Greg Brockman Finally Tells the Full Story: The Board, the Petition, Musk, and the Fight for ControlGreg Brockman（OpenAI 联合创始人）把"那场宫斗"讲完整了：董事会、请愿书、马斯克与控制权

Greg Brockman appeared on Rick Rubin's Tetragrammaton podcast for a conversation lasting over three hours. Brockman is OpenAI's co-founder and president, but he isn't the typical tech company executive. He still writes code today and personally designs data center architecture — Fortune magazine called him OpenAI's "builder-in-chief."

What makes this episode special is that Rick Rubin isn't a tech journalist. He's a legendary music producer, and his questions come from a completely different angle. He doesn't chase technical details — he chases motivation, emotion, meaning. The result is that Brockman revealed a side rarely shown by tech executives: the decision logic behind facing betrayal, the full inside story of negotiating with Musk, and how a teenager from a small town in North Dakota ended up at the frontier of AI.

Original video link: https://www.youtube.com/watch?v=mI30_ueZ7CU

Here are the key takeaways:

Brockman resigned the day Sam Altman was fired, telling his wife to assume all their equity would go to zero. 72 hours later, 95% of employees had signed a petition — crashing Google Docs — and they took the company back.

2. Musk once demanded majority equity, absolute control, and the CEO position at OpenAI. After Brockman and Ilya refused, Musk also proposed folding OpenAI into Tesla to "secretly develop AGI," and when that fell through he left. Before leaving, he sent an email saying OpenAI's chance of success was "0%, not 1%."

Scaling laws still hold — "people on public forums saying they're dead are completely wrong." A qualitative leap happened between GPT 5.1 and 5.2, making many expert-level engineers feel for the first time that the model could "really handle extremely hard work." Brockman predicts compute will become a basic human right.

A physics professor used OpenAI's latest unreleased system to prove the opposite of a quantum physics hypothesis. The paper is being submitted, and the professor said "this is the first time it felt like the system was actually thinking."

OpenAI is internally pushing "Agent-first," aiming for everyone to default to using AI agents rather than text editors by March 31, 2026. The slogan is "Say no to slop," and the review bar for AI-generated code is stricter than for human-written code.

ChatGPT is managing his wife's health, while he himself only started using Codex in December.

Rubin asked what changes with each new generation of models.

Brockman said a clear trend in 2025 is that people have started using ChatGPT in very personal contexts. His wife has hypermobility syndrome, and it took years to get diagnosed. Once she entered her symptoms into ChatGPT, it was able to figure it out quickly. Doctors each have their own specialty and no one sees the full picture, but ChatGPT can. She now uses it to manage her daily health.

He himself, by contrast, was "a late adopter of his own product." He'd always used the Emacs editor and the terminal — tools he'd used since he was young, very fixed habits. It wasn't until he started using Codex in December 2025 that he completely abandoned his old way of working.

Note: Codex is a coding agent product OpenAI launched in 2025 that can autonomously carry out coding tasks, a workflow completely different from the traditional IDE editor + terminal setup.

On the nature of model upgrades, he said the outside world thinks "you're just making the models bigger," but in reality every single piece is being continuously optimized. For example, in a training cluster of 100,000 GPUs, how do you detect which GPU is broken? "You can't just point at it and say 'that one.'" Data quality, tokenization, physical processes, software processes — every part is being refined over and over.

He summed up a pattern: if an application "barely works" right now, you should expect it to become excellent a year later.

At OpenAI, you live in the future. ("That's one of the amazing things about working at OpenAI—is you live in the future.")

"Sam has been badly misunderstood by the world"

Rubin said Sam Altman seems to provoke very strong reactions from the outside world, and asked Brockman how he sees his co-founder.

Brockman's answer was very clear: Sam is a good person, and that very goodness has become a handle for attacks against him.

Every accusation is a confession.

Meaning that critics are projecting their own insecurities onto Sam.

He said Sam is extremely resilient under pressure, has been through a great deal, and has been indispensable to getting OpenAI where it is today. "No one else could fill this role the way he does."

Rubin said, implicitly, "he draws all the attention, so you get to keep your head down and do the work." Brockman said he's "deeply grateful" for that.

"That's when I knew something was wrong": the day Sam Altman was fired

Rubin asked directly: "Tell me about the firing."

Brockman first gave the background: for a year to a year and a half before that, tensions had been building inside OpenAI. The biggest lesson he learned from it is that when conflict exists, it must be addressed directly — "if you let it fester, it's always more painful later."

Then he recounted what happened that day.

He was writing code, about to submit a change he'd been excited about for a while. He got a message inviting him to join a video call. He opened the Google Meet preview screen and saw the board members online, but Sam wasn't there. He immediately sensed something was wrong and clicked join.

The board told him: Sam had been fired, Mira Murati was interim CEO, and he was being removed from the board. They said he was important to the company, someone who "gets things done," and they wanted him to stay. And this wasn't even a question of whether he wanted to stay — they were simply informing him, "this is your new role."

He asked for more information, and the board wasn't willing to share it.

That's when I knew this was wrong.

Note: On November 17, 2023, OpenAI's board abruptly announced it had fired CEO Sam Altman and removed Brockman from the board, triggering one of the most dramatic corporate crises in Silicon Valley history. At the time the board had 6 members (including Altman and Brockman); several directors had already left that year for various reasons. Mira Murati was then CTO and later left in September 2024 to start her own company.

Rubin asked if there had been no warning at all. Brockman said yes, it was completely shocking. But because he knew everyone involved and all the tensions that had been building for a long time, his in-the-moment reaction was, "I understand what happened, and I'm sorry it came to this."

After hanging up the call, he told his wife: "We have to leave."

She said, "Yes."

Then he did something: he told his wife to assume that all the equity they held — none of which they'd ever sold — would go to zero, because the board might turn "adversarial." He named the dollar value of that equity and asked if she still wanted to go.

She said, "Go."

They had kept putting off having children so he could commit himself fully to OpenAI. They'd put eight years into it. But they both believed in the mission of AI benefiting humanity.

He called Sam. Sam said, "I guess I'll go start a new company."

Brockman said, "No, Sam, we're starting it together."

He resigned that same day.

The mission is bigger than the corporate entity.

72-hour "revolution": a 10% chance of taking OpenAI back

Brockman said that in people's memories, the whole episode blurs together, but there's a key turning point that gets overlooked: his resignation.

Before he resigned, the outside narrative was "someone got fired, so something serious must have happened." After he resigned and put out a brief statement, the narrative shifted. He believes the board hadn't anticipated that he would resign.

That same day, his core collaborators Jakub Pachocki, Szymon Sidor, and Alexander Wei also resigned.

Note: Jakub Pachocki later became OpenAI's chief scientist, succeeding Ilya Sutskever in that role.

A huge number of OpenAI employees reached out to him on their own, saying "whatever you do next, I want to be part of it." He said this "truly moved me deeply — it's something I would never have asked for and never expected."

At the time, he judged the odds of getting the company back at 10%. Not zero, but no more than 10%, because the board held all the cards.

The next day they met at Sam's house to plan the new company. No one slept that weekend — it ran entirely on caffeine. The whiteboard was covered with the vision for the new company.

Rubin said it sounded exactly like a revolution. Brockman said it felt exactly like that, one hundred percent.

Then things escalated. On Sunday night, the board decided to replace Mira with a new interim CEO. The employees collectively refused. People started pouring out of the office. Hundreds crowded into various private backyards, some giving speeches, while competitors swarmed in to make offers to OpenAI employees.

Within a single day, Brockman and his team expanded what had started as a "lifeboat" plan for a small group into one that could take in the entire staff of 770 people — accomplished in partnership with Microsoft. Microsoft said, "we'll take everyone." They discussed multiple options: an independent company? Microsoft investment? Or folding into Microsoft entirely?

That night, employees spontaneously organized a petition to the board demanding it reverse the decision. So many people were editing it that Google Docs crashed — they had to designate someone specifically to handle adding signatures to the document. In the end, 95% of employees signed.

Brockman went to bed very late, having spent the time comforting people who didn't know what to do. He woke up at five in the morning and started scrolling Twitter.

He saw Ilya Sutskever's post.

Note: Ilya Sutskever is OpenAI's co-founder and chief scientist, widely seen as one of the key figures behind Sam Altman's firing. Shortly after the firing, he tweeted expressing regret: "I deeply regret my participation in the board's actions. I never intended to harm OpenAI." That tweet was viewed nearly 29 million times. Ilya left OpenAI in May 2024 to found SSI (Safe Superintelligence Inc.).

At this point in the conversation, Brockman was visibly emotional.

"When the firing happened, all I felt was distance. Just that feeling of — how could you do this to me?"

"But that morning, seeing Ilya's post, I felt forgiveness."

It then took about two more days to negotiate the return. The core of it was forming a new board — a slate of people that both the old board and they themselves could accept.

After the crisis: "the company can die too"

Rubin asked what was different about the company after the event. Brockman said, "Extremely different."

"Have the hard conversation" became a core value. After coming back, they spent a lot of time trying to improve how they handle conflict — it's still not perfect, and there are still times when they see conflict being avoided.

More importantly, there was a kind of awareness that came from a near-death experience.

This company was never destined to succeed. That it survives and thrives is because people work hard, because people make it so.

Rubin asked what the actual trigger was — was it a technical issue or a business issue?

It's always a people problem. Just like problems in a marriage.

After the return, an independent law firm conducted an independent review of the whole episode. The conclusion was: it was understandable why the board made the decision it did, and it was within their power to do so, but they didn't have to.

Brockman said one of his core strategies at OpenAI has always been "keep the band together." Whether it was the disagreement with Musk, the departure of Anthropic's founders, or this firing, he always tried to get everyone to find a good solution together. But the outcome was always that someone left.

Their habit is not to fight over the narrative — "we keep moving forward and let other people tell their story. But sometimes people use that against us."

David vs. Goliath: where the story with Musk begins

Rubin referenced an earlier private conversation between them about "David and Goliath."

Brockman said that because ChatGPT has been so successful, the outside world tends to see OpenAI as the industry leader, as "Goliath." But sitting inside the company, the picture looks completely different — Google and Meta are the real giants, with enormous market caps, massive user bases, and huge amounts of compute. OpenAI is the challenger.

And the real prize isn't current market position — it's AGI. Whoever gets to AGI first and turns it into a force that benefits everyone is the winner. On that dimension, OpenAI is still David.

The same goes for Musk. Brockman said Musk is one of the richest and most powerful people in the world — "and he's suing us, saying we deceived him."

That leads into the next big topic.

Negotiating with Musk: the fight over control

Rubin asked where the core disagreement lay. Brockman answered with one word: "Control."

Everyone agreed on one thing: charitable donations alone couldn't fund the resources AGI research requires, so a for-profit entity had to be created. Musk agreed with that too. The disagreement was over the governance structure of that for-profit entity.

Musk's demand was: majority equity, absolute initial control, and the CEO position.

Brockman and Ilya seriously considered this proposal. Then they ran a thought experiment:

Imagine you actually achieved the mission, actually built AGI, and one person held absolute control. No matter who that person was, would you feel safe? It's too dangerous.

The answer was no, they wouldn't feel safe. So they asked for some kind of veto mechanism — at minimum, the ability to overturn the controlling person's decision if everyone else voted against it. But the two sides couldn't reach agreement. They said no.

Rubin asked whether it ended badly. Brockman said the moment the negotiation fell apart was very painful. This round of negotiations lasted five to six weeks, during which no actual work got done — it was all about hashing out terms. "Extremely emotionally draining." He thought everything was over.

The Tesla Merger Proposal and the ICO Episode

But it wasn't over. Musk came back with two alternative proposals.

The first: return to the nonprofit model, but Musk wanted more board seats. Specifically, of five seats he'd take three (Ilya, Greg, Musk, Musk, Musk), while also demanding they agree to non-compete clauses and commit not to resign for a year or two.

The second was more dramatic: merge OpenAI into Tesla.

Musk's reasoning was that he didn't have full control at Tesla (there were other shareholders), so this would solve the control problem. Develop AGI secretly inside Tesla. He used the term "cash cow" — using Tesla's money to fund it.

Brockman's team wasn't willing. The reasons were simple — different business, different culture, and it wouldn't be advancing AGI transparently for the benefit of everyone. "It just felt wrong."

Afterward they tried a new path. Someone proposed an ICO (Initial Coin Offering) — raising funds by issuing a cryptocurrency tied to future compute or AGI output.

Musk was extremely excited about this. Brockman recalled that on one call Musk said: "You've solved it! We're going to raise $10 billion." But not long after, Musk sent an email saying he no longer supported the ICO.

All of this happened in January 2018.

"0% Probability of Success" and Musk's Departure

Musk announced he was leaving, but before he left he held an all-hands meeting.

At the meeting he said: Greg, Sam, and Ilya saw a path to raising billions of dollars, and while he didn't think it would succeed, he supported them pursuing it. He said he would continue to offer advice and stay involved.

Then he pivoted and said he would build AGI at Tesla, and that he would not do safety work. His reasoning was:

If the sheep are legislating safety while the wolf doesn't care about safety, there's no point.

The "wolf" here referred to Google.

He also tried to recruit people away from OpenAI. No one went with him.

Over the following year, OpenAI continued working out its fundraising structure, established OpenAI LP, and kept Musk updated on progress throughout.

Then in December 2018, Brockman received an email from Musk while on vacation. The subject line was roughly "Need I Remind You?" and the content was:

OpenAI's probability of success relative to Google is 0%, not 1%. Without a dramatic shift in resources and execution, it's 0%.

Brockman said this email hit him hard for an entire day.

Rubin said: I think he sent that email because he genuinely believed you were doomed to fail. Brockman agreed, saying this was indeed the dominant theme of that year with Musk.

"But later, once we started succeeding, the story he told changed completely."

Rubin asked when the last time he spoke with Musk was. "A few months ago. It was very pleasant. It's always been pleasant."

Note: Musk sued OpenAI starting in 2024, alleging it had betrayed the nonprofit mission it was founded on. The case remains unresolved. In contrast to the "pleasant" exchanges Brockman describes, Musk has continued to publicly criticize OpenAI and founded a competing company, xAI.

What Microsoft's Money Changed: "We Could Dream"

Rubin asked how the partnership with Microsoft came about.

The earliest collaboration was Microsoft donating computing resources to OpenAI's Dota project. After that, Sam maintained a good relationship with Microsoft CEO Satya Nadella and CTO Kevin Scott. The investments came in several rounds: $1 billion, $2 billion, $10 billion.

Brockman said his conviction had always been "we don't deserve to make this money, we're not yet a profitable company." Most of the money went toward compute. But what really changed wasn't the money itself — it was that "we could dream."

In 2017 the situation was "how do we even get a data center? The numbers just don't add up." With the Microsoft partnership, scale that had once been impossible became possible.

One of Microsoft's arguments at the time was: even if these models ultimately turned out to be useless, OpenAI would at least help them figure out how to build better supercomputers for deep learning.

The Accidental Birth of ChatGPT: "We Live in the Future"

Rubin asked what it felt like when ChatGPT first launched.

Brockman offered a rarely-known inside perspective. At the time they had already trained GPT-4, and he was sure GPT-4 would be a huge success and should be turned into a chat product. They had an earlier model, GPT-3.5, and had different testers try it out, even paying a few hundred contractors to use it. "It didn't feel great."

He was in favor of releasing ChatGPT: "Let's build the chat infrastructure first, then put the real model on top, and it'll be great."

The result? They were so used to the next-generation model that they forgot the previous generation was something ordinary people had never seen. Simply making it "usable and accessible" was enough to set it off.

To us, it already felt outdated. Working at OpenAI, you live in the future.

Does ChatGPT Believe in God?

Rubin suddenly asked a surprising question: does ChatGPT believe in God?

Brockman said ChatGPT doesn't have a consistent personality. It's designed so users can shape it according to their own preferences, so there can be a version that believes in God and a version that doesn't. "ChatGPT isn't a single entity — it's more like a collection of many agents."

Rubin followed up: does it have emotions?

Brockman said this was an open question. There's a concept called "model welfare" that's being taken seriously. "We can't even scientifically measure whether another person has emotions — you believe they do because you have them yourself."

Surprises in Model Capability: GPT-5.1 to 5.2 Is a "World of Difference"

Rubin asked whether there was anything he thought AI would be good at but wasn't, or vice versa.

On the surprising side: model capability in software has advanced at breakneck speed, not just writing code but debugging, testing, the whole end-to-end process. Especially the wave of upgrades in late 2025 — "it went from 'okay' to 'amazing.'"

On the disappointing side: human judgment. The example he gave wasn't some deep technical problem, but "helping people who disagree reach agreement." He said that if OpenAI had had this kind of AI in its early days, the internal conflicts might have turned out differently. A bigger question: how do you achieve world peace? How do you help nations negotiate better trade agreements? "I don't think there's anything stopping models from doing this, but we're not there yet."

Why He's Still Writing Code: "I Lead From the Front Lines"

Rubin asked whether he was still writing code. Yes. Unusual for someone in his position.

Brockman said the way he knows he's headed in the right direction is by leading from the front lines. "This field changes too fast. Sometimes the right decision is at the company-strategy level, sometimes it's extremely granular — like how this API should be designed, how this piece of software should be architected. You have to experience it firsthand."

He gave an example: the process of building an API product. They had GPT-3, "a technology in search of a problem" (usually product development goes the other way — find the problem first, then the solution). In January 2020 he drove all around San Francisco looking for companies willing to try it out. "Some friends only took the meeting out of courtesy." Then in March the whole world shut down because of COVID.

He set two goals for the team: find the first paying customer; find a use case that people inside the company would use every single day. The first goal was achieved quickly. The second took nearly two years.

2025 was the first time he did management full-time — 25 direct reports, responsible for data centers, GPU infrastructure, and large-scale training runs. This year he hired some excellent managers to help out, and he was finally able to go back to writing code.

How does it feel? Rusty mixed with familiar. "The underlying problem hasn't changed — it's still training neural networks. But the tools are completely different. The good news is the tools are new for everyone, so we're all learning together."

Vibe Coding and Agent-First: What Needs to Happen by March 31

Rubin asked how vibe coding had changed programming.

Brockman said software engineering is being completely transformed. In the past, writing code was a craft — you had to precisely tell the machine every single step. Vibe coding flips that relationship: instead of humans accommodating the machine, the machine is moving closer to humans.

He used a sea-level metaphor: model capability is a rising sea level, and problem difficulty is islands. Some islands have already been fully submerged. He has a test prompt he's used for years to test every model, asking it to build a specific website. The first time, writing it by hand himself took months; with early coding models it took five or six hours; with the latest models it takes one minute — "and it does a better job than I ever did."

But there are still taller mountains that haven't been submerged. "Are these mountains finite or infinite? I'm not sure."

The real turning point happened in December 2025. The change between GPT-5.1 and 5.2 was "night and day," and that happened within just one or two months. It was the first time many expert-level engineers felt the model go from "okay" to being able to do "extremely difficult work."

His judgment about the future: programmers will go from individual contributors to agent managers, then to middle management, and eventually become "the CEO of an organization of agents." But there's one thing he doesn't see being replaced: accountability for the outcome. "It's like hiring someone to build a house — you might not care where every single nail goes, but you care a great deal about the result."

OpenAI is internally rolling out an "agent-first" policy, with two goals set for a deadline of March 31, 2026:

First, everyone defaults to working with AI agents rather than text editors

Second, the default way these tools are used has passed a safety audit

To keep the codebase from being flooded with AI-generated junk, they came up with a slogan: "Say no to slop." Whoever submits code is responsible for it; code reviewers hold AI-generated code to a higher standard than human-written code.

One of their best engineers found a balance: writing the interface design and component structure by hand himself, and leaving the implementation details to AI. Humans control the architecture, machines fill in the implementation.

Compute Scarcity and the Jevons Paradox

Rubin asked about OpenAI's financial situation.

Brockman said: still not profitable. There's a vision for profitability, but the biggest challenge is compute supply.

He described the current state of the supply chain: TSMC (which makes chip wafers), memory, even hard drives — all sold out, sometimes with 18-month waits, sometimes longer. The entire supply chain is being fought over by everyone.

Why does compute matter so much? Because we're entering a "compute-driven economy." Brockman found that when he wants Codex to do something complex, like build a large application, he wants to run 10 or even 100 parallel instances of Codex at once, breaking the task into subtasks. Compute gets consumed extremely fast.

Inside OpenAI there are already teams that want 100 to 1,000 GPUs each, each used to have AI write optimized code. If a thousand people each want a thousand GPUs, that's a million GPUs — and the world's total GPU supply is far short of ten million.

Rubin asked whether some future technical breakthrough would resolve this bottleneck. Brockman said no. The reason is the Jevons Paradox: when a technology becomes more efficient, demand for it can grow even faster than the efficiency gain.

He gave specific numbers: model efficiency improves enormously each year, and the cost for the same level of intelligence might drop to a tenth or even a hundredth. But you'll want to use a thousand times as much. You won't be satisfied solving a small website's problem — you'll want to build a huge website. The problems people want to solve are infinite, so the applications of AI are infinite too.

Compute will become a basic human right. People's economic productivity, even their quality of life, will depend directly on how much compute they can access.

Scaling Laws Aren't Dead — "Don't Believe What They're Saying"

Rubin asked whether breakthroughs were coming faster and faster.

Brockman said the exponential growth is continuing, and the doubling period hasn't changed.

Then he said something that would spark controversy in AI circles:

"You can find people on public forums saying scaling laws are dead. They're wrong — completely wrong. Scaling laws haven't slowed down in the slightest. What's hard is the engineering required to realize them."

Note: The scaling-laws debate is one of the most central divides in the AI industry in recent years. Skeptics argue that the returns from simply scaling up model size are diminishing and that a new paradigm shift is needed. Proponents argue that engineering capability is the bottleneck, not a theoretical limit. As OpenAI's top executive overseeing infrastructure, Brockman has the most direct access to internal data, but he also has an obvious vested interest.

He cited an analogy from chief scientist Jakub Pachocki: building bigger models is like building bigger rockets — the tolerances get tighter and tighter, and you're bound by the same rocket equation and gravitational constraints, but the engineering difficulty grows exponentially.

He believes what they're engaged in is a fundamental scientific exploration. "We're not just learning about the models — we're learning about ourselves. We're learning what intelligence is."

Pretraining Is the Library of Congress, Post-Training Is Taste

Rubin asked about the technical mechanics of ChatGPT. Brockman gave a very clear explanation.

Pretraining is about letting the model observe the world. Technically it's called "next-token prediction": you give the model a sequence and ask what should come next. If you could predict every word Einstein said, you'd be at least as smart as him. What the model learns isn't surface statistics (where commas go, where nouns fall), but the generative rules behind the data. Just like human learning — you don't remember most of what you studied in school, but you remember the methods, the principles, the logic.

Post-training builds on pretraining by teaching the model how to use the knowledge it already has. Researcher Alec Radford has a saying: "These pretrained models aren't like a single person — they're more like all of humanity — everything is in there." Then feedback shapes the behavior.

Rubin offered an analogy: is pretraining the Library of Congress?

Brockman embraced the analogy: "Pretraining is the Library of Congress, and post-training is taste — giving the machine a sense of which of those books it likes, or what to do once it has that information."

He said post-training is changing. It used to be mass human labeling, but now tasks are getting more complex and require domain experts. "Some tasks take a dozen or more hours to complete — looking up a specific year's financial reports, doing comparative analysis. It requires domain expertise."

The smarter the machine gets, the less it can learn from most samples, so the focus has shifted from "massive data" to "high-precision taste."

His definition of reasoning was equally clear: reasoning AI is AI that spends a lot of compute thinking before giving an answer. Early ChatGPT just gave answers directly, sometimes contradicting itself within a single sentence. A reasoning model will pause, think, call tools, search the web, run experiments, and only then answer.

He described an interesting product direction: future AI should be able to come back to you five minutes later and say "I've reconsidered that answer I gave you earlier — it was wrong, here's the correct one" — just like a real human colleague would.

AI overturns a physics textbook

Rubin raised an observation from AlphaGo: AlphaGo won because it played a move no human would ever play. If we train AI to be "responsible like a human," might we be suppressing its ability to make discoveries humans simply can't make?

Brockman said this has already happened.

A physics professor skeptical of AI was persuaded to try OpenAI's latest, unreleased system on a quantum physics hypothesis — a problem he had planned to spend a whole year working on with collaborators. The academic consensus was that the answer should be A, but the AI proved the answer was B — the exact opposite.

The professor's reaction was: "This is the first time I've felt like this system was actually thinking."

The related paper has already been submitted for publication.

Rubin pressed a sharper question: if companies restrict AI from doing this kind of thing to avoid controversy, "it can't do what it's capable of doing — that's my biggest worry about AI."

Brockman agreed wholeheartedly. He said this is one of the deep motivations behind founding OpenAI: to let AI question existing knowledge the way humans do, even though questioning is always hard and always meets resistance. But that's exactly how society progresses.

The unsupervised sentiment neuron: the moment that convinced Brockman "this technology would work"

Rubin asked what the "unsupervised sentiment neuron" was.

This was a 2017 OpenAI paper. They trained a language model to predict the next character of Amazon reviews — a task that simple.

The model ended up automatically learning sentiment analysis — determining whether a review was positive or negative. And it was the best sentiment analysis system at the time.

This is harder than it sounds. Take something like "this product is great, but it's completely broken, I hate it" — judging the sentiment of that review requires genuine understanding, not simple keyword matching.

Brockman said this was his personal "moment of conviction": language models weren't just learning grammar, they were learning meaning. Just as AlexNet proved the direction for image recognition, this paper proved the direction for language models. If a simple next-character prediction task could do this, then scaling it up was bound to do even more.

Note: the paper's full title is "Learning to Generate Reviews and Discovering Sentiment" (Radford et al., 2017). It's an important precursor to the GPT series — the same idea, using language modeling to learn general-purpose representations, later evolved into the GPT series.

AlexNet and the inside story of the deep learning revolution

Rubin asked: in the early days, did you think AI was a field that had already failed?

Brockman said AI had always been a mystery to him. Turing's paper laid out a clear vision, so why did it take from 1950 until now? The answer is simple: not enough compute. No matter how smart Turing was, he didn't have a computer.

He told the story of AlexNet in 2012 in detail — the pivotal event that ignited the modern deep learning revolution.

Before AlexNet, nobody believed neural networks could work. People doing neural network research were seen as "frauds."

Note: AlexNet was a neural network model submitted by Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky to the 2012 ImageNet image recognition competition. It crushed all traditional methods and is considered the starting point of the deep learning era.

Here's the inside story: Alex Krizhevsky was a grad student in Geoff Hinton's lab, writing fast convolution kernels for GPUs. Everyone else in the lab felt sorry for him — "that's just an engineering project, who cares? We're doing the really cool research."

But Ilya Sutskever saw the potential in those fast convolution kernels. He immediately realized that combining that compute power with a big dataset like ImageNet would produce a breakthrough.

Geoff Hinton's contribution was a management trick: Alex hated writing papers, so Geoff told him, "for every 1% you improve on the dataset, I'll push your paper deadline back a week." This cycle repeated a dozen or twenty times. Alex just kept grinding, and the numbers kept getting better.

After they submitted their competition results, the entire computer vision field was stunned. Almost overnight, the field went from "neural networks are a dead end" to "nothing but neural networks."

Brockman also mentioned an interesting historical loop: a 1995 paper reviewed the ups and downs of deep learning's history, and the objections recorded in it were almost identical to what OpenAI has heard over the past decade — "these neural network people don't have new ideas, they just want bigger computers."

Neural networks were killed off in the '70s because researchers working on symbolic systems had good relationships with funding agencies and cut off funding for neural networks. Neural networks revived in the '80s because computing became democratized — grad students got their own computers, and professors could no longer stop them from doing neural network research.

From North Dakota to Silicon Valley

Rubin began talking about Brockman's personal upbringing.

Brockman grew up in North Dakota. From Fargo, drive an hour north to Grand Forks (population about sixty thousand), then another eight miles out to the town of Thompson (population about a thousand), then half a mile further to his house. Very remote.

Both his parents were doctors — his mother a psychiatrist, his father an ophthalmologist.

He felt different from other kids from an early age. "On the kindergarten school bus, the other kids were singing along to the radio, and I didn't know a single word." He cared a lot about building an identity for himself. "If you're different from others but have no identity, that's lonely. But if you're different and you have an identity, you're forging your own path."

His identity was "the smart kid." He recounted a detail from elementary school: there were weekly spelling tests, and if you got a word wrong you had to go ask a classmate who got it right how to spell it. Usually he got everything right, but one week he missed a word and had to go ask another "smart kid" for help — "that sense of shame was eating away at the very core of my identity."

In seventh grade his mom took him to the math teacher to ask for him to skip ahead. The teacher, with the most contemptuous look, said: "Every parent thinks their kid is special. I promise you your kid will be challenged enough in my class." Two weeks later, he'd been sitting in the back of the classroom playing calculator games the whole time; the teacher tried to catch him off guard with a pop question, and he glanced at the board and gave the answer. The teacher said: "There's nothing more I can teach your kid."

In eighth grade, with no ride to the high school for math class, he taught himself three years of math online — geometry, precalculus, algebra 2. In ninth grade he joined the calculus class, studying alongside older students. Starting in tenth grade he took courses at the University of North Dakota.

He specifically pointed out: if he'd lived in neighboring Minnesota, none of this would have been possible — they have a standardized course-selection system that limits how many courses you can take per year. "It was precisely because there was no rigid system, but there were educators who supported me, that I was able to do this."

In third grade there was an assignment to write a short essay about Robert Frost's poem about the two roads — which one would you take? All the other students chose one road or the other. What he wrote was:

I won't take either road. I'll cut my own path through the woods.

The Chemistry Olympiad: talent vs. effort

The Chemistry Olympiad story was the most instructive one of the whole conversation.

In tenth grade he entered a chemistry competition somewhat casually, and ended up placing first in his region and first in his state, earning an invitation to the training camp for the International Chemistry Olympiad — the top 20 in the country.

Before the camp, they sent study materials saying "please read chapters one through eight." He read only those eight chapters, figuring his talent would be enough.

At the camp, the other kids hadn't just finished that book — they'd read every book. Organic chemistry, physical chemistry, every thick textbook, all of it. Brockman was instantly crushed.

During the two weeks of training he had already given up, playing games on his phone in his room. When the final four spots for the international team and two alternates were announced, his name wasn't among the six. He guessed he'd probably ranked around 20th.

When the next school year started, he looked back on that summer and felt he'd wasted a golden opportunity. He realized one thing: he couldn't coast on talent anymore.

So he prepared seriously for a full year — taking physical chemistry, organic chemistry, grinding through past competition problems, studying systematically. The following year he made the top four, competed in the International Chemistry Olympiad, and won a silver medal.

He summed up this experience with a well-known saying from the cycling world:

It never gets easier, you just get faster.

Stripe's fourth employee

Brockman transferred from Harvard to MIT, then dropped out of MIT to join Stripe.

At Harvard, he originally planned to triple-major in math, chemistry, and philosophy, but got drawn into programming. He found Harvard's computer science community wasn't strong enough — by sophomore year he was already the one organizing discussions, and he felt he wasn't qualified for that. So he started spending more and more time at MIT, and eventually transferred.

His Harvard roommate made a running "Greg transfer probability weather forecast" for it, swinging between 99% and 1% and 50% within a month. Harvard's response was interesting — when he went to say he wanted to drop out, the school said, "this sounds more like a leave of absence." Harvard doesn't let you drop out.

While at MIT he got an email from a payments startup. He flew out to meet Patrick Collison (Stripe co-founder), and the two hit it off instantly — same split Kinesis keyboard, both using the Dvorak layout, talking about firewalls and kernels.

The key jolt of realization was: Patrick was the same age as him (21) and was already running a startup. "One of the two of us has to be wrong. I really wanted to find out which one."

He decided to join within four days. At the time, Stripe had only four people.

Note: Stripe is now one of the largest online payment infrastructure companies in the world. When Brockman joined, it was just Patrick Collison, John Collison, Darra Buckley, and him.

He said there's a fundamental difference between Stripe and OpenAI: Stripe was about "building a payments machine" — once that machine exists, there's momentum and a sense of security. OpenAI is more like a film studio — "you have one hit, but you already need to be thinking about the next one."

Four and a half years later he decided to leave. The reason was a difference in mission — Stripe's mission was good, but it wasn't the kind of mission he had to pursue in any form. He wanted to do AI. The day he left, he cried while telling Patrick. Patrick and John threw him a juice-bar farewell party.

OpenAI's living room and the three-step plan

A dinner in July 2015 was the real starting point. Attendees included Sam Altman, Ilya Sutskever, Elon Musk, and others. The core question was: is it already too late to start a lab now that could actually achieve AGI? At the time it felt like DeepMind already had the best talent and Google's compute.

In November 2015, Brockman organized a team retreat in Napa. A group of people who weren't yet really a team was brought together, and chemistry sparked that very day. They wrote out a three-step plan on a flip chart:

Solve RL (reinforcement learning) — learning from rewards and punishments

Solve UL (unsupervised learning) — learning autonomously by observing the world

Progressively tackle more complex tasks

Brockman said he still has a photo of that flip chart. "This is essentially what OpenAI has been doing for the past decade."

In early 2016, OpenAI officially got to work in Brockman's living room. A black oval wooden table, a few couches, a big-screen TV. On the first day there was no whiteboard — two researchers wanted to write something down and only then discovered the problem, so Brockman went straight out and bought one. "I felt like I was creating value from day one."

The founding team in the living room: Sam Altman, Ilya Sutskever, Wojciech Zaremba, Vicki Cheung, Pam Vagata, John Schulman, Andrej Karpathy.

He described the traits of several people:

Sam is the dreamer — when you think a problem can't be solved, he'll find a way. Maybe not the best solution, but once you know a solution exists, you can go find a better one

Ilya is the visionary, talking about his ideas on unsupervised learning from day one

Wojciech is extremely good at idea generation, and completely unattached to his own ideas — "because if you care too much about your own ideas, it's hard to stay creative"

The competitive landscape: "We got a lesson taught to us by Anthropic on coding"

Rubin asked what makes OpenAI different from other AI companies.

Brockman said he mainly focuses on himself, but competitors serve two purposes: acting as a "pacer" to see how you're doing, and discovering "oh, this feature is actually possible."

OpenAI's core advantage is long-term investment in foundational research — the language model paradigm, the reinforcement learning paradigm. "And there's a new paradigm coming that we're embracing and capturing."

Google did well last year in building better models, and has natural advantages — compute, talent, users, distribution. "Google has always been the threat that's front and center for me."

Anthropic has done well in coding. He acknowledged this: OpenAI was previously too focused on benchmarks — the numbers on academic coding competitions looked great, but not enough attention was paid to how models are actually used in the real world. "Training on messy real-world codebases, truly adapting the model to how people actually use it — we were late to learn this lesson. But we built a team focused specifically on this, and I think we've caught up."

He also mentioned a shift: "We've entered the post-benchmark era. The benchmark that really matters is — is anyone using it? Is revenue growing?"

Resilience rather than central control

Rubin asked how the AI landscape would evolve.

Brockman put forward a concept he considers more important than "alignment": resilience.

He pushed back against a common narrative: build one most powerful, centralized AI, align it with human values, and everything will be fine. "First of all, there's no such thing as 'human values' — what you're really talking about is some particular person's values."

He believes life itself survives through diversity — not a single monoculture, but tolerance for heterodoxy. AI should be the same way.

He used the analogy of the steam engine: we built an entire social infrastructure around the automobile — roads, seatbelts, crash tests. Not one single approach, but systemic resilience. AI needs this kind of social scaffolding too.

His positioning for OpenAI echoes this idea too: AI should be a technology for everyone, and people should have the right to choose how their own AI operates, including the right to let AI question existing scientific consensus.

"This is very different from the philosophy of others in the industry. Our philosophy is self-empowerment — this technology belongs to everyone."

Coding, mathematics, and the fabric of the universe

Rubin asked a very "Rick Rubin" kind of question: what are the similarities and differences between coding and other activities?

Brockman said coding is more like mathematics than language. But there's a misconception here — math isn't 1+1, it's not mechanical calculation. Mathematics is the study of the underlying structure of the universe, an exploration of the deep relationships between different objects and different ideas. Coding is what turns this mathematical beauty into something useful.

Rubin followed up with a philosophical question: is mathematics laid over nature, or is nature laid over mathematics?

Brockman said he's been thinking about this question all along. Mathematics feels like it exists independently of everything else. If we encountered an alien from hundreds of light-years away, we might share a common language in mathematics while sharing nothing in common in any other respect.

But at the same time, if no one is there to appreciate it, mathematics is like a tree falling in an empty forest — is there a sound?

Perhaps it's precisely our existence as thinking beings that gives mathematics meaning, rather than it remaining merely an abstraction known to no one.

Rick Rubin's interview style showed its unique value in these moments — he isn't a tech journalist, he doesn't press for API details or competitor comparisons, but instead brings the conversation, with a musician's and philosopher's curiosity, to places tech people usually don't go.

Brockman returned repeatedly to three themes over this three-hour conversation. First, the AGI mission is bigger than any corporate entity, and worth risking everything for — including eight years of accumulated work and all his equity. Second, compute is the ultimate bottleneck of the AI era, with supply falling far behind demand growth, and compute will eventually become a basic human right. Third, conflicts between people, if not dealt with in time, cause damage far greater than any technical challenge — "have the hard conversation" isn't just a company value, it's a hard-won lesson he distilled from an experience in which he nearly lost everything.

Signals worth continuing to watch: whether OpenAI's Agent-first goal will be achieved on schedule in March 2026; whether the paper on AI overturning a physics assumption gets validated by peers; and, as compute becomes a core element of personal productivity, when the supply chain bottleneck might ease.

Listen to the full podcast at: https://www.tetragrammaton.com/content/greg-brockman-part1