George Hotz: Hacking the Simulation & Learning to Drive with Neural Nets | Lex Fridman Podcast #132

In his second appearance, George Hotz talks with Lex Fridman about whether we live in a simulation and whether it's worth trying to hack it, plus crypto, smart contracts, and immortality. The bulk of the conversation centers on comma ai's approach to self-driving: an end-to-end neural network trained on user driving data, contrasted with Tesla's multi-task 'data engine' feature-engineering pipeline. Hotz argues end-to-end and ideas like MuZero will ultimately win, defends driver monitoring as easy and essential, and explains comma's consumer-electronics business model. He closes with sharp opinions on NVIDIA's pricing, programming languages everyone should learn, book recommendations, drugs, love, and what counts as madness.

• Hotz dismisses UFO sightings as a psyop and frames Bob Lazar's alien-tech claims as likely MK-Ultra-style experiments to stimulate young physicists.
  00:11:47
• He confirms the legend: he walked into a restaurant, met the Optimism crypto team mid-deadline, and wrote a 300-line compiler diff to solve their Solidity-to-OVM problem.
  00:31:59
• Hotz bets Tesla is wrong and end-to-end learning is the only viable path to Level 5, calling task-by-task approaches 'feature engineering' that always loses.
  01:06:08
• He declares MuZero the cornerstone paper of the deep learning era and effectively 'the solution to self-driving cars'.
  01:10:19
• He bets $10,000 that Elon will ship driver monitoring before he ships Level 5.
  01:16:58
• He calls out NVIDIA for price-gouging A100s at $10K when they're 'not that different' from a $699 3080, predicting it opens the door to competitors.
  01:29:35
• Asked how he'd run Waymo, he says he'd get Anthony Levandowski out of jail and put him in charge, calling him unironically a genius.
  02:00:41
• He says he laughs in the faces of acquisition offers, calling M&A people demeaning and the money 'fedcoin'.
  01:51:45

• Hotz says a von Neumann probe plus a million years of sublight travel would let a civilization take over the whole galaxy, which is why he thinks intelligent civilizations blew themselves up.
  00:05:39
• He compares lawyers to Python: lawyers cost $1,000/hour and aren't deterministic, while Python is ~10,000x cheaper and reliable, making smart contracts compelling.
  00:41:16
• openpilot went from an unplanned disengagement every 10 miles to every 100 miles in about a year, a 10x improvement.
  01:01:59
• 70% of comma two buyers are daily active users, an unusually high retention rate.
  01:46:07
• Hotz can recall only three transformational days in his life: discovering Yudkowsky, the Hutter Prize (AI as compression), and the blog Unqualified Reservations.
  02:27:34
• He defines intelligence as lossless compression and notes a lossy compressor can be made lossless via an arithmetic encoder.
  02:32:46
• He calls GPT-3 overhyped, arguing cross-entropy loss on characters isn't the loss function of general intelligence and that GPTs lack long-term memory.
  02:34:50
• His recommended programming languages span four paradigms: assembly, C, Python (imperative), Haskell (functional), Verilog (hardware), and PyTorch (ML / 'programming 2.0').
  02:42:37
• He argues driving is more Markovian than language, so transformers may help but won't be a game changer for autonomous driving.
  02:36:57
• He criticizes San Francisco as a corrupt one-party city that votes against building houses, recommending San Diego instead.
  02:08:57