Wojciech Zaremba: OpenAI Codex, GPT-3, Robotics, and the Future of AI | Lex Fridman Podcast #215

Wojciech Zaremba joins Lex Fridman to discuss the science and philosophy behind OpenAI's language models. He frames intelligence and consciousness as forms of compression, defines love as optimizing each other's reward functions, and explains how GPT-3, Codex, and GitHub Copilot turn next-word prediction into powerful tools. He recounts OpenAI's robotics work solving a Rubik's Cube one-handed via simulation, and argues for iterative, gradual deployment of increasingly powerful AI. The conversation ranges widely across meditation, psychedelics, mortality, and how to generate good research ideas.

• Zaremba states his current belief that we may be alone in the universe, which makes conscious life more valuable.
  00:02:08
• He proposes consciousness is intertwined with compression and that self-consciousness may be 'metacompression' - a compressor compressing itself.
  00:23:27
• He defines love as the dissolution of boundaries where two people end up optimizing each other's reward functions.
  00:47:24
• He argues for iterative deployment of AI - releasing slightly better versions for public criticism rather than holding a powerful system back.
  01:24:11
• He says AGI should not be controlled by a small number of people and power should be distributed to a larger collective.
  01:25:48
• OpenAI solved a Rubik's Cube with a single robotic hand by training across a randomized range of simulations rather than one.
  01:45:58
• He says if building a robotics company today he would not dismiss supervised learning - he'd record human teleoperation trajectories first.
  01:52:44
• He admits as a child he synthesized nitroglycerin and detonated homemade dynamite with a friend.
  02:30:31

• Predicting the next word is equivalent to compression, which means learning a model of reality.
  00:21:56
• Marcus Hutter's compression prize treats intelligence as compression of Wikipedia; Hutter advised DeepMind co-founder Shane Legg.
  00:25:30
• Zaremba sketches Solomonoff-style optimal induction: enumerate all programs producing the observed bits and weight them by length.
  00:27:38
• More AI gains have historically come from compute than from algorithms, though you can't invest 10 trillion dollars in compute.
  00:38:30
• At a meditation retreat he describes reaching 'inbox zero' for the mind, after which the default state feels deeply peaceful.
  00:59:53
• GPT-3 goes off the rails on long text because it is trained on human writing, not on its own mistakes, so errors compound.
  01:18:34
• Code is more verifiable than language because programs can be executed and unit-tested automatically.
  01:41:21
• He blocks Wednesday-Friday as out-of-office for focused work, taking meetings only Monday and Tuesday, which boosted his mood.
  02:22:13
• He keeps a voice recorder by his bed and a separate app-free 'meditation phone' to capture ideas without friction.
  02:27:54
• He cites Bezos-style regret minimization: live so your deathbed self won't regret not having tried.
  02:45:34