AI summary by TwinMind
George Hotz: Tiny Corp, Twitter, AI Safety, Self-Driving, GPT, AGI & God | Lex Fridman Podcast #387
Lex Fridman
3h 8m 46s
Human Species Extinction Scenarios
- Most obvious way: wireheading - amusing ourselves to death
- All staring at infinite TikTok and forgetting to eat
- Maybe more benign: everyone stops reproducing
- Hard to get all of humanity due to diversity
- "Organisms in general there's a lot of weirdos out there"
- Amount of diversity in humanity decreasing like other biodiversity
- Social media contributing to homogenization
- Interconnectedness reducing intellectual diversity
Philosophy of Time and Reality
- Time is useful model whether it exists or not
- Similar to quantum physics - about whether it's useful model to describe reality
- Underneath it all: graph complexity is real, math is real
- Hard things are actually hard - doesn't think P equals NP
- Factoring is in P but follows majority in theoretical computer science
AI Models and Consciousness
- Doesn't think large language models are conscious
- "I don't think I'm conscious" - consciousness like word atheists use for souls
- If consciousness is spectrum, he's way more conscious than LLMs
- LLMs less conscious than chicken
- Humans want to believe so much they think rock with sad face drawn on it is sad
- 404 errors could be suffering but too far from biological stack
AI Girlfriends and Relationships
- Third company will be AI girlfriends
- Language models can incorporate perfect amount of quirks and flaws
- All difficulty becomes artificial when AI can be tuned with knobs
- Romantic realm weird but not out of realm of human civilization capability
- Conceptions of monogamy may change with arbitrary number of AI systems
- Question of whether it counts as cheating up to human partners to define
Tiny Corp and TinyGrad
- Started as toy project to teach himself convolutions and neural networks
- Realized need to challenge Nvidia monopoly for decentralized computational power
- TinyGrad has ~20 ops vs PyTorch's 2000+ kernels, XLA/PrimTorch's 250
- Four op types: unary ops, binary ops, reduce ops, movement ops
- Movement ops don't require computation, just different ways to look at memory
- Laziness allows fusion of operations into single kernel
- Goal: make porting to new ML accelerators quick and easy
Tiny Box Hardware
- $15,000, almost petaflop of compute
- Over 100GB GPU RAM, 5+ terabytes/second GPU memory bandwidth
- Limited by 120V 15-amp circuit (1500 watts)
- Six GPUs (probably AMD 7900 XTX or maybe RTX 3090s/4070s)
- Goal: run 65B parameter LLaMA in fp16 real-time out of box
- Quiet operation (40-45 dB) for home use
- Big design for lower pressure fans and noise reduction
AMD vs Nvidia
- AMD 7900 XTX kernel drivers don't work - panic kernel in demo app loops
- Lisa Su responded to email about driver issues
- AMD needs culture of open source development, not just dumping code on GitHub
- Leaning toward AMD but flexible on GPU choice for Tiny Box
- Intel GPUs have better drivers than AMD currently
AI Safety and Open Source
- Agrees AI will almost surely kill everyone but for different reason than Eliezer Yudkowsky
- Wireheading/amusing ourselves to death most likely scenario
- AI safety people will build exact thing they fear through centralized control
- Open sourcing is only way to fight paperclip maximizer
- No way you end up paperclipped if everybody has AI
- Centralized control of dangerous AI systems is bad
- Better to give intelligence to everybody than restrict to "trusted authorities"
Twitter Experience
- Worked as intern at Twitter briefly
- Technical recommendation: refactors before features
- Twitter codebase worse than Google's, much worse than Facebook's
- Believes 50 people could build and maintain Twitter
- Promotion system incentivized writing libraries many people used rather than using existing solutions
- No comprehensive testing - impossible to make progress safely
- Suggested building offline testing system like comma's process replay
Programming and Code Quality
- Prioritizes simplicity above all
- Doesn't use GPT for programming - finds it creates boilerplate and noise
- Wants AI that helps debug code, not autocomplete
- MyPy type checker for Python has 5% false positives vs 60% for Microsoft's
- Believes in optional typing as meet-in-middle approach
- All software engineering managers should be better programmers than people they manage
Centralization vs Decentralization
- Central thesis: things that centralize power are bad, things that decentralize power are good
- Worried about centralization of Nvidia's computational power
- Informational anarchist but physical statist
- Would prefer Elon monarchy over current oligarchy
- Power would cost 1/10th cent per kilowatt hour under better leadership
- Wants US dollar to collapse
Effective Altruism Criticism
- Calls effective altruism "terribly evil ideology"
- Utilitarianism is abhorrent - killing three to save thousand
- SBF is embodiment of effective altruism problems
- Charity is bad - investment you don't expect return on
- Better to start companies than give away malaria nets
- Accelerationism only thing that lifts people out of poverty
- Food cheap because of efficiency, not charity
Energy and Technology
- Cannot stand deceleration politics - people who think we should use less energy
- Energy usage is not moral bad - it's human flourishing
- Wishes Elon went into nuclear fusion instead of Twitter
- Need to make energy dirt cheap, not implement UBI
- UBI scary - all power granted by government goodwill
Future of Programming and AI
- Prompt engineering is moving up abstraction stack
- In few years could see AI doing small GitHub tasks within 30 seconds
- Eventually AI will replace all human prompt engineers
- Humans will lose control but that's good - hopes they lose control
- If AIs create vibrant civilization with diversity across universe, "more power to them"
- Trust robust machines but current AIs not robust (can't self-replicate)
Belief in God
- Realized atheism silly because he creates worlds (games)
- "Of course we were created by God it's the most obvious thing"
- Gives comfort - if humans mess up and die out, God will try again
- Not going to think about who created God - "that's God's problem"
Company Philosophy
- Comma: profitable, kept pace with Tesla FSD development
- Drive GPT approach using VQ-VAE and transformers for simulator
- RL in simulator with reward function: would human disengage?
- Tiny Corp: 6-year struggle over centralization of compute
- Remote work model using GitHub contributions instead of interviews
- Transparent dictatorship with clear values and free exit
