Altman calls this the best time yet to start a technology company, since major shifts such as AI can give upstarts an advantage over large companies. He sees abundant intelligence and energy as foundations for progress in both ideas and physical work. He says progress toward superintelligence could continue compounding for years, while acknowledging that this is a hope and a guess.
How To Build The Future
Speaker
Sam Altman, OpenAI CEO, discusses AI, startups, and technology’s future with a Y Combinator host on Y Combinator’s rebooted How to Build the Future.
Altman recalls that OpenAI openly aimed for AGI when that goal seemed implausible to many in the field. The team believed deep learning improved with scale and concentrated on those ideas despite having fewer resources than DeepMind. He says conviction helped guide action when evidence was limited, but teams should change course when results show they are wrong.
Altman says he was drawn to people who felt free to try things rather than accept how things were. He credits YC’s founder community with giving him ambitious peers, and advises young people to find that kind of group early. His first company, Loopt, did not find product-market fit, but gave him a rapid education in building a startup and navigating a platform shift.
Altman describes OpenAI’s early research as a series of bets and detours, including unsupervised learning, reinforcement learning, robots, and video games. He says GPT-4 was a moment when the team saw strong potential, but customer response was the real test: users quickly asked for more GPUs. He also describes OpenAI’s current work as more focused, with clearer research, infrastructure, and product goals helping the company move faster.
Altman outlines five levels of AI capability: chatbots, reasoners, agents, innovators, and systems that can perform at the scale of an organization. He advises founders to build with AI because models are improving and startups can react faster than companies tied to quarterly planning. He cautions that an AI demo is not a business: founders still need to create a valuable product and a competitive edge.
Takeaways
- Find peers who encourage you to attempt ambitious work, and build that circle early.
- When resources are limited, concentrate on a promising approach instead of spreading effort across too many bets.
- Use conviction to act before the evidence is clear, then adapt when results arrive.
- Test new technology with real users; an impressive demo alone does not establish demand.
- Move quickly with AI, but build lasting value and a competitive edge.
Mentioned in this talk
Search inside this talk
What was claimed 15
- built
“GPT-4 was one of the few moments where I thought, 'Okay, we have something really great on our hands.'”
YouTube Verified3 sources· Talk, Web· archived
- built
“they'd fine-tuned GPT-3 to write code.”
2:33inAnthropic Head of Pretraining on Scaling Laws, Compute, and the Future of AIYouTube Verified4 sources· Talk, GitHub, Web· archived
- builtAlec Radford built GPT-1
“Alec really had to persevere for years to make it work”
3:56inBob McGrew: AI Agents And The Path To AGIYouTube Verified2 sources· Talk
- feature
OpenAI: Pursued AGI from its beginning“from the very beginning, we said we were going after AGI”
YouTube Unverified1 source· Talk
- featureOpenAI: Early research goals included unsupervised learning and solving reinforcement learning
“figure out how to do unsupervised learning, solve RL”
YouTube Unverified1 source· Talk
- featureOpenAI: Believed deep learning worked and improved with scale
“The core beliefs were that deep learning works and gets better with scale.”
YouTube Unverified1 source· Talk
- featureGPT-4: Could handle legal workflows when prompts were divided into sufficiently small steps
“if he broke the prompts into small enough workflow steps, he could get it to do exactly what he wanted”
YouTube Unverified1 source· Talk
- featureGPT-3.5: Brought increased startup interest in using the model
“Then GPT-3.5 came along, and people—startups, YC startups in particular—started to show interest.”
YouTube Unverified1 source· Talk
- metricGPT-4: Users quickly asked how many GPUs they could get after receiving access
“we got the 'How many GPUs can you give me?' moment very quickly”
YouTube Unverified1 source· Talk
- featureLoopt: Mobile service for sharing friends’ locations
“It was basically Find My Friends geolocation”
YouTube Unverified1 source· Talk
- metricLoopt: Did not find product-market fit or a way to reach escape velocity
“we never really found product-market fit or a way to reach escape velocity”
YouTube Unverified1 source· Talk
- metricOpenAI: ChatGPT was less than two years old
“ChatGPT is less than two years old”
YouTube Unverified1 source· Talk
- featureOpenAI: The company’s research and infrastructure paths were fairly clear, and its product path was getting clearer
“our research and infrastructure paths are fairly clear, and our product path is getting clearer”
YouTube Unverified1 source· Talk
- featureCamper: Demonstrated iteratively improving an airfoil design using current AI models
“they were able to build something that would iteratively improve an airfoil”
YouTube Unverified1 source· Talk
- featureAI agents: Can carry out longer-term tasks with multiple environment interactions and ask people for help
“go off and do these longer-term tasks, maybe involving multiple interactions with an environment, asking people for help”
YouTube Unverified1 source· Talk
