Greg Brockman on Compute, Sora 2 and the Path to AGI

Matthew Berman begins by asking Greg Brockman how scaling Sora 2 differs from scaling language or image models. Brockman says the modalities seem very different, but both rely on deep learning, large compute budgets and related underlying computational processes.

Brockman says Sora uses a transformer with a different training process, including diffusion-related techniques. He finds the overlap between the computations used for text and video striking, even though the modalities and detailed implementation differ.

Asked whether transformers will lead to world models or AGI, Brockman expects more algorithmic innovation as well as continued scaling of compute and data. He would not be surprised if future AGI resembled current models, but would be surprised if its architecture were exactly the same.

Different models have different inference stacks and tradeoffs between memory, computation and specialized hardware, Brockman says. OpenAI has to distribute scarce systems expertise and capacity across models, even though many of their basic operations share common machinery.

On the AMD partnership, Brockman says OpenAI had invested in software through Triton for years. Training and inference both have substantial fixed costs for a new hardware platform; he says AMD inference performance had reached a point where the company was comfortable scaling it.

Brockman highlights the MI450 series while saying OpenAI also expected to scale NVIDIA for training and inference. His description is a statement of the company’s position in the 2025 interview, not an independent comparison of hardware performance.

Berman asks about newer chip providers such as Cerebras and Groq. Brockman recalls being excited by Cerebras in 2017 but says developing non-GPU architectures proved harder than OpenAI initially expected.

Early conversations with chip companies were sometimes difficult, Brockman recalls, because designers did not necessarily share OpenAI’s expectations about the workload or the growth of large models. He argues successful hardware teams pay close attention to where deep-learning workloads are going.

Asked about the biggest bottleneck, Brockman predicts sustained compute scarcity and a major energy constraint in the United States. He says other parts of the supply chain also had not adapted to the demand shock he anticipated.

Brockman describes a shift in OpenAI’s own thinking: an early hope that AGI was primarily a software idea gave way to an emphasis on compute and then physical infrastructure. He cites Stargate and data centers as examples of that expanding scope.

He says he would prefer the market to supply the required energy rather than OpenAI having to build it itself. What the company needs to undertake, in his view, depends on how the wider market responds to expected industry demand.

Berman asks how OpenAI allocates limited GPUs among research, consumer services, enterprise offerings and developer APIs. Brockman calls the tradeoff painful because promising projects compete for the same scarce resources.

Brockman describes a split between research and applied work made by senior leaders, followed by research-level allocation. Staff on his team also perform the practical scheduling: finding projects that are winding down and moving their capacity to new ones.

Allocation combines human judgment and tools such as spreadsheets, and Brockman wonders whether models could eventually assist. Because compute strongly affects team productivity, he says the question of who receives it carries considerable emotional weight.

Brockman adds that he expects compute allocation to become an economy-wide issue, not just an internal OpenAI problem. He says expanding supply could reduce the conflicts that emerge when many valuable uses compete for limited capacity.

Berman asks how far supply falls short of demand. Brockman does not give a measured ratio; he speculates that a tenfold increase in compute might produce a smaller, perhaps fivefold, increase in revenue because many products could not yet be released broadly.

The conversation shifts to agents and apps changing web browsing. Brockman argues that finding a single fact on a static page often forces people to sift through irrelevant material, a task a machine could perform on their behalf.

He expects dynamic applications inside an AI interface to reduce some of the button-clicking involved in online transactions. The broader aim, he says, is to preserve time for human direction, judgment, feedback and creativity.

Berman asks what happens to publishers whose revenue depends on human attention and advertising. Brockman says the future monetization model is not yet known: subscriptions, advertising and other arrangements may all have a place, but products must offer value to users.

Berman raises the risk of becoming dependent on a single intermediary, as publishers did during the mobile transition. Brockman says AI may not follow one familiar internet or app-store paradigm and questions whether a single portal can serve every interaction.

The interface may move closer to the person, Brockman suggests: instead of navigating URLs, users could ask for an outcome or receive proactive assistance. He sees a large surface area for products, while acknowledging that the shape of the entry point remains unsettled.

Berman asks about the balance between reactive chat and proactive AI over the next two years. Brockman expects more work initiated in the background and describes an aspiration for systems that can productively pursue hard problems for extended periods.

As an analogy, Brockman invokes Andrew Wiles’s long work on Fermat’s Last Theorem. A long project need not mean literal isolation: it can involve subproblems and interactions, while avoiding the need for a person to micromanage every step.

Berman distinguishes how long an agent runs from how much useful work it accomplishes. Brockman agrees that benchmarks can mislead and says both dimensions matter: some difficult problems require long, productive effort, but solving them quickly would be even better.

Asked for the longest autonomous run, Brockman says he does not know the record. He has seen reports of around seven hours but does not identify that as an established limit; the discussion concerns what was known at the time of the interview.

Turning to Sora 2, Berman asks why OpenAI chose a social experience. Brockman says the team’s product choices start from a model’s capabilities and limitations, then explore which interfaces make those capabilities useful.

Brockman recalls testing a post-trained GPT-4 model with a second question dependent on the preceding exchange. Its ability to use the prior context suggested a conversational interface. He describes a similar exploration of Sora 2’s strengths and possible product directions.

A base model contains a broad range of possible behaviors but can be difficult to use reliably, Brockman says. Product interfaces and post-training make behavior more consistent, while narrowing some of the raw model’s possible outputs.

He contrasts instruction-following interfaces with an earlier base-model workflow. A next-token model completes a prefix, so a user might provide several question-and-answer examples to make another answer likely; a lone question could be continued in other ways.

Brockman says this makes raw models hard to control as products. He describes post-training as a way to shape a more consistent set of behaviors and guardrails, while acknowledging the breadth that can be lost through that shaping.

Berman asks whether the social direction was chosen before post-training. Brockman describes an iterative process: explore a base model’s capabilities, identify an interesting behavior and then work to make it reliable enough to use.

The discussion turns to Sora cameos and the use of a person’s likeness. Berman says he made his cameo public; Brockman says his was not public at the time and predicts that other video tools will make likeness generation increasingly widespread.

Brockman says OpenAI tries to introduce the technology in a way it considers beneficial while recognizing it does not control every developer building video models. His expectation about future products is a prediction, not a guarantee of their capabilities or safeguards.

Berman asks whether language alone is sufficient for a world model, citing Yann LeCun’s skepticism. Brockman looks to empirical progress: tasks once thought to require more physical knowledge can become solvable as language models improve.

They discuss simple spatial questions involving objects and containers. Berman recalls earlier models failing an example that later models answered; Brockman treats improvement on an evaluation as useful evidence, while the anecdote alone cannot settle the broader scientific debate.

Brockman says arguments about whether models truly understand can become semantic. He prefers to examine concrete evaluations and whether performance improves on tasks that skeptics predicted the systems could not do.

Berman asks whether AI threatens his job as a creator. Brockman says many jobs will change, some may disappear and new ones may emerge, but he does not claim to know their eventual balance or shape.

Brockman predicts an economy of greater abundance and a changed social contract, including a good quality of life even without conventional economic work. He emphasizes that the outcome beyond this transition is uncertain; these are his hopes and forecasts.

He expects human connection to remain difficult for AI to displace and mentions skilled trades such as plumbing and electrical work as areas where adding value may be challenging. This is his assessment, not a guarantee that particular occupations are protected.

Berman asks about platform risk for developers building on OpenAI: a provider can release a product that overlaps with its partners. Brockman says OpenAI needs developers to connect models to real-world domains and cannot serve every sector itself.

He says OpenAI chooses areas where it has expertise or useful synergies. Coding is one example: improving coding tools can help outside developers while accelerating OpenAI’s own work. That rationale describes the company’s approach, rather than eliminating platform risk.

Asked whether code is the language of AGI, Brockman instead expects natural language, perhaps a more compact dialect, to remain important. He points to readable but terse mathematical proofs as an illustration of how model-generated communication can look.

On human involvement, Brockman says the purpose of the technology is to benefit people and other living beings. He wants systems to understand a person’s goals more directly, instead of requiring users to master the mechanical details of prompting and context engineering.

Berman asks whether software and every element of an interface could eventually be generated in real time. Brockman thinks this is possible and imagines an interface designed without today’s assumptions about buttons, files and folders, while admitting he does not know what it would look like.

Brockman uses a Sora cameo to explain how a person may remain present in a generated artifact without physically performing every depicted action. He believes an image becomes more meaningful when it connects to a person, family member or pet.

Extending the analogy to software, he imagines people conceiving and sharing an experience while an AI builds much of its implementation. Berman calls this taste; Brockman agrees judgment and knowing what one wants will matter, alongside practical skill in exploring model capabilities.

Berman asks about the Agentic Commerce Protocol and Brockman’s earlier experience at Stripe. Brockman says many underlying ideas were not new, but models had become capable enough to use them more reliably than earlier systems.

He recalls that earlier ChatGPT plugins could confuse the model when too many functions were active. In his view, stronger models change the timing of ideas that had already been considered, including shopping with an agent.

Looking toward the next developer conference, Brockman hopes for models that solve genuinely hard problems, potentially alongside expert humans. He compares the ambition to a surprising AlphaGo move that changed how people understood the game.

Brockman notes that difficult tasks may consume substantial compute, so their economic value will affect whether they can be funded. Some work might also be undertaken for broader public benefit; how to deliver such systems is still an open question.

Asked about 2030 and AGI, Brockman says the term is imprecise and gives a personal estimate of roughly one to three years from the 2025 conversation, leaning toward three. He says he would be surprised if the milestone were not reached by 2030. This is a forecast, not an established timetable.

His view of AGI has shifted from a single destination to a continuous process. Matching humans on economically valuable work would be one important milestone, he says, but not the end; he stresses follow-through, safety and delivering benefits to people.