OpenAI President Greg Brockman: AI self-improvement, the Super App gamble, the road to AGI, and the compute expansionOpenAI 总裁 Greg Brockman：AI 自我改进、Super App 豪赌、通往 AGI 之路、算力扩张

When Greg Brockman sat down in the Big Technology Podcast studio, OpenAI had just done three big things: killed Sora, secured $110 billion in funding, and finished pretraining its next-generation model, Spud. Cutting a marquee product, raising a sky-high round, and betting on a new model — three things happening at once, each demanding an explanation. Over the next 90 minutes, the OpenAI co-founder and president had a difficult task: repackage every seemingly reactive decision as a forward-looking strategic choice.

Did he pull it off? Mostly yes, but the cracks in the narrative hide some clues worth examining closely.

For the past 18 months, Brockman has mainly run the company's "Scale" division, overseeing GPU infrastructure, data centers, and supply chains. This interview spans an enormous range of topics: from why Sora was cut, to what the Super App will look like, to what the next-generation model can do, to how the $110 billion will be spent, to why he donated $25 million to MAGA Inc.

Original video: https://www.youtube.com/watch?v=J6vYvk7R190

Key takeaways

Was killing Sora "focusing on the core technical line" or stopping the bleeding? Brockman says Sora and GPT are different branches of the technology tree, and with limited compute you can only pick one. He didn't mention that Sora was burning about a million dollars a day and that its user base had fallen below 500,000.

The Super App will merge ChatGPT, the coding agent Codex, and the browser Atlas into one unified app, expected to ship within the next few months — but this vision directly overlaps with Microsoft 365 Copilot, and no one in the interview asked about that.

The next-generation model Spud crystallizes roughly two years of research breakthroughs. Brockman says it will raise both the ceiling and the floor of AI capability, but the descriptions are all impressionistic, with no concrete metrics.

He self-rates AGI progress at 70-80%, but also says the definition of AGI is "more like a vibe" — he never resolves the contradiction of precisely quantifying something that can't be precisely defined.

He admits OpenAI once lagged behind Anthropic on "last-mile usability" for coding tools, and claims they've now caught up — though third-party data shows Anthropic's enterprise market share is still growing rapidly.

Responding to Anthropic CEO Dario Amodei's criticism that OpenAI's infrastructure investment is "too aggressive": "I disagree." But he didn't address Amodei's specific argument that a one-year miss on revenue projections could lead to bankruptcy.

Killing Sora: a technical choice or a business bleed-stop?

Kantrowitz opened by asking: OpenAI leads in the consumer market, so why suddenly pull resources away from video generation?

Brockman's explanation framework differs quite a bit from outside speculation. He said Sora's model and the GPT series are "completely different branches on the technology tree," built in fundamentally different ways. In a world of limited compute, pushing two technology paths at once is extremely costly. He used a vector metaphor:

The sum of random vectors is zero — only aligned direction lets you move forward.

Scattered force cancels itself out; only concentrated force generates momentum. He believes the problem in deep learning isn't a shortage of opportunities but the opposite — too many — so spreading investment thin means not going far in any direction.

This explanation is internally consistent on technical logic, but it's only half the story. According to WSJ investigations and TechCrunch reporting, Sora had already failed commercially: its user count had plunged from a peak of one million to under 500,000, burning roughly $1 million a day in compute costs. Per WSJ reporting, Disney's planned $1 billion investment and character licensing deal with Sora was also called off, and Disney learned of the news less than an hour before it was announced publicly. Shutting down Sora had both a technical-focus rationale and real commercial pressure to stop the bleeding — Brockman chose to talk about only the former.

He stressed this isn't "shifting from consumers to enterprise" but a forced choice. He believes the two most important applications right now are: a personal assistant that understands you and is aligned with your goals, and an AI that can solve hard problems for you. Just these two things already exceed OpenAI's current compute capacity.

One technical detail: the image generation feature in ChatGPT is unaffected. Image generation is built on the GPT architecture, on the same technical line as text and voice; Sora uses a diffusion model, a different branch entirely. The Sora research team wasn't disbanded — it pivoted to world-simulation research for robotics.

Kantrowitz asked further: Google DeepMind's Demis Hassabis believes the model closest to AGI isn't a text model but an image generator, because it must understand how objects interact and how the world works. By abandoning that path, is OpenAI betting on the wrong track?

Brockman said this is a real risk: "Absolutely. In this field you have to make choices, you have to place bets." But he then pivoted, saying OpenAI is confident because it has "already seen the path to AGI." He gave an example: a physicist handed a long-unsolved problem to an OpenAI model, and 12 hours later got a solution — the physicist said it was the first time he felt the AI was "thinking." However, Brockman didn't reveal which physicist or which problem, so the case can't be independently verified. [Note: in public information, mathematician Terence Tao published a paper in March 2026 whose core proof was completed by ChatGPT Pro, but that was in mathematics and doesn't directly correspond to the case described here.]

The Super App: a grand vision, but who answers the Microsoft question?

Having cut video generation, OpenAI is placing its bet elsewhere. Kantrowitz asked what the Super App actually is.

Brockman said: integrating ChatGPT (chat), Codex (coding agent), and the browser into one app.

The computer should adapt to the person, not the other way around.

He compared the Super App to a laptop: is your laptop for personal use or work use? Both. The Super App is the same — both a personal assistant and a work tool. The personal side includes everything ChatGPT already does, plus deeper memory and context capabilities. ChatGPT already has a feature called Pulse (personalized content push) that delivers content daily based on what the AI knows about you.

But Brockman said what users see is just "the tip of the iceberg." What matters more is unifying the underlying technology: over the past two years, AI development has shifted from "just looking at the model" to "looking at the whole system" — how the model gets context, how it connects to the outside world, what actions it can take. OpenAI internally had multiple separate implementations of these things; now they're being consolidated into one general-purpose "AI layer." In general, users shouldn't need a dedicated finance AI or legal AI — the Super App should be general enough. If this vision is realized, it could genuinely redefine what an AI product looks like.

The problem is that Microsoft 365 Copilot is already doing almost exactly the same thing — embedding AI into workflows, unifying entry points, connecting various tools. OpenAI is simultaneously Microsoft's model supplier and its direct competitor — how is that tension resolved? Kantrowitz didn't ask, and naturally Brockman didn't answer. Add to that the fact that Sora just proved OpenAI's track record on consumer products isn't exactly stellar, and the "ship within a few months" timeline deserves a question mark.

Codex: how far is the road from programmers to everyone

If the Super App's ambitions are to be realized, Codex is one of the key pieces. It was originally designed for software engineers, but OpenAI has already seen a lot of spontaneous use by non-engineers internally. Brockman said Codex is fundamentally two things: a general-purpose agent framework (capable of calling tools) plus an AI that can write code. Connect it to spreadsheets or Word documents, and it can do knowledge work.

Specific examples: someone used Codex for video editing, and the AI wrote an Adobe Premiere plugin itself to automatically chapter the footage; someone on the internal communications team connected Codex to Slack and email to consolidate feedback. If true, these cases suggest AI's capability as "universal glue" connecting different tools is taking shape. But these are all internal use cases recounted by Brockman, with no external user verification, and no mention of whether they're publicly available.

Brockman said that for non-programmers, Codex's usability is currently low — if you hit an error during setup, a developer knows how to handle it, but an ordinary user is just stumped. But he believes the hardest part is already done (building a truly intelligent AI), and what remains is the "much easier" part: lowering the barrier to entry. Whether this holds depends on how you define "hard." For people who've built consumer products, the "last mile" is often harder than the ninety-nine miles before it.

He also mentioned OpenClaw founder Peter Steinberger joining OpenAI, as a signal of Codex expanding to non-programmers. Steinberger is an Austrian developer who, in late 2025, built an open-source AI agent that can autonomously manage email, order food, and control smart-home devices, reaching 196,000 GitHub stars and 2 million weekly active users. In February 2026, Altman announced Steinberger had joined OpenAI to lead "next-generation personal agents" — the role itself signals that OpenAI is taking "getting AI beyond the programmer crowd" seriously.

Catching up to Anthropic: admitting falling behind, claiming to have caught up

Kantrowitz asked: Anthropic already has Claude chat, Claude Cowork, and Claude Code — essentially having built its own Super App first. Did OpenAI only wake up to this after seeing what Anthropic did?

Brockman admitted that OpenAI has historically had the best competition results in coding, but underinvested in "last-mile usability." A model can be brilliant in coding competitions but has never seen the messy state of real-world codebases.

Around the middle of last year, OpenAI formed a dedicated team to tackle this: studying all the "dirty work" of real-world codebases and building training environments that simulate the messy scenarios disrupted in strange ways. He said that at this point they've caught up, and in head-to-head comparisons users increasingly prefer OpenAI. But he also said the front-end experience still lags.

Third-party data paints a different picture. According to reporting from firms like Synergy Research, Anthropic's enterprise market share rose from a low base in 2023 to about 40% by late 2025, while OpenAI's fell from 50% to about 27%. Claude Code has earned extremely strong word of mouth among software engineers. When Brockman says they've "caught up," what exactly is being compared — model capability or market share? He cleverly confined the battlefield to the former.

The bigger change is internal. Brockman said OpenAI used to treat research and deployment as "almost two separate things," and is now integrating them into one process — thinking about how the product will use it even while doing the research. This might be the single most substantively informative line in the whole conversation, hinting that OpenAI's prior organizational structure really did have problems.

"We won" is the scariest moment

Shifting from product competition to internal culture, Kantrowitz changed angle. He asked: OpenAI has been leading since 2022, and now competition has intensified — has the internal mood changed?

Brockman told a story. At the holiday party after ChatGPT launched, he sensed a "we won" atmosphere pervading the company. That was the moment that scared him most.

No, we're the challenger — we always have been.

This is standard tech-executive rhetoric — a $300 billion company calling itself a challenger, much like Google once called itself a startup; believe it or not as you like. But he followed up with something more grounded: when things are going well, don't believe people telling you how great you are; when things are going badly, don't believe people telling you how bad you are either.

Spud: it feels right, but where's the data?

Beyond culture, next comes technology. On the next-generation model Spud, Brockman confirmed it's a new pretrained foundation model, crystallizing "roughly two years of research results."

He first corrected a framing: what matters isn't any single model, but the whole "engine of progress." Model development happens in stages: first pretraining, training a base model on vast amounts of data — the most expensive stage; then post-training, using reinforcement learning to teach the model to solve various problems; and finally tuning behavior and usability. Spud is the result of a new round of pretraining, and will go through extensive post-training optimization afterward.

As for what Spud can actually do, Brockman gave no concrete metrics. He coined a concept called "big model smell" — when a model is truly strong enough, you can "smell" that quality. You ask it a question and it doesn't go off track, doesn't need you to explain things repeatedly. He said Spud will raise both the ceiling (solving more open-ended, longer-horizon problems) and the floor (usefulness on everyday tasks) at the same time.

No benchmark scores, no comparisons with competitors, no task-specific demo data. This purely qualitative description may be for confidentiality reasons—keeping models under wraps before launch is standard in the industry. But at a time when competitors are frequently showing off benchmarks, relying on "feel" alone has limited persuasive power. OpenAI deliberately withheld certain benchmark details when GPT-4 launched too, and later won back the narrative through actual product performance. Whether Spud can follow the same path depends on how it performs after launch.

December 2025: The Inflection-Point Narrative

Kantrowitz asked: what happened in December 2025? Coding agents seemed to move from theory to practical use.

Brockman said that model release pushed AI from being able to do "20% of your tasks" to "80%." It went from a "nice-to-have" to something where "you have to reorganize your workflow around AI."

He shared a personal test he'd run for years: having AI help him build a website—one that took him months to build back when he was learning to code. For most of 2025, it took four or five hours and multiple rounds of back-and-forth. In December, one prompt, done in one shot.

Slowly, slowly, slowly, then all at once.

This phrasing comes from Hemingway's The Sun Also Rises, where the original line was about "how did you go bankrupt." It's an apt way to describe the nonlinear growth of AI capability.

Another example: an engineer he works closely with, doing low-level systems engineering, couldn't use AI at all in the GPT-5.2 era. By 5.3, given a design document, the AI could implement the code, add monitoring metrics, run a profiler, and optimize on its own—producing exactly what the engineer wanted.

On the point that "public reaction to the GPT-5 launch was somewhat disappointing," Brockman said every release produces two camps: one that sees a night-and-day difference, and another whose use cases weren't intelligence-bottlenecked, so they don't notice a difference. The real shift is that people's mental models of AI capability update more slowly than the technology itself.

AGI: 70-80% There, But the Definition Is a "Vibe"

Nvidia CEO Jensen Huang recently said, "I think we've already achieved AGI." Kantrowitz asked Brockman if he agreed.

Brockman said AGI means different things to different people. Current technology is "very jagged": it already surpasses humans at tasks like writing code, but AI still fails at basic tasks that humans handle easily. Where you draw the line is "more of a vibe and a feeling than a science."

He gave his own self-assessment: "70-80% of the way there."

This self-contradictory statement actually reflects a dilemma facing the whole industry. If the definition of AGI isn't even scientific, more of a "vibe," then how was 70-80% calculated? Everyone is quantifying a concept they themselves admit they can't precisely define. Huang's definition of AGI is extremely narrow—he himself admits that 100,000 agents couldn't build another Nvidia; Lex Fridman has proposed "can it found and run a billion-dollar company" as a test. Brockman is more cautious than Huang, but falls into the same definitional trap.

He believes AGI will arrive within the next few years, with AI able to do nearly all the intellectual work you do on a computer, though capability will remain jagged.

Automated AI Researcher: Using AI to Accelerate AI

OpenAI plans to launch an automated AI researcher this fall.

Brockman calls the current stage "takeoff"—a term with a specific meaning in AI safety circles, referring to the stage where AI enters rapid recursive self-improvement. Using it in this context at least shows he's aware of its weight. His meaning has two layers: technically, the better AI gets, the faster it can accelerate its own improvement; in application, chip makers are ramping up investment and companies across every sector are exploring AI applications—all that energy is accumulating.

As for the AI researcher specifically, Brockman's definition is: taking the complete end-to-end work of an OpenAI research scientist and running it on silicon.

But he stressed this doesn't mean just letting AI run on its own. He used an analogy: it's like managing a junior researcher—leave them unsupervised too long and they'll wander down useless paths. Senior researchers provide direction and review; AI does the execution.

In the Agent Era, Human Responsibility Can't Be Outsourced

Kantrowitz quoted something Brockman had said before: when you use AI agents to get things done, "you become the CEO of a fleet of thousands of agents executing your goals and vision, but you don't understand exactly how everything gets solved. This new way of working might make you feel like you've lost the pulse on the problem."

Brockman said this cuts both ways, and the key is to "acknowledge the strengths, mitigate the weaknesses." AI agents give you enormous leverage, but ultimately there's a "responsible party." If your agent messes up a website and affects users, that's not the agent's fault—it's yours.

Kantrowitz asked: but how do you reconcile "losing the pulse" with "being responsible"?

Brockman used a home renovation analogy: you hire a general contractor to renovate your house, and there are details you genuinely don't need to worry about because you trust the professionals. But if something goes wrong, you should know about it. "You can't just blindly accept 'I can stop paying attention.' You need to actively stay on top of the problem."

The $110 Billion Compute Gamble: Revenue Engine or Casino Chips?

Back from technical roadmap and product vision to business reality.

Brockman compares compute to hiring salespeople: as long as the product sells and there's a scalable way to sell it, more salespeople means more revenue. Compute isn't a cost center, it's a revenue center. He recounted an internal conversation on the day ChatGPT launched:

"How much compute should we buy?" "All of it." "No, no, seriously, how much should we buy?" "No matter how much we build, I know it won't keep up with demand."

He said every year since then has proven that call right. ChatGPT currently has over 400 million weekly active users, and consumer subscription revenue keeps growing—so from the demand side, that judgment does hold up. But the challenge is that compute purchases have to be locked in 18 to 24 months in advance, meaning you have to predict the future with precision.

As for revenue mix, consumer subscriptions are still the largest source for now, but enterprise knowledge work is growing extremely fast. He doesn't think future revenue will simply split into "consumer" and "enterprise" categories—it's more that users have an entry point into a digital world, and revenue comes from that entry point.

Kantrowitz brought up Dario Amodei's suggestion that some competitors are burning cash "YOLO-style," turning the "risk dial too far."

Brockman's response: "I disagree. We've always been extremely deliberate, and we were among the first to recognize this trend and position ourselves ahead of it."

But after "I disagree," he didn't actually engage with Amodei's core argument: if revenue projections are off by even a year, the company could go bankrupt. That's a question that needs numbers, not attitude—and the numbers aren't entirely on Brockman's side: of the $110 billion in financing, only about $25 billion is confirmed as immediate cash; the rest is either compute (Nvidia's $30 billion is mostly GPUs) or comes with milestone conditions (Amazon's $35 billion). According to estimates from outlets like The Information, OpenAI's annual cash burn will reach $57 billion by 2027. The financing figure looks large, but the actual runway may only be 18-24 months. By comparison, Anthropic has likewise committed to $50 billion in infrastructure, but is pacing it strictly to revenue growth.

On whether even larger-scale pretraining is still needed, Brockman's position is clear: improving pretraining makes every downstream step easier. But he also said there's been an important shift over the past 24 months: no longer just chasing raw capability, but also considering inference efficiency at the same time, aiming to find the optimal point of "intelligence × cost."

Safety and Public Trust

On AI risk, Brockman proposed a "resilience" framework: many participants jointly develop the technology, while also building social infrastructure to ensure the technology heads in the right direction. He used electricity as an analogy: many people generate electricity, and electricity is dangerous too, but we've built a whole system of safety standards, regulation, and inspectors around it.

On negative public sentiment toward AI (YouGov data shows Americans who think AI has had a negative effect on society outnumber those who think it's had a positive effect by three to one), Brockman said the key is showing people concretely how AI helps them. He gave a medical example: a family's child had symptoms like headaches, and their insurance denied an MRI; using ChatGPT to research the symptoms, they found arguments that persuaded the insurance company, got the MRI, discovered a brain tumor, and treated it in time to save the child's life. He said stories like this "happen every day."

On the data center controversy, Brockman claimed that OpenAI's Abilene facility—one of the world's largest supercomputers, with over 100,000 GPUs deployed—uses "about as much water annually as a single household." This claim is questionable. For comparison, Microsoft's 2024 environmental report disclosed global data center water use of about 7 billion liters annually, and Google's figures are on the same order of magnitude. Even with a state-of-the-art air-cooling setup, the heat load from 100,000 GPUs far exceeds what a single household's water use could plausibly handle, and this claim has not been independently verified. On electricity, he pledged that OpenAI would "foot the bill itself" rather than driving up local residents' electricity rates, citing North Dakota as an example where the arrival of data centers actually helped lower electricity costs—a claim likewise lacking third-party verification.

Political Donations: The $25 Million Controversy

On the controversy over his $25 million donation to MAGA Inc, Brockman said he and his wife have also donated to bipartisan super PACs. He describes himself as a "single-issue donor" focused on supporting politicians who "genuinely embrace AI technology." But MAGA Inc. is a PAC that broadly supports Trump's agenda, not a dedicated tech-policy organization—there's a clear tension between the "single-issue" framing and the recipient of the donation. Brockman also donated $25 million to Leading the Future, a bipartisan PAC promoting AI development. The $25 million was the largest single donation MAGA Inc. received in the second half of 2025, and it sparked the #QuitGPT boycott movement in early 2026.

What This Interview Really Reveals

Unpacking the narrative Brockman built in this interview, you can see OpenAI facing pressure on three fronts.

The first is product pressure. Cutting Sora did stop the bleeding, and ChatGPT remains the world's largest AI product by user count, but the next bet, the Super App, faces a far harsher competitive environment than Sora did. Microsoft 365 Copilot on the enterprise side, Anthropic on the developer side, Google on the consumer side—OpenAI has to build a "unified entry point" while being squeezed from all three directions. Brockman's response to this is "ship it within months," but the lesson of Sora is that having the technology doesn't mean you can execute a good product.

The second is a communication bind. Throughout the interview, whenever Brockman was asked the hardest questions, he repeatedly substituted emotional language for data. Spud's capabilities are judged by "feel," AGI progress by "vibe," and the evidence for catching up to Anthropic is an undisclosed internal comparison. This reflects a structural contradiction facing OpenAI right now: as a company that just raised $110 billion, it needs to prove to investors that it's technologically ahead, but it can't leak specific metrics before a product launch. So "feel" becomes the only safe mode of expression. The problem is that feelings can't substitute for facts, especially when competitors are speaking in benchmarks and market share.

The third is scale risk. The $110 billion financing figure sounds intimidating, but it's less secure once you break it down. Brockman's rebuttal to Amodei's "YOLO" criticism stopped at "I disagree," never directly addressing the core question: what happens if demand growth comes in a year slower than expected? He himself said compute purchases have to be locked in 18-24 months ahead—which happens to be exactly the runway analysts estimate OpenAI actually has.

Brockman's advice to AI skeptics is "just try it first." But whether a $110 billion bet is the right call isn't a question a personal trial experience can answer. Over the next six months, whether Spud delivers on its promises and whether the Super App can ship while squeezed between Microsoft and Anthropic will determine whether this narrative turns out to be vision or spin.

Original video: https://www.youtube.com/watch?v=J6vYvk7R190