Sam Altman on Model Improvements, Agents and Startup Opportunities

At OpenAI DevDay, CEO Sam Altman answers audience questions with 20VC host Harry Stebbings. They discuss model improvements, startup strategy, agents, infrastructure and leadership. Altman's claims about OpenAI's plans, capabilities and economics are the views of the company's leader, while his predictions are uncertain.

Altman says OpenAI wants improvements across its models but regards reasoning models as a particular strategic priority. He hopes they will help with difficult coding and scientific work, and expects progress in the o-series. These are aims and expectations, not demonstrated results for all such tasks.

He expects tools for experienced programmers to become more productive first, followed eventually by higher-quality no-code tools. Building an entire startup from a simple no-code instruction, he says, will take longer; he does not give a timeline.

Stebbings asks how far OpenAI will move up the application stack and whether work such as tuning retrieval systems could become obsolete. Altman offers a general principle rather than a product-by-product roadmap: a business built only to patch a small current model weakness is exposed if future models improve.

Conversely, he recommends building products that become more valuable when models improve. He uses an imagined much more capable future model as a test of a founder's assumptions. OpenAI believes progress will be steep, but that trajectory is a company expectation, not a guarantee.

Asked where OpenAI could 'steamroll' startups, Altman predicts substantial new value in products and services enabled by AI. He distinguishes OpenAI's effort to improve general models from applications built on top, while not promising that the company will never enter particular application categories.

He recalls that many early GPT-3.5-era startups tried to work around limitations OpenAI expected GPT-4 to reduce. He now sees more founders betting on model improvements, citing potential tutoring and medical-advice applications. These are examples of possible products, not evidence of clinical safety or efficacy.

Stebbings asks about a proposed $9 trillion capital-expenditure figure and $9 trillion annual economic benefit. Altman declines to validate the precise numbers, saying even order-of-magnitude estimates are difficult. He expects large investment and value creation, but the amounts and timing are speculative.

He imagines a future in which someone can describe a company's software and have it built much more cheaply, while emphasizing that such a capability is still some way off. He sees similarly large potential in education and healthcare if AI can deliver genuinely useful outcomes.

Altman sees a place for open-source models alongside integrated services and APIs. He says users can choose the delivery mechanism that fits them; his answer does not announce a commitment to release a particular OpenAI model as open source.

His off-the-cuff definition of an agent is a system given a long-duration task that needs minimal supervision while executing it. He cautions that the field does not yet have a clear intuition for what agents will ultimately become.

A restaurant booking illustrates one possibility: an agent might compare many restaurants in parallel, especially if other agents answer them. Altman finds the more consequential possibility to be a capable collaborator that works for days or weeks, asks questions when needed and returns useful work. Both are hypothetical examples, not claims that such systems already work reliably.

Asked how agents might change per-seat software pricing, Altman explicitly says he is speculating. One possibility is payment based on the compute continuously working on a customer's problems, rather than a fixed price per human seat or per agent. He also expects significant infrastructure and scaffolding around agentic systems.

Altman agrees that individual models depreciate as newer ones arrive, but argues that OpenAI's revenue can justify training costs. He says learning to train one model can help the next and that ChatGPT's large user base helps amortize costs. He acknowledges that similar economics may not hold for every model developer; no audited unit economics are provided.

Reasoning remains OpenAI's most important current focus, he says, alongside multimodal capabilities and other features. He hopes reasoning will extend to images and expects progress in vision models, without disclosing specific product commitments.

He declines to detail OpenAI's research methods, saying it is easier to replicate a result after someone has shown it is possible than to make an unproven discovery. He regards a culture that repeatedly pursues new work as one of the organization's strengths—an assessment from its CEO.

Altman thinks talent is wasted when capable people lack opportunities, whether because of their organization, location or life circumstances. He hopes AI could help more people reach their potential and says he would like eventually to write about building organizations that support original work.

Rapid growth forced OpenAI to learn how to serve customers and build a company faster than a conventional startup might. Altman says he underestimated how much work it takes to pursue the next tenfold expansion while still handling current operations.

That requires internal communication, planning and coordination across different time horizons. Compute capacity, research, products and even office space can have long lead times; a team cannot simply repeat last week's work and expect to operate at ten times the scale.

Stebbings raises advice to hire very young people. Altman rejects an age-only rule: talented people early in their careers can bring fresh ideas, while very complex infrastructure may need deep experience. He wants a high talent bar across ages and says inexperience does not mean a person lacks value.

Asked about Anthropic's coding models, he acknowledges that they are strong and says developers often use multiple models. He expects discussion eventually to move from individual models toward larger systems, though he says that transition will take time.

Altman believes the trajectory of model capability will continue improving for a long time, while acknowledging past uncertainty, failed training runs and difficult research transitions. He recalls unresolved problems during GPT-4 development and a long research path to o1 reasoning models; the future trajectory remains his prediction.

During setbacks, he says a shared mission to build AGI helps motivate the team, and a belief in deep learning has sustained it through obstacles. He does not suggest progress has been a straight line.

Many leadership decisions reach him because they are close calls. Rather than rely on one adviser for every question, Altman tries to consult a range of people with strong judgment and context in different domains. Some high-stakes product or infrastructure choices are difficult to reverse.

Semiconductor supply and international tensions concern him, though he does not identify them as his single largest worry. He describes a broader challenge: aligning power, networking, chips, research readiness and products capable of supporting expensive infrastructure.

Asked whether $100 billion is required to enter frontier model development, he thinks the cost can be lower without giving a precise estimate. He cautions against taking analogies with the internet or electricity too literally: building models is capital intensive, while many application companies can use models supplied by others.

Of imperfect historical comparisons, he favors the transistor: a general capability that spread through many products, supported by expensive industrial processes and complex supply chains. He uses it to illustrate possible broad effects, not to claim that AI will follow the same history.

If starting a company, Altman says he might build an AI-enabled vertical product such as a tutor, legal tool or engineering assistant. These are startup ideas and aspirations; they do not establish that current systems can safely or reliably perform every role.

He would like AI that can understand a person's life and use relevant personal data, though he does not prescribe a specific technical approach. He also praises Cursor's user experience and highlights OpenAI's realtime API; the latter is an OpenAI product he has an interest in promoting.

On latency versus accuracy, he favors a user-controlled trade-off. A quick conversational answer may be appropriate in one setting, while a difficult research task might justify much more time. He frames this as a desired product capability rather than a guarantee of accuracy.

Altman says he feels uncertain about details of OpenAI's product strategy and considers product leadership an area he wants to improve. He praises product leader Kevin for discipline, focus and deciding what not to build.

Looking five years ahead, he predicts rapid advances in technology and science while society may change more slowly in the near term. He invokes a rough Turing-test comparison, but his characterization and timeline are debated; neither establishes a measured test result or a guaranteed societal outcome.

Stebbings and Altman close the onstage conversation by thanking the audience. Altman's longer-term scenario remains a forecast, with substantial uncertainty about capability, adoption and effects.