Skip to content
TLexDR
Episodes / George Hotz: Hacking the Simulation & Learning to Drive with...

George Hotz: Hacking the Simulation & Learning to Drive with Neural Nets

05-28-26 ▶ 3h 8m 📖 7 min read
Core Takeaways
Comma AI's Comma 2 device, priced at $1,000, supports 91 car models and aims to reach 100 by year-end.
Why it matters This expansion could significantly increase Comma AI's market share and influence in the autonomous driving sector.
George Hotz argues that Tesla's multitask learning approach to autonomous driving is flawed compared to end-to-end learning. ▶ 1:20:00
Why it matters If correct, this could shift industry focus towards end-to-end learning, impacting future AI development strategies.
MuZero is highlighted as a cornerstone paper for self-driving technology, with full autonomy expected in the late 2020s. ▶ 1:40:00
Why it matters This timeline suggests a realistic horizon for achieving full autonomy, influencing investment and regulatory planning.
Hotz critiques NVIDIA's pricing strategy for high-end chips as unsustainable and potentially harmful long-term. ▶ 2:00:00
Why it matters Such pricing could limit accessibility and drive customers to competitors, impacting NVIDIA's market dominance.
Ethereum's layer two solutions improve scalability by running contracts off-chain, posting only results on-chain. ▶ 2:20:00
Why it matters These solutions enhance Ethereum's capacity, crucial for handling increased transaction volumes and maintaining network efficiency.

Detailed Insights

Autonomous Driving
+
Comma AI's Comma 2 supports 91 car models and aims for 100 by year-end.
Tesla's multitask learning approach is critiqued as flawed.
MuZero is pivotal for self-driving technology, with full autonomy expected by late 2020s.
AI Hardware
+
NVIDIA's pricing strategy for high-end chips like the A100 is criticized.
Google's TPU terms restrict autonomous vehicle training, favoring NVIDIA.
Cryptocurrency and Ethereum
+
Ethereum's layer two solutions improve scalability by running contracts off-chain.
Smart contracts offer a reliable alternative to traditional legal agreements.

How the conversation moved

Lex Fridman began the conversation by framing the central question around the potential for intelligent life in the universe and the implications of civilizations self-destructing. George Hotz introduced the idea that the IQ required to destroy the world decreases annually, suggesting that technology democratizes destruction. This led to a discussion on the Fermi Paradox and the likelihood of intelligent civilizations wireheading themselves into oblivion. Hotz's perspective was that if intelligent civilizations existed, they would have already taken over the universe unless they self-destructed.

Hotz's main argument centered on the flaws in Tesla's approach to autonomous driving, advocating for an end-to-end learning model over Tesla's multitask learning. He supported his position by highlighting the historical trend in AI where feature engineering approaches are eventually replaced by end-to-end methods. Hotz also pointed to MuZero as a cornerstone paper that could significantly impact self-driving technology, projecting full autonomy in the later 2020s. This claim was backed by his belief that end-to-end learning offers a more scalable solution for achieving level five autonomy.

Lex did not challenge Hotz's critique of Tesla's approach, though the obvious counter-position would be that Tesla's vast data collection and real-world testing offer a practical advantage. Hotz also pushed back against the notion that NVIDIA's high pricing strategy is sustainable, arguing that it could limit accessibility and drive customers to competitors. The tension here lay in contrasting business strategies: maximizing short-term profits versus long-term market dominance.

The conversation pivoted to the technical aspects of Ethereum and its layer two solutions, which Hotz believes enhance scalability by running contracts off-chain. Hotz's critique of NVIDIA's pricing strategy remained unresolved, but he emphasized the importance of open competition and innovation in the cryptocurrency space. The discussion concluded with Hotz's vision of a world where smart contracts replace traditional legal systems, highlighting the potential of decentralized technologies to reshape industries.

Surprising moments

George Hotz
George Hotz claimed that the IQ required to destroy the world falls by one point every year, suggesting technology makes destruction more accessible.
Share this quote X Bluesky LinkedIn Email Download card
George Hotz
Hotz argued that Tesla's multitask learning approach to autonomous driving is fundamentally flawed compared to end-to-end learning.
George Hotz
Hotz critiqued NVIDIA's pricing strategy for high-end chips as unsustainable, arguing it could harm their long-term prospects.

Topics Covered

Autonomous Driving AI Hardware Cryptocurrency and Ethereum

Memorable Quotes

"The IQ required to destroy the world falls by one point every year." — George Hotz
"If you can make smart contracts, this whole concept of code is law. I love, and I would love to live in a world where everybody accepted that fact." — George Hotz

Still open

Unresolved by the end of the conversation

  • Hotz questioned whether Tesla's multitask learning approach can achieve level five autonomy, leaving its viability open-ended.
  • The conversation raised doubts about NVIDIA's long-term pricing strategy without resolving its potential impact on market competition.

Jargon glossary

wireheading
A hypothetical scenario where intelligent beings prioritize artificial pleasure over survival.
MuZero
A deep reinforcement learning model that learns a model of the environment to make decisions.
end-to-end learning
An AI approach where the system learns directly from input to output without intermediate steps.
layer two solutions
Scalability solutions that run off-chain computations to improve blockchain efficiency.

References & Resources

Infinite Jest by David Foster Wallace book
Atlas Shrugged by Ayn Rand book
The Metamorphosis of Prime Intellect by Cory Doctorow book
Hutter Prize by Hutter other
Unqualified Reservations by Curtis Yarvin other

For the specialist

What a senior practitioner would find new

  • Hotz modified the Ethereum compiler with a 300 line diff to streamline safe bytecode creation, reducing complexity from a 3000 line transpiler.
  • Hotz argues that lossless compression is equivalent to intelligence, suggesting a novel perspective on AI capabilities.

Ask this episode Deep

A preview of how Deep chat answers, grounded in this episode with citations and timestamps:

Cite this episode

For papers, blog posts, anywhere.

Copied!

Related episodes

Where to go next from this conversation.

AI-generated summary · last refreshed 2026-06-06 22:05:27 · how we make these

Quotes are matched verbatim against the source transcript; references are checked to resolve to real URLs. Even so, AI can misread structure or attribute claims imperfectly. If you spot an error, please let us know.

Report an inaccuracy →