Sam Altman on building OpenAI, conviction under pressure, and the path to AGI

Executive overview

Most incumbents are sleeping on AI — the same way Facebook nearly missed mobile. The window for startups to move fast and build on rapidly improving models is open now, and it won't stay that way.

Sam Altman traces OpenAI from a scrappy 2015 research bet to a company with a clear path to AGI. The core insight: deep learning scales predictably, and the right response to that is extreme concentration on one bet rather than spreading across many.

Conviction without data is a starting point, not a permanent stance — but you need the courage to operate before the data arrives.

The founding thesis

• Announced pursuit of AGI publicly at a time the field considered it irresponsible to say aloud.
• Core beliefs at founding: deep learning works, and it gets better with scale — both considered heretical at the time.
• The scaling property was initially a hunch; the data confirming its predictability came years later.
• Against better-resourced rivals like DeepMind, OpenAI's strategy was concentration: pick one thing and push hard on it.
• Early team assembled one by one over nine months; Ilya Sutskever was recruited after Sam cold-emailed him, got no response, then showed up at his conference.
• The original whiteboard goals: unsupervised learning, solve RL, stay under 120 people. Hit two of three.

The scaling bet

• Established players argued scaling was wasteful or irresponsible — that it would cause an AI winter by consuming too many resources.
• OpenAI's response: look at the results, not the critics. The results kept getting better.
• The framing that stuck internally: they had "discovered a new square in the periodic table" — something fundamental worth pushing on even without full understanding.
• Scale as an underappreciated property generally: "when in doubt, if you have something getting better with scale, scale it up."
• More is more. The field wanted "less is more" — they were wrong.

Product inflection points

• GPT-3 was impressive as a demo; almost no great businesses were built on it (copywriting was the exception).
• GPT-3.5 shifted the dynamic — startups started pulling instead of being pushed.
• GPT-4 triggered immediate GPU demand from developers, confirming product-market fit before public launch.
• The real test is always users: internal excitement is not a signal.
• Early signal on language models came from Alec Radford noticing a single neuron flipping positive/negative sentiment during unsupervised Amazon review generation — this led to the GPT series.

The five levels of AGI

• Level 1 — Chatbots: conversational AI (current widespread deployment).
• Level 2 — Reasoners: systems capable of complex, multi-step reasoning. OpenAI believes O1 reached this.
• Level 3 — Agents: systems that handle longer-horizon tasks, interact with environments, ask for help, collaborate.
• Level 4 — Innovators: scientist-level systems that explore poorly understood phenomena over time and generate genuine new understanding.
• Level 5 — Organizations: innovator-level capability at the scale of an entire company or institution.
• Sam expected the jump from level 3 to 4 to require significant new ideas; recent demos (including a YC hackathon team iteratively improving an airfoil to competitive lift) suggest current models may get further than expected through creative application alone.

How OpenAI works now

• First time Altman feels the company "basically knows what to do" — research path, infrastructure path, and product direction are clear.
• For a long time it was a true research lab operating without that clarity, which limited speed.
• Alignment of the full organization on a single direction is a significant determinant of how fast you can move.
• The company compressed a decade of typical tech company scaling drama into roughly two years.
• Teams that excel at zero-to-one are often not the right teams for one-to-ten or ten-to-hundred.

The age of intelligence and what's coming

• ASI (Artificial Super Intelligence) is potentially thousands of days away if current rates of progress compound.
• The two key unlocks for everything else: abundant intelligence and abundant energy.
• Abundant energy path: solar + storage is on a good enough trajectory even without a nuclear breakthrough; eventually physics gets solved.
• Robotics plus intelligence on tap could unlock nearly all physical labor, not just knowledge work.
• Companies of under 100 people — possibly 20 or fewer — making billions of dollars per year already appears to be happening.

Advice for startup founders

• Bet on AI now: models will improve dramatically and quickly; the saturation point is far off.
• Your edge over big companies is speed, focus, conviction, and the ability to react to how fast the technology moves — this is the number one startup advantage generally, but especially right now.
• Don't confuse short-term growth from riding a new platform with having built a durable business. The laws of business still apply.
• Everyone can build an incredible demo. Building a business is different.
• React to new models and capabilities the same day rather than putting them into a quarterly planning cycle.

Peer groups and early conviction

• Finding a high-ambition peer group early is one of the most important things a young founder can do — more important than most realize at the time.
• The value of YC was as much the peer cohort as any individual mentor.
• Stanford competed on investment banking internships; YC competed on ideas. The difference was stark.
• No one is immune to peer pressure — the only lever is choosing good peers.
• Being high-conviction and wrong is fine, provided you update when the data arrives. Many people hold conviction past the moment of data.