Sam Altman on OpenAI's origins, AI agents, and building for the future

Executive overview

Most people in AI are building the same things as everyone else — and that's the wrong bet. OpenAI itself started as eight people with no product, no revenue, and no clear path, in a field that 99% of the world thought was a pipe dream.

The core argument: the biggest opportunity in AI right now is not replicating ChatGPT but building entirely new products on top of reasoning models, memory, and agents — tools that are barely being used to their potential.

The window to build something genuinely new is open right now, and the startups that move on unexplored ground will win.

Starting OpenAI and the contrarian bet

• Starting at all was nearly a coin flip — DeepMind looked impossibly far ahead; AGI seemed like a pipe dream.
• The 1% who believed in AGI turned out to contain a lot of the world's best researchers — with nowhere else to go.
• Doing something no one else is doing is a talent magnet; doing the same thing as everyone else is not.
• Nothing that becomes big starts that way — the difference between a $0M and $0B startup is zero dollars of revenue and a few people in a room.
• Elon Musk told the team they had a "zero percent chance of success" after seeing GPT-1. That kind of doubt is normal; resilience through it is what matters.

The current product opportunity

• Model capability has entered a new realm; the products people have built to use that capability are far behind — a large overhang.
• Even if models stopped improving today, there is a huge amount left to build.
• API costs are falling fast — last week o3 cost five times more than it does this week.
• A new open-source model is coming that will surprise people on what's possible to run locally.
• Reasoning models are genuinely different from prior models; startups are only now beginning to build for that difference.
• Memory is the first feature that makes the shift from chatbot to persistent AI companion visible to users.

The vision for ChatGPT and future interfaces

• The end state: a persistent AI entity that knows you, connects to all your data, acts proactively, and surfaces information without being asked.
• Not just request-response — it will run in the background, send you messages, and act on your behalf.
• New devices and deep integrations across services are part of this, not just software.
• Jony Ive is working on new hardware; the opportunity for a third major interface paradigm (after keyboard/mouse and touch) is real.
• The ideal interface mostly melts away — the computer does the work, surfaces only what matters, and earns trust to act autonomously.
• Local and cloud compute will be mixed; pushing workloads to local devices would help OpenAI as much as users.

Agents and the agentic shift

• ChatGPT used to feel like an advanced search engine; now tools like Codex and Deep Research feel like a junior employee working on a task for a few hours.
• The fraction of knowledge work that can be done in computer-based, few-hour chunks — with a human review at the end — is large.
• MCP (Model Context Protocol) launched at OpenAI on the day of this talk; deep integration across data sources is central to the roadmap.
• Just-in-time software — LLMs generating interfaces on demand rather than static SaaS flows — is coming, and it levels the playing field for startups.

Defensibility and startup strategy

• Don't build a ChatGPT clone — OpenAI will do that better with a large head start.
• The most enduring companies are usually not doing what everyone else is doing.
• OpenAI itself had no defensibility strategy for a long time — just the only good product in the market — and built brand and memory-based moats only later.
• OpenAI wants to be a platform: traffic from ChatGPT to startups, a potential app/agent store, and a "sign in with OpenAI" that brings personalised models to third-party products.
• When the industry clock cycle changes this much, startups almost always beat big companies — lower cost, faster iteration, fewer structural advantages for incumbents.

Hiring

• Smart, driven, curious, self-motivated, team-compatible, vision-aligned — that gets you 90% of the way there.
• Early-stage: prefer young and scrappy over senior and polished; the eminent administrator rarely works as a first hire.
• Hire for slope, not Y-intercept.
• Look at what people have actually built and their velocity — not their resume or institution.

Long-term view: AI for science and radical abundance

• In 10–20 years: unimaginable superintelligence, absent something going badly wrong.
• Most personally excited about: AI for science — accelerating discovery is the highest-leverage path to improving human lives.
• All long-term sustainable economic growth traces back to new scientific discovery plus reasonably good governance.
• Energy and AI are now clearly the same vector: energy is the fundamental limiter on how much intelligence can be built.
• The correlation between energy access and quality of life across human history is one of the most important charts there is.
• The transistor is the best historical analogy for AI — society will figure out how to extract value from it, and the ramp will be faster and steeper.

Lessons from the journey

• Conviction and resilience over a long period are underrated and hard to maintain — your reserves wear down.
• Trust that it will eventually work out; most people give up after one failed startup.
• Develop and trust your own instincts; refine them over time.
• Have the courage to work on things that are out of fashion but that you believe in.
• Being a founder is like having a child: the good parts are better than you imagined, the hard parts are harder than anyone can explain — and you keep going.