notes from the latest george hotz podcast

these are transcriptions and notes about interesting points in the podcast that i can share to friends to discuss about them

#artificial_intelligence

this one

memes #

• superhuman ai can generate super scary memes
• "infinite jest" book, about a tape once you watch it once you only ever want to watch that tape, to the point of doing self harm
  • we're going to build that, i think, but it wont be just a static tape, maybe something more dynamic.
  • the tiktok you can't look away from
  • imagine tiktok but for porn
  • "i dont even know what it will look like"
  • it's not the machine that's gonna do it. it's going to be humans using the machine to do this to you
• eliezer yud
  • "agree?" "yes, but maybe for a different reason"
  • nuclear weapons didn't kill everyone
    • it's hard to accomplish tactical objectives with them
    • if you nuke a place you just obliterate it, lose all resources on it
    • winning a war is total subjugation and domination of the people in it
  • the "red button" of nuclear weapons wasn't pressed
  • but the ai one will
    • it's not because the ai will kill us, but a human will use ai to kill us
  • idea where shit happens
    • most obvious idea is "wireheading", staring at that infinite tiktok and forgetting to eat
    • even more bedign, we just stop reproducing
    • to be fair, it's hard to get all of humanity. there's a lot of weirdos out there
      • "i think the diverity in the world is decreasing" -gh
      • "the interconnectedness is what's doing it" -gh
      • "there's still going to be a guy in texas" -lex
      • "i think ai kills everything that represents society today" -gh
      • "but restarting society is tricky. most people don't know how to do most things" -lex
      • "some of us do, and they'll be okay and rebuild" -gh
    • a new society will be religiously against electricity and power
    • "some kind of amish looking thing, taboos against technology" -gh
  • "we haven't built a machine capable of reproducing"
    • you need a fab to produce chips
    • fabs are big
    • life will keep going in the absolute absecnse of civilization
    • if an ai kills us, they will die too
    • its hard to self-replicate, different life stack (biology vs technology)

tinycorp #

• origin story
  • started as a toy project, akin to karpathy's micrograd
  • and then i started thinking about ai chips
  • "if nvidia becomes a monopoly, how long until nvidia is nationalized?" -gh
    • "it's not to challenge nvidia, i like nvidia" -gh
    • "it's to make sure power stays decentralized" -gh
      • "computational power" -lex
    • "if nvidia gets 10x better than everyone else, you give a big advantage to somebody who can secure nvidia as a rersource" -gh
    • "in fact if Jensen watches this podcast you may want to consider this" -gh
    • "you may want to consider making sure his company is not nationalized" -gh
    • "do you think thats an actual threat" -lex
      • yes -gh
    • "we have amd" -lex
      • "so we have nvidia and amd. great" -gh
    • "what about google" -lex
      • google does not sell tpus
      • they love renting them
  • there's three kinds of computation paradigms
    • cpu
      • cpu can do everything
        • add, multiply
        • load, store
        • compare, branch
    • gpus
      • simpler
        • no compare and branch (absurdly slow)
        • can do load, store
    • dsp
      • just add and multiply
      • only static load and store
    • 95% neural networks are dsp
      • statically scheduled adds and multiplies
    • tinygrad extends this idea
  • whole stack is turing complete
    • llvm, cuda, python, etc
    • "I want to get turn completeness out of the stack entirely" -gh
    • "because once you get rid of turn completeness you can reason about things. Rice's theorem and the halting problem do not apply to Admiral machines" -gh
    • "the reason you need to do Branch prediction in a CPU" -gh
    • "and the reason it's prediction and the branch predictors are I think they're like 99 on CPUs why do they get one" -gh
    • "percent of them wrong? well they get one percent wrong because you can't know right? that's the halting problem. it's'" -gh
    • "equivalent to the halting problem to say whether a branch is going to be taken or not um I can show that but"
    • "the admal machine the neural network runs the identical compute every time the only thing that changes is the data"
    • "so when you realize this you think about okay how can we build a computer and how can we build a stack that takes maximal advantage of this idea"
  • about 20 primitive operations
    • compare to XLA or PrimTorch
    • torch has 2000 kernels
    • PrimTorch (in torch 2.0) has around 250
    • it's about risc vs cisc
    • tinygrad is risc
  • it's really easy to see the kernels being sent to the gpu
    • torch doesnt give you that, no flop usage, how many memory acesses
    • tinygrad? debug=2, every kernel that's run is then shown
  • tinygrad solves the problem of porting new ml accelerators quickly
    • sequoia, tenstorrent, graphcore, cerberus
    • all of these build chips, the chip is good, but software is terrible
    • its hard to wite a pytorch port, you need 250 kernels, and tune them all for performance
  • frist: write a performant nvidia stack. target amd tho lol!
  • amd sucks for ai
    • kernel drivers dont work. demo apps in loops panics the kernel
    • they're trying to change
    • intel gpus have better drivers than amdgpu
  • spearhead the diversification of gpus
    • "and then to everything" -gh
  • gh hopes meta ships their internal accelerator chip
• tinybox
  • specs
    • 738 FP16 TFLOPS
    • 144 GB GPU RAM
    • 5.76 TB/s RAM bandwidth
    • 30 GB/s model load bandwidth (big llama loads in around 4 seconds)
    • AMD EPYC CPU
    • 1600W (one 120V outlet)
    • Runs 65B FP16 LLaMA out of the box (using tinygrad, subject to software development risks)
    • $15,000
    • preorder $100
  • "the best deep learning box that you can plug into a wall outlet" -gh
  • "im leaning towards amd gpus" -gh
    • 7900 XTXs, maybe 3090s, maybe A770s
    • "still exploring" -gh
  • "we're also going to sell the tiny rack which like what's the most power you can get into your house without arousing suspicion uh and one of the answers is an electric car charger" -gh
    • "a wall outlet is 1.5kW, a car charger is 10kW"
  • "experience a chat with the largest language model that you can have in your house yeah from a wall plug" -gh
  • cooling
    • "i want the cooling to be good and quiet" -gh
    • "you want it to coexist with you if it's screaming it's 60 DB you don't want that in your house you'll kick it out 60 DB" -gh
    • the trick will be to make it big, but not as big as UPS Large boxes, slower fans lead to less noise
  • no assembly on location, you get the box, you run the box
  • ubuntu, not arch
  • tinygrad builtin

using llms for programming (and then killing google) #

• lots of stuff skipped here
• "it's actually weird the way that we do uh language models right now"
• "where all of the uh information is in the weights and human brains don''t really like this it's like a hippocampus and a memory system so why don't llms have a memory system"
• "and there's people working on them I think future llms are going to be like smaller but are going to run looping on themselves and are going to have retrieval systems and the thing about using a retrieval system is you can cite sources explicitly"
• "that's gonna kill google"
  • "when someone makes an llm that's capable of citing its sources it will kill Google." -gh
  • "LLM that's citing a sources because that's basically a search engine" -lex
  • "I would put a lot more money on Mark Zuckerberg what is that because Mark Zuckerberg's alive" -gh
  • "there's this old Paul Graham essay: startups are either alive or dead. Google's dead" -gh
  • lex: "and facebook's alive, meta"
  • gh: "meta!"
  • gh: "this is Mark Zuckerberg reading that Paul Graham essay and being like I'm gonna show everyone how alive we are I'm gonna change the name"
  • lex: google doesn't have a pivoting engine
  • gh: "when I listen to Sam Altman podcast um he talked about the button everyone who talks about ai talks about the button. the button to turn it off"
  • gh: "do we have a button to turn off Google? is anybody in the world capable of shutting Google down?"
  • gh: "does Sundar prashai have the authority to turn off google.com tomorrow?"
  • gh: "they have the technical power but do they have the authority?"
  • gh: "let's say Sundar pashai made this his soul Mission came into Google tomorrow and said I'm going to shut google.com down. I don't think you keep this position too long"
    • revenue zero, shareholders, etc
  • lex: "so the the capitalist machine prevents you from having the button"
  • gh: "yeah, and this is true for the AIs too"
  • gh: "now, does Mark Zuckerberg have that button for facebook.com? i think he does"
  • gh: "this is exactly why i bet on him so much more than google"
  • lex: "the positive side of the button is innovating aggressively"
  • gh: "i would bet on something [..] that looks like midjourney but for search"
  • lex: "I mean it just feels like one model can take off"