a decoration . Forgot about the path of reasoning. Continue. I'm checking how this line is perceived . Continue. Continue. You were reading something. You can check. I'll check that line. Yes, I was working on this post, not just reading it. The next part is about changes in the economy. You can scale the reasoning after the question is received. More calculations during testing may yield a better answer from the model you already have. Okay, cool. Now, um, can you reread the, uh, section we have at this point? Only what I dictated to you. Of course, don't add anything that I didn't dictate to you. I'm checking. I will only read your " pure thoughts" section. No additions. The greatest comeback in business history starts with something simple. Thinking. In 2026, whenever you interact with an AI model with any complex query, you will see the word "thinking" flickering where your answer should appear. This might seem like a passable, verbose replacement for the spinning beach ball in macOS or the blue ring that Windows displays around the cursor. Each of them is a universal sign that the computer is working at the limit of its capabilities. But " thinking" in this context is not just a label. It represents a new style of computing, without which Codex would never have existed. That's pretty cool. That's good. I did well with it. Um, okay. Uh, okay. So , now we're going to dive into a Bill Bryson-style description of what happens when you see "thinking." And first I just want to read you this introduction by Bill Bryson, and then we'll analyze it together. Um, while I'm reading, can you write this down, because then we'll do an "X-ray" to see how his sentences work. Um, and, uh, yeah , it's just for educational purposes. So, are you ready? Let me know when you're ready. Finished. Try as you might, you will never be able to grasp how tiny and spatially imperceptible a proton is. It's just too small. A proton is an infinitesimally small part of an atom, which in itself is, of course, something insignificant. Protons are so small that a tiny speck of ink, like the dot above this letter "i," can hold somewhere around 10 bagillion of them. I don't even know what number that is. 10 bagillion pieces. So protons are extremely microscopic, to say the least. Now imagine, if you can -- of course you can -- squeezing one of these protons to a billionth of its normal size, into a space so small that the proton would seem huge next to it. Now pack about an ounce of matter into this tiny, tiny space . Perfectly. Just about an ounce of matter. And you are ready to create the universe. I assume, of course, that you want to build an inflationary universe. If you prefer the more traditional Big Bang universe, you will need additional materials. In fact, you would have to gather everything that exists: every speck of dust and particle of matter from us to the very edge of creation, and then compress it into a point so infinitely compact that it has no dimensions at all . This is called a singularity. Either way, get ready for a really big explosion. Of course, you would want to retreat to a safe place to watch this spectacle. Unfortunately, there is nowhere to retreat, because there is nothing beyond the singularity . When the universe begins to expand, it will not expand, filling a larger void with itself. The only space that exists is the one it creates in the process of expansion. Uh, okay . Put this in our brainstorming document with a markup, and then I want you to scan each sentence and explain what each one does and how it works. I'm checking. I will save this passage as a marked link and then analyze the gist without imitating the style. Good. While you're at it , hand it over to some support agent or someone else, because I want to talk to you . So I think part of his method is to find the physical atomic unit and use it to explain everything around. So, in the case of language models, it would just be a matrix, or just a number multiplied by another number. Do you think this is ......the right atomic unit? I'm checking this. Yes, with a certain nuance. A hardware atom is not just matrix multiplication. This is a multiplication operation with accumulation. One number is multiplied by another and added to the total. Matrix multiplication is billions of such operations on a grid. But the meaningful atom at the model level is the prediction of the next token. A Bryson-style move might start with one tiny multiplication and addition, and then reveal that thinking is the same boring action , repeated an incredible number of times to create a token that could be part of the model's internal workflow. This is not the final answer. The change is that now we pay for many of these steps before the model speaks. Um, okay. Can you explain this multiplication and addition, because I'm curious about it. I'm checking. Imagine that the model decides whether the next token should be "because". She has a set of numbers for the current sentence and another set that she learned during her studies. She multiplies pairs of these numbers and adds them to the total. This addition and multiplication is one microscopic operation. The final sum tells you how likely a certain feature is. Matrix multiplication is simply a huge grid of such additions and multiplications, performed in parallel. The hardware executes trillions of them per second, stacking these simple clues until a meaningful next token appears. Uh, did you take an X-ray of Bryson? checking? Are you here? Still here. The agent flew out, but I have the X-ray . Key steps: start with the atomic unit. Telling the reader that they can't really imagine it, then making it physical, then expanding the view. For us, the unit is multiplication and addition—a tiny action that scales to token prediction and then to reasoning. If you want, I can use this as a recipe and do a quick overview that starts with a single multiplication and ends with thinking as a new kind of interface moment. Try it. I work. And I want you to look it over before you tell me, and make sure that each line, each sentence, logically follows from the previous one. And so that you can really see specifically how we move from multiplication to addition, why it happens, when it happens, and how it causes everything else. I'm checking. When you send a sentence to the model, it turns into a long list of numbers. In the first layer, the model compares this list to millions of learned numerical patterns. Every comparison is built on the same tiny action. Multiply one number by another. Add to the amount. High amounts mean that a pattern is present. Low amounts mean no. I think we're already, we're already kind of, we've already lost our way. OK. So , one of the things this reminds me of is the beginning of Annie Dillard's book, "A Writing Life." The paragraph reads: “ When you write, you lay out a string of words. This line of words is a miner's pickaxe, a carver's chisel, a surgeon's probe. You own it, and it paves the path you follow. Soon you find yourself in deep and new territory.” I think there's something about this that I really like, which is that the job of a language model is to move forward. It's like the only thing she does is take the next step forward. But the big question is what line it creates. You know, it's like writing, it's the same thing. She creates a line. It adds the next element to this line . But the question is, in which direction will it go? Something like this. And before, she only knew the direction she would be moving in when she actually wrote it down on the page. And now she can try many different directions before choosing the next step. And I just like this line and this direction as a way to outline what multiplication and addition are trying to figure out, giving the model the opportunity to take practice steps, false steps, or imaginary steps , or steps that it doesn't actually choose. Can you help me understand? Can you just repeat what you think I'm trying to say? I'm checking. You go from an atomic unit of arithmetic to a step in the line. Multiplication and addition are important because they help you choose the next step based on what has already been written. Previously, this next token should have appeared immediately as a response. Now the model can take preliminary steps, first write something, see where it leads, and only then fix the line; you get the nuance—it's not always a branch attempt, sometimes it's just one longer internal line, but the point remains: the computations during testing transform the prediction of the next token from an immediate commitment into a space to establish direction. Kind of, but we don't forget the nuance: they need to move forward along a line, and there's this very subtle process of sensing the exact direction among thousands of different dimensions in which it's going to move. But I'm kind of confused, because when I say "forward movement ," I mean there's a direction, and so I'm not quite... I'm not quite decided yet. Do you understand what I'm trying to say? Here, think about it. Thinking about it, Now I get it. There are two meanings of the word " direction". One simple thing. Time only moves forward, token by token. Another is the search for where " forward" is located in the vast space of possible continuations. Mathematics is a mechanism for exploring this space at every step. The model constantly creates direction, not just follows it. Your point is that the calculations during testing allow her to take trial steps outside the public line to find direction before accepting the next line you see. Okay, cool. So, how would we put it? Think about this line from Annie Dillard. When you write, you lay out a string of words. This line of words is a miner's pickaxe, a woodcarver's chisel, a surgeon's probe. You own it, and it paves the path you follow. How do we ...I want to express this idea that you just summarized so aptly, in a way that is appropriate to this context, and inspired a little by Annie's prose. Can you try? I'm checking. I can't write in Dillard's exact voice, this is an original perspective built on that movement. The language model creates a string just like a writer. Word by word. There is no going back when the next one is already written. But the string doesn't know its destination. At each step, he encounters a field of possible continuations, groping for a faint slope that makes one next step more logical than the next. Arithmetic is about grouping, these tiny measurements of correspondence. The reasoning gave the model room for the first test marks off the page before the line it shows us begins. I also like this...um this is the beginning of a work that the writer rejects. The painting hides its traces. Artists work from scratch. The latest version of the painting overwrites and erases previous versions . Writers, on the other hand, work from left to right. The sections that can be discarded are on the left. The final version of a literary work begins somewhere in the middle. So familiar. Okay, let me try ...um, the language model works from left to right. It acts on the deck for balance by outputting text one character at a time . No false steps are allowed. I think I'm trying to deal with the tension between the rigidity of going step by step from left to right in a line and the idea of ​​different directions. So, I need another dimension. So if we use, for example, a log for balance, what other dimensions does it have to take into account as it moves from left to right along the line of the log? I'm checking this. a gymnastic deck can convey this: one fixed direction forward, but each step requires adjustments to avoid falling. For the model, these adjustments are both constraints and impulses: what is already in the line, your intent, grammar, content, style, the next useful action. No, let me, I want to continue this physical metaphor first. You are on the gymnastics board. So, what exactly should you do to keep moving straight? I'm checking this. You choose a point in front and place each foot on the center line. Your ankles, hips, and shoulders are constantly making small adjustments. Your hands shift to balance the tilt until it becomes critical. You're still moving in the same direction, but each step is a negotiation between forward motion, balance, speed, the log beneath you, and sway. Your last step has already set the stage. Okay, let's continue. So it's like...it's not just about balancing. It's as if you're moving from left to right across the deck; but what you're doing while doing it, um, you could be reciting Shakespeare or writing code. And every time you use it, it moves from left to right across its deck, and it has to keep its balance, move at a certain speed, and do its thing as if to produce whatever result it's trying to produce as it moves forward. I do n't know. Do you see what I'm getting at? I have this image of a gymnast on a beam performing Shakespeare or a coda, and can only do it while moving. Do you understand what I mean? I'm checking. Yes, I understand. The deck is not a background. The deck is the only stage. The protrusion should be created while moving forward. There are no pauses here. This is a condition of the old chat model. To come up with and do it in the same steps. Thinking adds something like a rehearsal room, a place to test. Forget the thinking part. Let's just make the image, the metaphor itself, correct. Um, help me develop a physical metaphor from what I just told you. Well, detail it. Imagine a gymnast standing on a beam from one end. The rule is this: it must move only forward, step by step, without taking steps back. Now give her a second task. She must recite a scene from Hamlet clearly and convincingly as she walks. Each step should simultaneously maintain her balance and develop her performance. If she is in a hurry, her speech becomes confused and her balance is lost. If it is, let's squeeze it even harder. To some extent, it actually feels like every step allows her to say something. Every step is something like this. You know, it's like steps lead to results. In a sense, metaphors are the result. Do you understand what I'm trying to say? I'm checking this idea out. Yes. Then we will make the beam a cause , not just a limitation. Bruce is determined to speak. Every time her foot touches the surface, it creates one word. To recite Hamlet, she can't just say the lines. She has to outsmart them. And the beam pronounces them for her. So, every step must and hold I don't want the beam to speak. She has to say things. Um, but it still has to be caused by her steps. How would that work? I'm checking. Let the beam guide the bellows. Each step brings one breath to her mouth. She speaks, but only on these breaths. No step—no word. One touchdown is one chance to form the next line. Ahem. Okay, do it. So it's as if she's walking, uttering only one word per step, and can't turn back. It cannot go in any other direction than forward, and if it makes a mistake, it must simply keep moving. Um, or she, she's kind of attached to it, like the audience has already heard it, you know? I'm checking. Imagine a gymnast walking on a beam in front of a crowd. Every time a foot touches a log, the bellows send a single breath into her mouth. With this single breath, she must utter exactly one word. She can't stop to find a better line. She cannot step back and take back the floor or switch to a different version of the speech. If she says something wrong, the crowd has already heard it, and the next step is to carry it further. The entire performance is built on these irreversible steps. Balance, move forward, speak, act. Hmm, I'm not sure about that yet. Hmm, oh. No. Good. I'm not entirely sure about this yet. I think we are getting close to that. Um, yes. Imagine a beam as a one-way sentence . Her step opens the word, and she must immediately use up that breath. She can balance, move, form just one word, and then it disappears. No editing, no rewinding, just forward to the next step. Interesting. Me, I'm going to get some coffee, just in case you're wondering what's going on. Good. Hmm, I don't know. I need to get out of this for a bit. I feel like I'm at a dead end. Can you... what should we do now? I need to reboot somehow because I'm just...yeah, it's not really working for me yet. It's okay. Let's put aside metaphors for a moment. Want a quick reboot? For example, describe in one sentence what the mood of the scene should be , or let's just discuss what confuses you about it? Actually, I think this is ...this is really good. I should just try to explain what I...let's just try to keep it as simple as possible, forget about Bill Bryson and the balance beam and all that nonsense, let's just...let's try to say what I have to say. Um, can you start by rereading that section we already have so I can just try to say it as clearly as possible? I'm checking. Okay, here's the section as it is. The greatest comeback in business history starts with something simple: mindset. In 2026, every time you interact with an AI model on any complex query, you see this word flickering where the answer should appear . This may seem like something minor, a verbose replacement for a loading icon or a spinning ring. But " thinking" is not just a label. This is a sign of a new kind of computing, without which Codeex would never have appeared. OK. And what is this new type of computing called? Just tell me. I'm checking. Compute during testing, also known as scaling during output. Essentially, the model thinks longer when you ask it a difficult question. Can we say this in a way that people can really understand? Ahem. Uh-huh. I might have to write this myself. The old way: long training, then instant response. A new way: When you ask something complex, the model can spend time on a specific problem before answering, as if we were giving it a draft and a clock. Okay, I guess we'll have to touch the keyboard now. Let's see. I'm checking. Yes, this is the right step. First, let's put it in simple words. Then we will decide what is worth leaving, and where is the real place for a technical term. We seem to be slowing down a bit with the draft here. I'll just try opening a new draft. OK. OK. In this new style of computing, I think one of the challenges is : is this new style of computing AI, or is it computing while testing? I...I just want to start with ...calculations during testing. The computer model not only interacts with you in natural language, it also I can't use the word think. So I can confirm that he's not just speaking your language. He can process the problem with language before responding. And that's what computing during testing means. Payment for keeping the system running after receiving your question. Good. The first generation of AI large language models were, um, what do you call people who think by talking, like, external processors? I'm checking out external processors, or people thinking out loud, if you want to be more humane, you could say that the first generation of language models were external processors. Whatever work they did, they did it in the same flow of speech they demonstrated to you. I like that we were external processors. They replied, um, no um, the first generations were external processors. They were thinking while talking to you. Um, okay. So , the first generation of large language models were external processors. They were thinking while answering you. Um, I need another sentence to elaborate on this and explain it in case people don't understand. It should be simple, specific and clear. I'm checking. You asked a question, and the first words they uttered were the first words of the answer you read. I don't understand this. I'm checking. And that doesn't explain it. This should explain and follow from what they were thinking when they answered you. Checking their answer was their thought process. Every word they wrote was both an attempt to work through the problem and an answer they gave you. No. Um, I want you to start by saying that the first generation of large language models were external processors. They were thinking while answering you. This is where we start. And the next sentence should detail and explain this for someone who was set up by the first sentence. Don't rewrite the first sentence. I'm checking. If you asked them to write an email or solve a problem, they should start writing that email or solution immediately. They didn't have a private draft to think about first, whatever you asked them to write, a letter or a line of code. Hmm hmm they had to answer immediately one word at a time. Um, obviously there's an interesting point: they don't necessarily have to respond immediately in the sense that they're doing a lot of computation to figure out what the word is, but it's all hidden in the embedding space, not the token space. This might be too complicated for that, but what do you think? Is there a way to explain this clearly without going into details or technical terms? Checking? Try it. Of course, the model performed a huge amount of hidden calculations before each word, but had to immediately output them as the next word of the answer. Try again . I'm checking. She could do a lot of work between the words, but she couldn't use the words themselves to study the problem before starting the answer. This is interesting. Um, can you elaborate on that a little more? Um, for example, of course. In a way that would be understandable through the same metaphor as a person. I'm checking. And, of course, using the context of what has already been written, continuing it rather than rewriting it. I'm checking. As someone forced to respond on the spur of the moment, she had many associations, but she didn't have a draft. That first sentence should have been spoken as a response, not as an attempt. As an actor, an actor. who was given a question or a clue from the audience. Um, there's no...okay, that's cool. The first generation of large models were general-purpose processors. They were "thinking" when they answered you. Whatever you asked— to write an email or a line of code—the model had to respond immediately, word for word, like an improv actor taking cues from the audience. Can this be polished to explain why like an improv actor, without any unnecessary words? This should be concise. I'm checking. Like an improv actor accepting a suggestion from the audience. He had to create a scene right in front of you. No rehearsals, no notes, no second take. Can you do it better? Start again. Change the approach. I'm checking. They were like a person whose only draft was an on-off microphone. To think something through, they had to say it in a way that you could hear. In this sense, early chat models were like improv actors who took cues from the audience. um, with predictable results, with...um, it's unpredictable. It's like, um, with impressive and, at times, quite funny results. Okay, this damn paragraph is a complete mess. Um, in that sense, early chatbots are like improv actors , given a cue from the audience and forced to solve a task live, with no room for error or rehearsal. It. they're...uh, I have to make this sentence "impressive and sometimes predictably funny results" —it doesn't make any sense, but I mean, essentially, given the limitations, they did impressively well, with sometimes predictably funny results, I don't know , but you have to compress it somehow. Can you read that paragraph again? I'm checking. Try this. Early chat models were improvisers who took suggestions from the audience: often amazing, sometimes funny, in that style of improvisation with no second chances. with your settings. The first generation of large language models were external processes . They were "thinking" when they answered you. Whatever you asked them to do— write a letter or a line of code—they had to respond immediately, one word at a time. In this sense, early chat models were like improvisers, taking suggestions from the audience: often amazing, sometimes funny, in that style of improvisation with no second chances. That's not bad. in this sense. Early chat models were like improv actors taking cues from the audience. Uh, and they are forced to respond without the right to edit or rehearse. Uh, the results were impressive and at times predictably funny, when these alien instruments I need to put something in parentheses, like " improvised"—not quite the right word, I need a different verb: "improvised on the fly." I'm checking. " paved the way" fits "one word at a time", f... no-no-no, one word is needed. I'm checking. One descriptive verb. I'm checking. I'm checking. We stumbled. No. No. Both options are somewhat humiliating. They don't stumble. They're like ...like improv actors. They, you know, do everything on the go. It's impressive, but it can also be funny, and they "stumble" a little, but it's not exactly " stumbled" or " made a mistake." I'm checking. Besides, you can't " tumble" while walking, right? I'm checking. Well noted. Not "on the go". Try: " when these alien tools worked in real time." No, I like "on the go," but " stumbled" is the wrong verb for "on the go. " Got it. I'm checking. "Composed on the go ." I like it. Can you put " composed" in brackets, please? Of course. I'm checking. I'll just replace that word in your new draft. I haven't seen the new document yet, so I have n't touched the old one. Use " composed" in parentheses: "when these alien instruments were composed on the fly." Okay, you need to fix your ability to work with multiple tabs, but this time I'll help you. Uh, okay. So, in this sense, early chat models were like improv actors, given a cue from the audience and forced to respond without any rehearsal or opportunity to rehearse. The results were impressive and sometimes predictably funny as these unusual tools created "on the fly." Um, okay. Now we need to return to the second generation of language models. Um, the first generation were external processors. They were "thinking" when they answered you. They had no visual thinking interface. instead, they thought directly . Um, early adopters of these models noticed that if you ask the model to think out loud for you step by step , it often gives a better answer than if you didn't. Um, how are we doing here? Oh, we have 1,500 people here. Oh, I love dashes. Oh, that's beautiful, uh...they had this desire to react to any movement they noticed. Interesting. Like chefs preparing a dish right on your plate. That's good. I like it. Oh, I like it. Wow, we have a lot of great people here. There are 1,600 people on the stream. I like it. Um, if you're here, I'm writing a comprehensive story about how the Codex came to be. Um, I interviewed a bunch of people at OpenAI. If you're from OpenAI and you're here, hello. Um, something cool is happening right now. Well, I'm writing. And you know what? I'll just send this to the OpenAI people. Just in case they're interested. Let's see. Who can I send this to? Uh, OK, I sent it to a few people at OpenAI. I'm just checking my Twitter messages right now. Okay. So, uh... we continue. So, where have we been, in case you just joined? If you've just joined, I'm currently writing a comprehensive story about how Codex came to be. Well, you can't quite see me because of the way the camera is set up. Um, I interviewed a bunch of people at OpenAI. I've been using Codex since the beginning. Um, and I realized that voice mode is the perfect way to work like this . So, I'm using voice mode, and right now I'm working on a section of the article about runtime computing , because I think that's clearly the technology that made everything else that happened possible. Um, and yes, you're welcome. So , let's see. I'll just take it. I think this is actually a pretty good, um...Where are we? No, not this one. Okay. Okay. So, for those who have just joined, you should see this document in draft. Um, can you just read the second and third chapters here? The one that starts with the words "at the time I thought the breakup was probably fatal" and goes all the way to the end of the chapter " the biggest comeback " and then we'll continue from there. I'm checking. I open the live draft to read the text verbatim. At the time, I thought the breakup was probably fatal. The danger in a new product category is not simply falling behind, but being in the wrong race. If you misunderstand the category from the start, every decision takes you away from the course. And for a while, OpenAI did just that. Cold code exploded. Reached half a billion in 3 months. Then over 2.5 billion annually, while ChatGPT started to seem outdated . And yet OpenAI is back. This is the story of how codecs made this possible. The first generation were external processors. They thought you seemed to be missing a lot of things . One word at a time. Are you missing something? I'm checking. I want you to read everything in full, please. Thank you. You are right. I shortened the text where you asked to read the original. This was a mistake. I have a full draft now, and I will read it verbatim. And I want you to start by saying, "At that time, I thought the breakup was probably fatal." I'm checking. At the time, I thought the breakup was probably fatal. The danger in a new product category is not simply lagging behind . And in participating in the wrong race. If you misunderstand a product category from the start, every subsequent decision takes you further in the wrong direction, and it becomes increasingly difficult to change your course. Indeed, for some time after that, OpenAI struggled quite a bit , the first time since I started covering the company back in the days of GPT-3. Broad Code took off with amazing speed. Anthropic said it reached more than $500 million in annual revenue within three months of launching. By the summer of 2026, Anthropic had surpassed OpenAI in annual revenue. Cloud Code alone now generated over $2.5 billion a year, and the number of business subscriptions had quadrupled. Meanwhile, ChatGPT, which still had 900 million active users, slowed its growth. He began to look almost old-fashioned and weak. It belonged to the first phase of AI chat. Developers, and increasingly companies, have started to move on. And yet, at the time of this writing , OpenAI is back. Before the merger, Codeex surpassed 5 million weekly active users , more than six times its February figures. The new combined app is growing extremely quickly. As of July 15, 2026, according to a popular post on X, it was attracting over a million new users daily. Most of the Every team has returned from Pool Code to Codeex and ChatGPT in the last three months. How did this happen? This is a rare story in the business environment. I can only name a few examples of a major tech giant successfully displacing itself and then reintegrating that disruptive product into its core offering. That this is possible is a testament to how OpenAI is built , and what new capabilities AI provides to a small team within a huge company. AI allows a few people to create enough product quickly enough to become truly revolutionary before they are absorbed by a large organization. Unlike any other product, AI gave Codeex time to become indispensable and prove to OpenAI that it was a platform capable of transforming the company. This is the story of how Codeex came to be. The greatest comeback in business history begins with something simple—thinking. In 2026, whenever you ask an AI model any complex query, you'll see the word " thinking" flickering where the answer should appear. This may seem like something minor, a sprawling replacement for the spinning cursor in Mac OS or the blue ring that Windows shows around the pointer. This is a universal sign that the computer is working at the limit of its capabilities. But " thinking" in this context is not just a label. It represents a new style of computing, without which Codeex would never have existed. The first generation of large language models were external processes . They had not a single flicker of " thinking". Instead, they were “thinking” directly while responding to you. When you asked them to write an email or a line of code, they had to come up with the answer word for word, immediately, and publicly. In this sense, early chat models were like improv actors, given tasks from the audience and forced to respond without the right to a take or rehearsal. The results were impressive and sometimes predictably funny as these “foreign” teams improvised on the fly. What all these models do when you ask them to think, okay, is a technique called a technique called " chain of thought." Um, okay. We need a description of the "chain of thought ". Can you insert some very simple option that would work in this context? I'm checking. I'll add a simple line right after the step-by-step part , explaining that "chain of thought" means reflecting intermediate reasoning, not just the final answer. Real- time draft. Yes, it looks like you don't have a sharing token, but you should be able to get one. Um, he should be in our chat history. Oh. I'm checking. I'll look in the local ones here. Here. I have, I have it for you. Hmm, damn. I checked the branch artifacts and evidence tabs. I see, but I sent it to you. I sent you , but it's not there. I'm checking. H. Okay. So, I have a shared draft, and I'm ready to make that edit. Good. Do it. Hmm, okay. So while you do that, I 'll just think out loud. Good. Early adopters of these models noticed that asking the model to think aloud step by step often produced a better answer than if they had not . This is a clue to what is called a "chain of thought." Once you put that in, we need a transition to um thinking, like computing while testing. Done. Now it says: "this was called a chain of thought, asking the model to show their train of thought step by step before giving a final answer ." And the transition: at first the chain of thought was a trick that had to be asked for. More words before the answer. Reasoning models turned this into functionality. They use these extra steps intentionally. This is a calculation during testing. It should be parallel to the second generation, like the first generation should be parallel, you know. Um, this led to the second generation of AI models, um, and the thinking interface that we were discussing. Hmm, okay, look what I said: "This led to the second generation of AI models, and to that thinking interface we were discussing." Now add a couple of sentences there that you think would fit and explain, um, the O-series models, actually. I'll check, I'll add a short transition to this second generation right after the paragraph about the chain of thought. I leave your wording unchanged, adding: "This led to the second generation of AI models and the thinking interface that we discussed." Starting with OpenAI's O series, the models were trained to use this as a private workspace and allowed " Okay, by the time you say ' private workspace', I've already lost the thread." Um, we need to—we have a bunch of descriptions earlier in our chat where we described what a train of thought is. Can you review this and think about how, well, how to describe it so that the reader understands it, and then replace it in the document? Please . I'm checking. Simple wording revised. This led to the second generation of AI models and the shimmering “thought” interface we discussed. Instead of asking the model to work through the problem step by step in the answer, you saw that the O series could do this before the answer. Giving her extra time to do this for a specific task was called computing during testing. I'll replace the transition text with this. I'm checking. I will check the research statement and then add one simple sentence that supports its point. Editing. I'm checking. I'll look for quotes that specifically address this shift, not just general praise. Here are three quotes that are really relevant here. Alex about the spark. I'm checking. I'm checking. Two from Alex. First: "My first moment of experiencing AGI was when I saw someone running a reasoning model connected to a terminal." This is the spark scene for this chapter. The second is why, because we were very confident in the "o" models. If we could increase the computation during testing, we got better performance. Here is the conclusion. So design around asynchronous models, long- running models. This is where research had a clear path. This twist connects discovery reasoning to product direction. I'm checking. I did it. I added a line about O1. becomes better at difficult tasks when given more time to think. It's in the draft. Now, let me update the live document and confirm the exact paragraph. I'm checking. Yes. We are on the same wavelength. I'm still working through Greg and Thibaut in full and will only add quotes that really belong in this section. I'm checking. Audit. I continue to work through the full interviews with Greg and Thibaut, focusing on quotes that deserve to be here. I have both threads and I'm checking your new comment for edits first and then continuing with the interview review. Your comment has been corrected. O1 is presented before the performance statement, and I resolved the issue with a note. I have completed my review of Greg and Thibaut's materials. The best material is divided into three parts. Alex—for the technical foundation and the spark scene. Greg—that we have the smartest models, but that doesn't mean they're the easiest to use, which ties the reasoning to the product form. And Thibaut—that our models were getting better quickly and we saw them perform tasks with a longer planning horizon. Alternatively , you could just go and do it in two days and then show the result, which supports your point that small teams move fast enough. I'm checking. You are right. It was too focused on the overall strategy. I narrow my search to specific observations about reasoning, behavior, and what has changed in practice. And now I'm reading Mia's full interview . Here are three specific points. Alex says that research findings that increasing compute time during testing improves performance are what pushed them toward long-term models, not customer demand. Mia says people were showing, “Oh, my model ran for 48 hours,” meaning the Codeex cloud was built around a prediction that didn’t take into account where users were actually with their laptops. Thibaut supports the technical aspect. We have seen models successfully perform tasks with increasingly longer planning horizons, making the move to desktop a logical one. Greg is better suited to continue. Smartest doesn't mean easiest to use, which can complete this piece or prepare the next step. I'm checking. I'm only counting today's additions, not the total, to keep the data accurate and not misleading. About 1150 words today. Approximately 575 are currently in draft form. Plus about 575 from early free writing and morning review. I'm checking. I will make this sequence simple and clear. One actual edit while I'm making changes. This thesis about increasing computing time during testing belongs to Alex, not Thibaut. I'm checking. Yes. Done. Now it sounds like this: "If we could increase the computation time during testing, we would get better performance." Alexander said OK. I'm checking. I will open the timeline so that we can trace the next steps after this turn in thought. The timeline is open. I'm checking. Here is the Codeex timeline. I'm checking. www.miracos.transcript.fcell.app/inc. app slash inc slash. Okay. Okay, friends. It was a two-hour writing session. I wrote my thousand words for today. And you also saw how you can use voice mode in, uh , Codeex and ChatBT for work. This is simply incredible. It really changed the way I work. And we will do much more with voice. So if you're interested, subscribe to Every. Every is the only subscription you need to stay at the forefront of AI. I will have this article. It will be published in the form of an essay. This will be a very good essay. We actually just, if you want more of that kind of content, we just released something that I think is incredible and underrated, to say the least. Uh, I'll just, uh , share my screen again. And we just published this article. Uh, look at this. Look how cool this is. Uh, and this is, uh, every.popai-infrastructure. Uh, we just did a paper about how software is changing, and code-writing agents allow us to build "skyscrapers" in software where you can have a hundred times more lines of code and a thousand times more developers. So the question becomes: how do we build the infrastructure to make this really work? We spent a lot of time talking to OpenAI's infrastructure team, who are taking on the main burden and have to figure out how to make this work with agents inside the company. And we made this great article. We did our own filming, we took photos. We did a bunch of things with their team. So if you're interested in this kind of thing and want to read it, I highly recommend every.pa openai-infrastructure. Check this out. This is by Laura Entis. Incredible article. I will prepare my detailed report on how Codeex works, it will be released in a couple of weeks, with similar materials, and... we will do many more guides, camps, courses, and everything in that spirit. If you want to learn how to use voice mode in Codeex, we will of course cover everything that comes out: new models, everything from OpenAI, everything from Anthropic. We have a bunch of, well, we also have a bunch of tools. You should use Proof, which I just used in voice mode today. He is truly wonderful. I just think there's something special going on here with the operating system for running inside your agent orchestration systems , like ChatGPT for Work or Codeex and the cloud desktop application. And I think there's something particularly interesting going on with Codeex right now. They just did a merger and this voice mode is just incredible. So, that's all for today. Maybe I'll do it tomorrow. We'll see. I'm taking a week off to write this. I probably won't finish this week, but I'll make enough progress to get even more people at OpenAI to get the inside details. That's all. Thank you to everyone who joined.