Sam Altman: OpenAI CEO on GPT-4, ChatGPT, and the Future of AI | Lex Fridman Podcast #367

Lex Fridman talks with Sam Altman about what GPT-4 is, how reinforcement learning with human feedback (RLHF) made ChatGPT usable, and why OpenAI builds in public to let society adapt. They dig deep into AI safety, alignment, bias and the politics of who decides a model's values, plus consciousness, jailbreaking and the system message. The conversation broadens to AGI takeoff scenarios, the economic and political transformations ahead (UBI, falling cost of intelligence and energy), the impact on programming and jobs, and OpenAI's unusual capped-profit structure. Altman also addresses Elon Musk's criticism, the Microsoft partnership, Silicon Valley Bank's collapse, and offers reflections on truth, power, and meaning.

• Altman says the pivotal moment wasn't the model but ChatGPT's usability, driven by RLHF and the interface, not the underlying capability.
  00:05:41
• States his core principle: he cares that the degree of alignment increases faster than the rate of capability progress.
  00:24:22
• Describes a 'platonic ideal' where every person on Earth deliberates, constitutional-convention style, on where to draw an AI's boundaries.
  00:35:17
• On Eliezer Yudkowsky's claim AI could kill all humans: 'I think that there's some chance of that and it's really important to acknowledge it.'
  00:55:02
• Admits ChatGPT surprised the team, but GPT-4 'has weirdly not been that much of an update for most people.'
  00:58:41
• Says his nearer-term worries are disinformation and economic shocks, not a machine 'waking up' and deceiving us.
  01:09:39
• Insists any version where one person controls the technology is 'really bad' and he wants no special voting or board control.
  01:18:29
• Walks back his own famous 'How to Be Successful' advice, saying he 'mostly got what I wanted by ignoring advice.'
  02:18:33

• RLHF aligns models with remarkably little data compared to the massive pre-training process, which is far less understood scientifically.
  00:06:13
• Going from GPT-3 to GPT-4 was not one breakthrough but 'hundreds of complicated things' multiplied together across data, training and architecture.
  00:43:37
• The viral '100 trillion parameters' GPT-4 meme originated from one of Altman's own YouTube talks, taken out of context.
  00:43:50
• GPT-4 was called possibly 'the most complex software object Humanity has yet produced' that will be trivial to make in a couple of decades.
  00:45:11
• Altman's point that despite Deep Blue beating Kasparov ~25 years ago, chess has never been more popular than it is now.
  00:52:26
• Ilya Sutskever's test for consciousness: train a model with zero mention of consciousness, then see if it recognizes a described subjective experience.
  01:05:01
• OpenAI's capped-profit cap started at 100x for early investors and is much lower for new ones; everything beyond flows to the nonprofit.
  01:14:21
• Altman helped start Worldcoin and OpenAI funded what he calls possibly the largest, most comprehensive UBI study, finishing end of 2023.
  01:39:41
• His SVB diagnosis: management chased returns by buying long-dated instruments funded by short-term variable deposits in a zero-rate world.
  02:05:34
• Altman, who once jailbroke the first iPhone as a kid, compares jailbreaking AI to how piracy gave birth to Spotify.
  01:56:48

Books, products and media the guest or host genuinely endorsed here — with the buy link.

Affiliate link — we may earn a commission at no extra cost to you.