# Commoditizing the Petaflop — with George Hotz of the tiny corp

- **Show:** Latent Space: The AI Engineer Podcast
- **Published:** 20 Jun 2023
- **Source:** https://lspace.beehiiv.com/podcast/s/latent_space_the_ai_engineer_podcast/commoditizing_the_petaflop_with_george_hotz_of_the_tiny_corp
- **Transcript:** https://podcasts.opennash.com/shows/latent-space/commoditizing-the-petaflop-with-george-hotz-of-the-tiny-corp-3189c2/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

George Hotz argues that AI compute should be cheap, open and accessible to everyone, and that this is worth building a company around. His company, the tiny corp, makes tinygrad, a deliberately small ML framework. It also sells 'tiny box' computers. His technical bet is that neural networks do the same fixed set of operations every run, so simple, statically scheduled hardware and software can beat complex general-purpose designs.

The central tension is between open, simple tools and closed, complex ones. Hotz is blunt about the gap: tinygrad is still about 5x slower than PyTorch on NVIDIA, and AMD's drivers crashed his machine. He is more upbeat about the future of work (tools that supercharge people), AI hype (OpenAI's secrecy hides unremarkable tricks), and his long-term goal of merging with machines.

## Key insights

1. **Neural nets are predictable, so simpler chips suffice** Hotz says a neural net runs the identical loads, stores and arithmetic every time, with only the data changing. That means compilers can schedule everything in advance, so hardware can drop branch predictors, warp schedulers and complex caches.
2. **Chip makers must first master software on NVIDIA** Hotz argues a new chip will be harder to use than an AMD GPU, so if you cannot write a fast framework for existing GPUs you cannot for your own chip. He cites Google's TPU as the exception, because Google wrote its ML framework first.
3. **AMD's driver problems, not speed, nearly ended his plans** He says AMD demo apps crashed his entire computer with kernel panics, which he called an unacceptable basis for building. Lisa Su replied to his email and AMD sent a pre-release ROCm 5.6 that fixed it, but he sees AMD as lacking open-source culture.
4. **Tiny box is hard to build, not just assembly** Hotz says putting six GPUs in one quiet machine that runs from one wall outlet requires cooling, PCIe redrivers (PCIe 4.0 extenders failed in his tests) and power limiting. He says limiting GPU power can keep about 80% of performance at half the power.
5. **He bets on small models trained far longer** Hotz thinks the best chatbot models will be smaller ones trained many times and fine-tuned, not ever-bigger ones, because inference cost matters more. He claims GPT-4 is just an eight-way mixture of experts, which he calls 'what you do when you're out of ideas'.

## Questions this episode answers

### What is tinygrad and why does George Hotz say it is different from PyTorch?

Tinygrad is a small machine-learning framework from Hotz's company, the tiny corp. It breaks every model into a small set of basic operations (about 25 versus about 250 in XLA or PrimTorch) and lazily fuses them so memory round trips are avoided.

### What does Hotz mean by 'commoditizing the petaflop'?

A petaflop is a measure of computing speed, and Hotz wants to sell compute at the lowest cost per flop and per watt so anyone can afford it. He says his company's business goal is to sell computers for more than they cost to make.

### Why did Hotz criticize AMD, and did that change?

He said AMD's GPUs and demo apps crashed his entire computer with kernel panics, which made them not worth building on. After he emailed CEO Lisa Su, AMD responded and sent a pre-release driver that fixed the crash, though he still criticizes its open-source habits.

### What is a tiny box and who is it for?

It is the tiny corp's first product, a quiet multi-GPU computer meant to sit under a desk and run large models like Llama. Hotz says it suits individuals and companies, and for most model-training jobs it could be 5x faster than an H100 box at the same price, though not when huge interconnect bandwidth is needed.

### What does Hotz think about AI replacing jobs and AI risk?

He says coding is 'above the API line', so it will be supercharged by tools rather than replaced for now, though he expects machines to replace most things in about 20 years. He thinks AI could kill many people, but because companies and governments aren't aligned with you, not because of paperclip-style machines.

## Facts

- **Alessio Fanelli** [0:00:20]: "So you traded the first ever unlocked iPhone for a Nissan 350Z and three new iPhones."
- **George Hotz** [0:03:33]: "And yeah, it can run all ML models with basically 25 instead of the 250 of XLA or PrimTorch."
- **George Hotz** [0:04:26]: "So MidJourney is trained on TPU, right? Like a lot of startups do actually train on TPUs."
- **George Hotz** [0:08:16]: "Think of a grid of mullax, and then the grid can multiply and then shift, multiply, then shift, multiply, then shift."
- **George Hotz** [0:09:40]: "a 32 megabyte binary blob called libTPU on Google's cloud instances. It's all closed source."
- **George Hotz** [0:10:10]: "For a long time, TinyGrad had a hard limit at a thousand lines of code."
- **George Hotz** [0:11:07]: "We have better support for Onyx than Core ML does."
- **George Hotz** [0:18:38]: "But on NVIDIA, it's about 5x slower than PyTorch right now."
- **George Hotz** [0:19:03]: "And TinyGrad is about 2x faster on the GPU than Qualcomm's library."
- **George Hotz** [0:21:37]: "Intel GPUs have a stable kernel driver and they have all their hardware documented."
- **George Hotz** [0:22:43]: "But no, Lisa Sue reached out, connected with a whole bunch of different people. They sent me a pre-release version of Rock M 5.6."
- **George Hotz** [0:36:32]: "NV-linked A100s are going to have, and then you can even get like full fabric and NVIDIA really pushes on that stuff, 600 gigabytes per second."
- **George Hotz** [0:36:32]: "If you're training on a tiny box, your limit's going to be about 7 billion."
- **George Hotz** [0:40:41]: "The tiny box has 60 gigabytes per second. And then your internet has 125 megabytes per second."
- **George Hotz** [0:41:58]: "A person is 20 petaflops. It's about 30 petaflops. We built out a little compute cluster. And we paid double what you theoretically could per flop."
- **George Hotz** [0:42:46]: "Well, it's a lot cheaper now. So like I said, Comma spent about half a million on our person and a half."
- **George Hotz** [0:43:20]: "Well, okay, so GPT-4 is 220 billion in each head, and then it's an eight-way mixture model."
- **George Hotz** [0:46:28]: "I got into AI through the Hutter Prize. 2014, I'm trying to build compressive models of Wikipedia."
- **George Hotz** [0:48:36]: "I changed them to be log2 and exp2 because hardware has log2 and exp2 accelerators."
- **George Hotz** [0:55:58]: "Except, of course, we made the comma version of it. Tesla uses 20 actuators. We use two."

## Transcript

**[0:00:09] swyx:** Hey everyone, welcome to the Latent Space podcast. This is swyx, writer and editor of Latent Space, and Alessio is taking over with the intros. Alessio's partner, NCTO-in-Residence, and Decibel Partners.

**[0:00:20] Alessio Fanelli:** Hey everyone, today we have Geohot on the podcast, aka George Hotz for the human name. Everybody knows George, so I'm not going to do a big intro. a couple of things that people might have missed. So you traded the first ever unlocked iPhone for a Nissan 350Z and three new iPhones. You were then one of the first people to break into the PS3 around arbitrary code. You got sued by Sony. You wrote a rap song to fight against that, which is still live on YouTube, which we're going to have on the show notes.

**[0:00:47] Alessio Fanelli:** Did not go to Tesla to build vision. And instead you started Com.i .i, which was an amazing engineering feat in itself until you got a seasoned disease from the government to not put these things on the street. turned that into a research-only project. You know they're out there. Yeah, yeah. No, no, no. They're out there. But you market them as a research kind of like no warranty.

**[0:01:06] George Hotz:** Because I use the word dev kit, that's not about the government. That has nothing to do with the government. We offer a great one-year warranty. The truth about that is it's gatekeeping. What's the difference between a dev kit and not a dev kit? Nothing. Just the question of do you think it's for you? And if you think it's for you, buy it. It's a consumer product. We call it a dev kit. If you have a problem with that, it's not for you.

**[0:01:28] Alessio Fanelli:** That's great insight. I was going through your blog post to get to the day. You've heard this post about the hero's journey. And you linked this thing called the portal story, which is kind of the set of stories in movies and books about people living this arbitrary life. And then they run to this magic portals, kind of takes them into a new, very exciting life and dimension. When you've heard that post, you talked about Tiny Grad, which is one of the projects you're working on today.

**[0:01:53] Alessio Fanelli:** And you mentioned this is more of a hobby, something that is not going to change the course of history. Obviously, you're now going full speed into it. So we would love to learn more about what was the portal that you run into to get here.

**[0:02:04] George Hotz:** Well, what you realize is, you know what made me realize that I absolutely had to do the company? Seeing Sam Altman go in front of Congress. Why? What are the odds they nationalize NVIDIA? You know, what are the odds that large organizations in the government, but of course I repeat myself, decide to try to clamp down on accessibility of ML compute? I want to make sure that can't happen structurally. So that's why I realized that it's really important that I do this.

**[0:02:29] George Hotz:** And actually from a more practical perspective, I'm working with NVIDIA and Qualcomm to buy chips. NVIDIA has the best training chips. Qualcomm has the best inference chips. Working with these companies is really difficult. So I'd like to start another organization that eventually in the limit either works with people to make chips or makes chips itself and makes them available to anybody.

**[0:02:48] Alessio Fanelli:** You share kind of three core thesis to TinyCorp. Maybe we can dive into each of them. So XLA, PrimeTorch, those are the complex instruction system. TinyGrad is the restricted instruction system. So you're kind of focused on, again, TinyGrad being small, not being overcomplicated and trying to get as close to like the DSP as possible in a way where it's at more.

**[0:03:08] George Hotz:** Well, it's a very clear analogy from how processors developed. So a lot of processors back in the day were CISC, complex instruction set, System 360, and then x86. Then this isn't how things stayed. They went to now the most common processor is ARM, and people are excited about RISC-V. No one's excited about RISC-V. RISC-V is even less complex than ARM. No one is excited about CISC processors anymore. They're excited about reduced instruction set processors.

**[0:03:33] George Hotz:** processors. So TinyGrad is we're going to make a risk offset for all ML models. And yeah, it can run all ML models with basically 25 instead of the 250 of XLA or PrimTorch. So about 10x less

**[0:03:46] Alessio Fanelli:** complex. Yeah. You talk a lot about existing AI chips. You said if you can write a fast ML framework for GPUs, you just cannot write one for your own chip. So that's another one of your core insights. I don't know if you want to expand on that. Yeah. I mean, your chip is worse, right?

**[0:04:00] George Hotz:** There's no way the chip that you're going to tape out, especially on the first try, is going to be easier to use than an AMD GPU. And yet there's no good stack for AMD GPUs. So why do you think you can make one for your chip? You can't. There's one other company, aside from NVIDIA, who's succeeded at all at making training chips. What company? AMD? Intel? No, no, no. Who's trained to model on AMD or Intel?

**[0:04:26] George Hotz:** Cerebris. Cerebrus, I'm talking about, you might know some startups who trained models on these chips. Oh, TPU. Exactly, yes, right? So MidJourney is trained on TPU, right? Like a lot of startups do actually train on TPUs. And they're the only other successful training chip aside from NVIDIA. But what's unique about Google is that they also wrote their own ML framework, right? And if you can't write your own ML framework that is performant on NVIDIA, there's no way you're going to make it performant on your...

**[0:04:53] Alessio Fanelli:** When they started from TensorFlow and then they made the chip after.

**[0:04:56] George Hotz:** Yeah, exactly. Exactly. And you have to do it in that direction. Otherwise, you're going to end up, you know, a Cerebus? What are those things? A million? I've never seen a Cerebus. No one's ever like, oh, I trained my model on a Cerebus. Most people are like, I trained my model on GPUs. Some people, 20 % are like, I trained my model on TPUs.

**[0:05:14] Alessio Fanelli:** And then the third one, which is the one that surprised me the most, is true incompleteness is harmful. It should be avoided. It made sense once I read it, but maybe tell us a bit more about how you got there.

**[0:05:25] George Hotz:** Okay, so CPUs devote tons of their silicon and power to things like reorder buffers and speculative execution and branch predictors. And the reason that you need all these things is because at compile time, you can't understand how the code's going to run. This is Rice's theorem. This is the halting problem and its limit. And this is not like, oh, the halting problem is theoretical. No No. It's actually very real. Does this branch get taken or not?

**[0:05:51] George Hotz:** Well, it depends on X. Where does X come from? Forget it, right? But no branches depend on X in a neural net. Every branch is a static loop. Like if you're doing a matrix multiply, it's a static loop over the inner dimension. And neural networks are even better. No loads even depend on X, right? So with a GPU shader, right? You're like, your load might depend on which texture you're actually loading into RAM. But with a neural network, your load is, well, I load that weight.

**[0:06:15] George Hotz:** Why? Well, because I load that weight the other million times I ran the same net. Every single time you run the net, you do the exact same set of loads, stores, and arithmetic, the only thing that changes is the data. And this gives you a very powerful ability to optimize that you can't do with CPU-style things, which have branches, and even GPU-style things, which have loads and stores. Well, GPUs, if you want GPU-style stuff, you have load based on X, you now need a cache hierarchy.

**[0:06:40] George Hotz:** And not an explicit cache hierarchy, an implicit cache hierarchy with eviction policies that are hard-coded into the CPU. You start doing all this stuff and you're never going to get theoretically good performance. Again, I don't think there's 100x. Some startups will talk about 100x and they'll talk about absolutely ridiculous things like clockless computing or analog computing. Okay, here, analog computing just won't work. And clockless computing, sure, it might work in theory, but your EDA tools are, maybe AIs will be able to design clockless chips, but not humans.

**[0:07:11] George Hotz:** But what actually is practical is changing cache hierarchies and removing branch predictors and removing warp schedulers, right? GPUs spend tons of power on warp scheduling because we have to hide the latency from the memory. We don't have to hide the latency if everything's statically scheduled.

**[0:07:24] Alessio Fanelli:** Yeah. Why do you think people are still hanging on to Turing complete?

**[0:07:27] George Hotz:** Well, because it's really easy. Turing completeness is really easy to just, ooh, you know, it would just be so nice if I could do like an if statement here and actually branch the code, right? So it requires a lot more thought to do it without Turing completeness. And would this be qualitatively different than TPUs? So TPUs are a lot closer. TPUs are a lot closer to what I'm talking about than like CUDA. Okay, so what is CUDA?

**[0:07:51] George Hotz:** Well, CUDA is a C-like language, which compiles to an LLVM-like IR, which compiles to PTX, which compiles to SAS, which are all turn-deplete. TPUs are much more like this, yeah. Their memory is pretty statically managed. I did some reverse engineering on the TPU. It's published in TinyGrad. It has like a VLIW instruction, and it runs them. So it's similar. I think the TPUs have a few problems. I think systolic arrays are the wrong choice.

**[0:08:16] George Hotz:** I think they have systolic arrays because that was the guy's PhD. And of course, Amazon makes... Could you summarize systolic arrays for those? Systolic arrays are just... Okay, so basically you have like... It's a way to do matrix multiplication. Think of a grid of mullax, and then the grid can multiply and then shift, multiply, then shift, multiply, then shift. And they are very power efficient, but it becomes hard to schedule a lot of stuff on them if you're not doing like perfectly sized dense matrix multiplies.

**[0:08:42] George Hotz:** which you can argue will design your models to use perfectly sized dense mazorick multiplies. Sure.

**[0:08:47] swyx:** Thanks for indulging on these explanations. I think we need to keep our audience along with us by pausing every now and then to explain key terms.

**[0:08:55] George Hotz:** When I say explain a systolic array, I just immediately get a picture in my head of like tilting a matrix and shifting it. It's like hard to kind of explain. Yeah. We'll do something. We'll do something. We'll do something.

**[0:09:06] swyx:** And we edit it in visuals.

**[0:09:07] George Hotz:** Yeah, yeah, yeah. There's some great graphics that just show you, oh, so that's what a systolic array is. but it's a mullack shift machine that looks kind of different from the typical like alu sort of machine i think the right answer is something that looks more like queues that feed into alus and then you can like prefetch the loads from the memory put in a bunch of queues and then the queues just like and feeds into another queue over here but yeah but that's not even the main problem with the main problem with tpus is that they're closed source not only is the chip closed source but all of XLA is open source, but the XLA to TPU compiler is

**[0:09:40] George Hotz:** a 32 megabyte binary blob called libTPU on Google's cloud instances. It's all closed source. It's all hidden stuff. And, you know, well, there's a reason Google made a closed source. Amazon made a clone of the TPU. It's called Inferencia, or they have some other name for it, a training. Trainium, yeah, yeah, yeah. And you look, it's a clone of the TPU.

**[0:09:56] Alessio Fanelli:** Like Google software, at least it kind of works. So those are kind of like the three core pieces. The first thing you're working on that you've been working on is TinyGrid. And one of your Twitch streams is that just the best thing you've ever written? Yeah, tell us a bit more about that creation.

**[0:10:10] George Hotz:** For a long time, TinyGrad had a hard limit at a thousand lines of code. And what this would force you to do is really make sure you were not wasting lines. I got rid of the restriction because it became a little code golfy at the end. But once like the core framework of TinyGrad was there in those thousand lines, but like the core framework, the ideas are expressed with no boilerplate. If you go read Piedtorch, You know, PyTorch, I think it's actually pretty good code.

**[0:10:36] George Hotz:** I think Facebook's pretty good. But there's so much boilerplate. Go in PyTorch and try to track down how an LU actually works. Just a lot of distractions. Oh, you're going to be diving down a long stack from Python to C to custom libraries to dispatchers to... And then I don't even know how to read TensorFlow. Like, I don't even know where's an LU in TensorFlow. Nobody knows. Someone at Google knows, maybe. Google as an organism knows.

**[0:10:59] George Hotz:** I don't know if anyone individual at Google knows.

**[0:11:01] Alessio Fanelli:** What are the important ergonomics for a developer as you think about designing the TinyGrad API?

**[0:11:07] George Hotz:** So the TinyGrad front-end looks very similar to PyTorch. There's an even higher level front-end you can use for TinyGrad, which is just Onyx. We have better support for Onyx than Core ML does. And we're going to have, I think we're going to pass Onyx runtime soon too. People think Onyx runtime, that's a gold standard for Onyx. No, you can do better. Pass them in what specifically? Test, compliance tests. So Onyx has a big set of compliance tests that you can check out.

**[0:11:28] George Hotz:** and we have them running in TinyGrad and there's some failures. We're below Onyx runtime, but we're beyond CoreML. So that's where we are in Onyx support now. But we will pass Onyx runtime soon because it becomes very easy to add ops because you don't need to do anything at the lower levels. You just do it at this very high level and TinyGrad compiles it to something that's fast using these minimal ops. You can write most concretely what TinyGrad can do that PyTorch can't really do is if you have something like A times B plus C, right?

**[0:11:57] George Hotz:** If you write that in naive PyTorch, What it's going to do on the GPU is, well, read A, read B in a kernel, and then store A times B in memory, and then launch another kernel to do A times B plus C. Okay, got to do those loads for memory. And now I did a whole extra round trip to memory that I just didn't have to do. And you're like, yeah, but you can use the Torch JIT and it corrects this.

**[0:12:17] George Hotz:** Yeah, for that one example, for that one example of Molac, but, oh, now you did three multiplies, six multiplies, it won't compile arbitrary code. And if you looked into like the other

**[0:12:28] swyx:** approaches like PyTorch Lightning to accelerate PyTorch itself? Well, PyTorch Lightning, my

**[0:12:33] George Hotz:** understanding is it's mostly a framework around PyTorch, right? PyTorch Lightning is not going to fix this fundamental problem of I multiply six tensors together, why is it going to memory any more than a single read from each and a single write to the output? There are lower level things in PyTorch that are, I'm not exactly sure what Dynamo does, but I know they're generating some Triton stuff, which is going to generate the kernels on the fly.

**[0:12:54] George Hotz:** But you know, PyTorch Lightning is at as at a higher level of abstraction. So TinyGrid's front-end stuff looks like PyTorch. I made a few tweaks. There's a few things I don't like about PyTorch. Why is ReLU a class? Really, where's the state? You make a class when there's a state. Everything should just be Torch functional and then ReLU, but just.ReLU on the tensor. Also, there's things in Torch where you have to do tensor dot and not a tensor dot, right?

**[0:13:17] George Hotz:** It just shows an API that's not perfectly refined. But when you're doing stuff TinyGrid style where you don't have lines, well, it has to work this way. Because even the lines to express the, well, you can't use the where operator unless in the where operator in PyTorch. Why is it true case, condition, false case? Ugh, the worst. That's like how Python expresses ifs. It's disgusting, right? Turner operators are much nicer. It should be, I can do my like a less than zero dot where a comma one, right?

**[0:13:46] George Hotz:** The very pandas like API. Yeah Yeah. It looks like Torch, NumPy, pandas. They're all very similar. I tried to take like the cleanest subset of them and express them. But like I said, you can also interact with it using Onyx. I have a rewrite of Stable Diffusion. I have a rewrite of Llama. I have a rewrite of Whisper. You can look at them. They're shorter than the Torch versions and I think they're cleaner.

**[0:14:05] George Hotz:** And you stream them all.

**[0:14:05] Alessio Fanelli:** Yeah, very nice. Laziness is kind of the other important concept that you're leveraging to do operation fusing.

**[0:14:11] George Hotz:** Yeah, you have basically like a few different like models for the simplest one is Eager. As soon as the interpreter sees A times B, it actually dispatches A times B, right? Then you have graph, like TensorFlow, which will put A times B into a graph, and then we'll do absolutely nothing until you actually compile the graph at the end. I like this third choice, which is somewhere in the middle, laziness. Laziness is you don't know when the ops are going to dispatch, and don't worry about that.

**[0:14:41] George Hotz:** You don't have to worry about this as a programmer. You just write out all your stuff, and then when you actually type.numpy, it'll be ready by the time you copy the thing back to CPU. or you can do dot realize and it will actually like force that tensor to be allocated in RAM. And if you think about it, PyTorch is kind of lazy in a way, but they didn't extend the paradigm far enough, right? When I do A times B in PyTorch, it's going to launch a CUDA kernel to do A times B, but it's not going to wait for that CUDA kernel to complete.

**[0:15:08] George Hotz:** So you're getting the worst possible worlds. You're getting the same laziness, but you also can't get fusion because PyTorch doesn't know that I'm then going to do plus C. There's no way for it to be like, whoa, whoa, whoa, don't launch that CUDA kernel. Whoa, just do this one. too, right? Again, PyTorch is working on this. And, you know, it's a little bit harder. Like in comma, I felt like I was competing against a lot of idiots.

**[0:15:28] George Hotz:** Here, I'm competing against, you know, smart, smart, very smart people who've made... Program language compiler people. Yeah. Who've made some, I think, different trade-offs. Whereas if you're trying to build something that is just straight up good on NVIDIA, and we have a lot of people and complexity to throw at it, yeah, PyTorch made a lot of the right choices. I'm trying to build something that manages complexity. Like you can always make your software do more.

**[0:15:50] George Hotz:** The magic is when you can make your software do more without adding complexity, right? Because, you know, complex things eventually collapse under their own weight. So it's kind of... How does fusing actually work? There's this thing called lazy.py. And when you do like A times B, that's, it's put into a graph, but it's a very local graph. There's no global graph optimizations. And even this can change, right? Again, like the programming model for TinyGrad does not preclude eagerness, right?

**[0:16:15] George Hotz:** Laziness is not guaranteed laziness. It's just going to try its best. So you put in A times B, and that's a binary op, right? And then you put in A times B, like that's a node in the graph. It's a virtual node because it's not realized yet, plus C. Okay, here's a new node, which takes the C tensor in here and takes the output of A times B. It's like, whoa, wait, there's two binary ops. Okay, we'll just fuse those together.

**[0:16:35] George Hotz:** Okay, here I have a kernel. This kernel has A, B, and C as inputs. It does A times B plus C in the local registers and then outputs that to memory. And you can graph.1 in TinyGrad. Another amazing thing that TinyGrad has that I've not seen in any other framework is two things. Graph equals one, which is an environment variable. It will output a complete graph of all the operations. People are like, oh, you can use PyTorch, export it to Onyx and use Netron.

**[0:16:59] George Hotz:** Yeah, you can. Like, what? That's not what's real. Graph.1 will show you the actual kernels that were dispatched to the GPU. You can also type debug equals two, which will print those kernels out in your command line. and it will tell you the exact number of flops and the exact number of memory accesses in each kernel. So you can immediately see, wait a second. Okay, this kernel used this many flops. This was the gigaflops. This is how many bytes it read.

**[0:17:25] George Hotz:** And this was the gigabyte per second. And then you can profile without having to like, okay, I mean, in theory in PyTorch, sure, use the NVIDIA Insight Profiler. No one does that. No one does, of course, because it's so difficult, right? Like, actually, NVIDIA used to, I think CUDA 9 was the last one they had it. They had a command line one, but now it's like, okay, I'm going to generate this blob, use this NVIDIA GUI tool to convert it into a Chrome trace and then load it.

**[0:17:49] George Hotz:** Yeah, no one does that, right? I'll just type debug equals two in any tiny grad model, and it will show you all the kernels that it launches and the efficiency of each kernel, basically.

**[0:17:58] swyx:** Yeah, this is something that John Carmack has often commented about, is that when you code, you need to build in your instrumentation or observability right into that. I wonder if whatever John is working on, he's adopting this style. And maybe we can sort of encourage it by, I don't know, naming it and coining a certain kind of debugging style. If you would like to start contributing to TinyGrad, I'd be so happy.

**[0:18:21] George Hotz:** I've chatted with him a few times. I'm not really sure what his company's doing. But no, I mean, hopefully we get TinyGrad to a point where people actually want to start using it. So TinyGrad right now is uncompetitive on NVIDIA. It's uncompetitive on x86.

**[0:18:35] swyx:** And specifically, what do you care about when you say uncompetitive?

**[0:18:38] George Hotz:** Speed. Straight up speed. It's correct. The correctness is there. The correctness for both forwards and backwards passes is there. But on NVIDIA, it's about 5x slower than PyTorch right now. Like 5x, wow, this is insurmountable. No, there's reasons it's 5x slower. And I can go through how we're going to make it faster. And it used to be 100x slower. So we're making progress. But there's one place where it actually is competitive. And that's Qualcomm GPUs. So TinyGrad is used to run the model in OpenPilot.

**[0:19:03] George Hotz:** Like right now, it's been live in production now for six months. And TinyGrad is about 2x faster on the GPU than Qualcomm's library.

**[0:19:10] swyx:** What about Qualcomm architecture?

**[0:19:11] George Hotz:** Oh, what makes it doable? Well, because the world has spent how many millions of man-hours to make NVIDIA fast? And Qualcomm has a team of 10 Qualcomm engineers? Okay, well, who can I beat here? What I propose with TinyGrad is that developer efficiency is much higher. But even if I have 10x higher developer efficiency, I still lose on NVIDIA, right? You know, okay, I didn't put 100 ,000 man-hours into it, right? If they put a million, that's what I'm saying.

**[0:19:36] George Hotz:** But that's what I'm saying we can get. And we are going to close this speed gap a lot. I don't support TensorCourse yet. That's a big one that's just going to massively close the gap. And then AMD. I don't even have a benchmark for AMD because I couldn't get it compiled. Oh, and I tried. Oh, I tried. I spent a day. I spent actually a day trying to get PyTorch. And I got it built. I got it kind of working.

**[0:19:59] George Hotz:** Then I tried to run a model. There's all kinds of weird errors. and the rabbit hole is just so deep on this. I'm like, you know, you can compare the speed. Right now, you can run Llama. You can run anything you want on AMD. It already all works. Any OpenCL backend works. And it's not terribly slow. I mean, it's a lot faster than crashing. So it's infinitely times faster than PyTorch on AMD. But pretty soon, we're going to start getting close to theoretical maximums on AMD.

**[0:20:22] George Hotz:** That's really where I'm pushing. And I want to get AMD on MLPerf in a couple months, hopefully.

**[0:20:26] Alessio Fanelli:** Now that you bring up AMD. Yeah, let's dive into that. Because when you announced the Semicorp Fundraise, you mentioned one of your first goals is like build the framework runtime and driver for AMD. And then on June 3rd on Twitch, you weren't as excited about AMD anymore. Maybe let's talk a bit about that. And like, you compared the quality of like commit messages from like the AMD kernel to like the Intel work that people are doing there.

**[0:20:50] Alessio Fanelli:** What's important to know?

**[0:20:51] George Hotz:** When I said I wanted to write, I want to write a framework. I didn't never intended on writing a kernel driver. I mean, like I flirted with that idea briefly, but like realistically, like there's three parts to it, right? There's like the ML framework, there's the driver, and then there's the user space runtime. I was even down to rewrite the user space runtime. I have a GitHub repo called CUDA IO control sniffer. It's terribly called, but you can actually launch a CUDA kernel without CUDA.

**[0:21:15] George Hotz:** So you don't need CUDA installed. Just the NVIDIA open source driver and this open source repo can launch a CUDA kernel. So rewriting the user space runtime is doable. Rewriting the kernel driver? I don't even have docs. I don't have any docs for the GPU. Like it would just be a massive a reverse engineering project. I wasn't complaining about it being slow. I wasn't complaining about PyTorch not compiling. I was complaining about the thing crashing my entire computer.

**[0:21:37] George Hotz:** It panics my kernel. And I have to wait five minutes while it reboots because it's a server motherboard and they take five minutes to reboot. So I was like, look, if you guys do not care enough to get me a decent kernel driver, there's no way I'm wasting my time on this, especially when I can use Intel GPUs. Intel GPUs have a stable kernel driver and they have all their hardware documented. You can go and you can find all the register docs on Intel GPUs.

**[0:21:57] George Hotz:** So I'm like, why don't I just use these. Now, there's a downside to them. Their GPU is $350. You're like, what a deal. It's $350. You know, you get about $350 worth of performance. And if you're paying about $400 for the PCIe slot to put it in, right, like between the power and all the other stuff, you're like, okay, nevermind. You got to use NVIDIA or AMD from that perspective. But I sent an email to Lisa Su.

**[0:22:18] George Hotz:** She responded. And I've had a few calls since. And like what I tried to do, first off, like, thank you for responding. It shows me that like, like, if you don't care about your kernel panicking I can't. This is just a huge waste of my time. I'll find someone who will care. I'm not asking for your 7x7 Winograd convolution when transposed to be fast. I'm not asking for that. I'm asking literally for - The basics. To not get running.

**[0:22:43] George Hotz:** Oh, and this isn't TinyGrad. This is your demo apps. I ran their demo apps in loops, and I got kernel panics. I'm like, okay. But no, Lisa Sue reached out, connected with a whole bunch of different people. They sent me a pre-release version of Rock M 5.6. They told me you can't release it, which I'm like, why do you care? But they say they're going to release it by the end of the month, and it fixed the kernel panic.

**[0:23:08] George Hotz:** The guy managed to reproduce it with the two GPUs and the computer, and yeah, sent me a driver, and it works. I had that experience, and then I had another experience where I had two calls with AMD's communication people. I tried to explain to these people open source culture. It's not open source if you dump the source code on a GitHub repo and then forget about it until the next release. It's not open source if all your issues are from 2022.

**[0:23:34] George Hotz:** It's just no one's going to contribute to that project, right? Sure, it's open source in a very technical sense. To be fair, it's better than nothing. It's better than nothing. But I fixed a bug in Nickel. There's a fun fact, by the way. If you have a consumer AMD GPU, they don't support peer-to-peer. And their all-reduced bandwidth is horrendously slow. because it's using CUDA kernels to do the copy between the GPUs. And it's putting so many transactions on the PCIe bus that it's really slow.

**[0:23:58] George Hotz:** But you can use CUDA memcopy. And there's a flag to use CUDA memcopy. But that flag had a bug. I posted the issue on Nickel. I expected nothing to happen. The NVIDIA guy replied to me within an hour. He's like, try this other flag. I'm like, okay, I tried the other flag. It still doesn't work. But here's a clean repro. And I spent like three hours writing a very clean repro. I ended up tracking the issue down myself.

**[0:24:20] George Hotz:** But just the fact that somebody responded to me within an hour and cared about fixing the issue. Okay. You've shown that it's worth my time. And I will put my time in. Because let's make this better. I'm here to help. But if you show me that you're like, you're the Colonel Panix. This is just expected. Okay. Well, it sounds like EMD is getting a message. They are. And I don't really think they've had someone explain to them.

**[0:24:42] George Hotz:** I was like, you're going to build in public. What's an example of building in public? I'm like, go look at PyTorch. Go look at PyTorch. I have two minor things merged into PyTorch. because it's very responsive, you know?

**[0:24:52] Alessio Fanelli:** So that's kind of like the lowest level of the stack. And then at a slightly higher level, obviously there's TinyGrad, there's Modul, there's GGML. How are you thinking about breadth versus like depth, like where you decided to focus early on?

**[0:25:06] George Hotz:** So GGML is very much like a, okay, everyone has M1s, right? Actually, I was thinking in the beginning, I was thinking of something more like GGML, focus on the M1s, but GGML showed up and was just like, we're actually just focusing on the M1s. And actually, M1 PyTorch is considerably better than AMD PyTorch. M1 PyTorch works. It only gives wrong answers sometimes, and it only crashes sometimes. But like some models kind of run. When I was writing the Metal backend, I was comparing to MPS PyTorch, and I had like a discrepancy.

**[0:25:33] George Hotz:** TinyGrad checks all its outputs compared to Torch. And I had one where it didn't match. I'm like, I checked the matrix by hand. It matches TinyGrad. I don't understand. And then I switched PyTorch back to CPU, and it matched. I'm like, oh. Well, there's like bugs. It's like if you like transpose the matrix, because like, I think it's like has to do with like multi views and PyTorch and like weird under the hood stuff that's not exposed to you.

**[0:25:54] George Hotz:** Like there's bugs and maybe they fix them. But like, you know, it seems like there was a lot of momentum again, because you're getting how many engineers care about making PyTorch work on M1, right? Thousands, tens of thousands. Yeah. And you have an open development process and guess what? It's going to be good. How many engineers care about AMD working, PyTorch AMD working? You got 10 guys that work for AMD and then like a couple hobbyists.

**[0:26:15] swyx:** You revealed an interesting detail about how you debug. You hand-check the matrix math?

**[0:26:19] George Hotz:** No, I don't hand-check it. One of the best tests in TinyGrad is a file called testops.py. And it's just 100 small examples written in TinyGrad and PyTorch. And it checks both the forwards and backwards to make sure they match.

**[0:26:33] swyx:** Good test suite. Yeah. Very important.

**[0:26:35] George Hotz:** That's, I mean, that's one of them. Where you like, I really, I put a lot of effort into the CI for TinyGrad. I think CI is super important. Like, I want that green check to mean I can merge this. Yeah. Like, I don't want my tests to. And if the green check, if you somehow manage to introduce a bug and get the green check, okay, we're fixing the test. Top priority. Mojo? It's closed source. No, I'm not that interested.

**[0:26:54] George Hotz:** Do you know what I mean? Like, look, I like Chris Lattner. I think he's going to do great things. And I understand the, like, kind of the wisdom even in keeping a closed source. But, you know, I'm interested when it's open.

**[0:27:05] swyx:** Yeah. You have an interesting design deviation from him because he's decided to be a, well, promised to be a superset of Python. And you have decided to break with PyTorch APIs. And I think that affects learnability and transportability of code.

**[0:27:18] George Hotz:** You know, if the PyTorch thing ends up being like a stumbling block, I could write a perfect PyTorch. Instead of import PyTorch, instead of like, yeah, import Torch, you type import tiny Torch as Torch. And if that really becomes the stumbling block, I will do that. No, Chris Lanner went much further than PyTorch. Replicating the PyTorch API is something I can do with a couple, you know, like an engineer month or two. Right, like a shim, yeah.

**[0:27:46] George Hotz:** Replicating Python? There's a big graveyard of those projects. How's Piston going? How's Jython? PyPy? You can go way back.

**[0:27:59] Alessio Fanelli:** So your core mission is commoditizing the petaflop. And then your business goal is to sell computers for more than it costs to me, which seems super reasonable. And you're going to have three tiny boxes? Red, green, blue?

**[0:28:12] George Hotz:** No No. know that was my look you know a lot of people like i love you know leaning into like saying i'm giving up right it's great to give up right giving up is this wonderful thing it's so liberating and then like you can decide afterward if you really give up or not there's very little harm in saying you give up except like you know great twitter haters have something to talk about and all press is good press kids so just read only read yeah any box red tiny box red unless amd you know upsets me again and then we're you know we're back to we're back to other colors we have other

**[0:28:40] Alessio Fanelli:** colors to choose from. When you think about hardware design, what are some of the numbers you look for? So Teraflops, it's Persegon is one, but like memory bandwidth is another big limiter. Like how do you make those straightoffs? Well, I mean, fundamentally I'm limited to what

**[0:28:54] George Hotz:** GPUs I can buy, but yeah, for something that I think a lot of people are going to want to reasonably do with a coworker of mine described them as luxury AI computers, right? Like luxury AI computers for people. And that's like what we're building. And I think a common thing people are going to want to do is run like large llama, right? Or large like Falcon or whatever. FP16 llama. FP16, exactly, exactly. You know, int8, I think, can work.

**[0:29:17] George Hotz:** I think that like what GGML is doing to go to like int4, like this doesn't work. Like, have you done, and maybe they have, but like I read what it was and I was like, this isn't from any paper. This is just some, like you made -

**[0:29:29] swyx:** Squeezing as much as possible.

**[0:29:31] George Hotz:** Yeah, you made up some quantization standard to make it run fast. And like, like maybe it works, but okay, where's like the hellaswag number, right? Where's your -

**[0:29:38] swyx:** The thesis is right that if you have hundreds of billions of parameters, that the individual quantization doesn't actually matter that much.

**[0:29:44] George Hotz:** Well, the real way to look at all of that is to just say you want to compress the weights, right? It's a form of weight compression. Quantization is a form of weight compression, right? Now, this is obviously not lossless. It's not a lossless compressor, right? If it's a lossless compressor and you can show that it's correct, then okay, we don't have to have any other conversation. But it's a lossy compressor. And how do you know that your loss isn't actually losing the power of the model?

**[0:30:03] George Hotz:** Maybe int 4.65b llama is actually the same as fb16.7b llama, right? We don't know. Maybe someone has done this yet, but I looked for it when it first came out and people were talking about it. And I'm like, it's not from a paper, right? The int 8 stuff is from a paper where they, like some of the int 8 stuff is from a paper. There's one paper, I think it's like llm.int8, where they actually do all the tests.

**[0:30:27] George Hotz:** And they didn't go fully int 8. they made like 90 % of it in date and kept like 10 % of it in FP16 for what they called like the like outliers or whatever so I think that this is not quite so easy and I think being able well so first off if you're training no one's gotten training to work with in date yet there's a few papers that vaguely show it but if you're training you're going to need BF16 or float16 so this is why I target that now the thing that you're going to want to do is run these large language models out of the box on your hardware in FP16 and that's memory bandwidth so you need

**[0:30:56] George Hotz:** large amounts of memory bandwidth too So ask how I trade off memory bandwidth and flop. So what cheap use can I buy?

**[0:31:01] Alessio Fanelli:** So first of all, you have this hiring process, which is you got to solve one of the bounties that are open on TinyGrad. There's no technical interview. One of them is Intate support. Do you already have some things you want to test on?

**[0:31:13] George Hotz:** We have Intate support. What I'd like to see somebody do is just load the GGML Intate Llama into TinyGrad and then benchmark it against the FP16 one. Intate already works in TinyGrad. It doesn't actually do the math in Intate. It does all the math still in FP32. So intake can mean you just have your weights in intake, or intake can mean you actually do math in intake. And doing your math in intake, the big gain that people care about is actually having your weights in intake because weights in intake mean less memory and less memory bandwidth.

**[0:31:43] George Hotz:** Whereas the math, keep it in FP32. On M1s, it doesn't matter what data type you're doing in the GPU. I'm not even sure it can do intake, but FP16 and FP32 is the same teraflops. So yeah, no, that's one of the bounties. One of the bounties is get intake, llama running with the indate weights. And then actually what you could even do, if you really want to test this, just take the FP16 weights, convert them to indate, then convert them back to FP16, then compare the unconverted and converted.

**[0:32:10] George Hotz:** Oh, that's a nice hack. Oh yeah.

**[0:32:12] swyx:** Right. Like, like, like, it should be lossless in the other direction.

**[0:32:15] George Hotz:** Yeah. I think FP16, it should be lossless in the other direction. I'm actually not 100% about that. Why not? Oh, cause like you ever try to like, like if you want to represent, If it was like int 16, it's not lossless. Sure. I think all of int 8 can be represented in FP 16, but I'm not 100 % about that. Just draw the bytes. We just have to do it, right? Just literally do it.

**[0:32:33] George Hotz:** There's only 256 to check. But yeah, either way. I mean, int 4, definitely. So do your int 4, convert it back, and now see, even with int 4 weights and FP 32 math, like, okay, how much does your performance degrade of this model?

**[0:32:47] Alessio Fanelli:** I think like you're planning to release the first tiny box ship them in like two to six, eight months, something like that. What's the top of mind for you in terms of building a team? Who should, who are you calling for? So as the GPU is picked out and you're like,

**[0:33:01] George Hotz:** well, I could make that computer with the GPUs. And my answer is, can you? Do you know how hard it is to put six GPUs in a computer? And people think it's really easy and it's really easy to put one GPU in a computer. It's really easy to put two GPUs in a computer, but now you want to put in eight. Okay. So I'll tell you a few things about these GPUs. They take up four slots.

**[0:33:20] George Hotz:** You can buy the nicest super micro you can't put eight of those in there you need two slot blowers if you want to use one of those those four you super micros you need two slot blowers right or water cooling right if you're trying to get the four slot cards in there you're going to need some form of water cooling there are some like chinese 4090s that are blowers right you have any blowers or water cooling if you're trying to get it in those things right so you are you doing water no i'm not using that chassis okay so now you want to get six gpus in a computer so that's a big challenge you're like oh just use a

**[0:33:44] George Hotz:** pcie extenders i saw it online as tech tips it works great. No, it doesn't. Try PCIe extenders that work at PCIe 4.0. And interconnect bandwidth is super important. They're going to work at 3.0. No PCIe extender I've tested, and I've bought 20 of them, works at PCIe 4.0. So you're going to need PCIe redrivers. Now, okay, how much is that adding cost, right? Like these things all get really hard. And then tiny boxes, I've even had another constraint to it.

**[0:34:09] George Hotz:** I want this thing to be silent. Not totally silent, but my limit is like 45, maybe 50 dB, but not super micro machine, 60 dB. We have a small, we have a compute cluster at comma. You got to wear, you got to wear your protection to go in there. Like it's some videos

**[0:34:25] swyx:** where you give a tour. Oh yeah. It's noisy. It's super loud. You got all these machines just

**[0:34:29] George Hotz:** screaming all those. Like if you have a blower, what is that thing? 10 ,000 RPM just screaming. Like I want to be able to use the normal big GPU fans and make this thing so it can sit under your desk, plug into one outlet of power, right? Six GPUs, your GPUs are 350 watts each. Can't plug that into a wall outlet. Okay. So how are you going to deal with that? Good questions, right? And you're not sharing them.

**[0:34:52] George Hotz:** Well, that one, I mean, that one is pretty obvious. You have to limit the power on the GPUs, right? You have to limit the power on the GPUs. Now you can limit power on GPUs and still get, you can use like half the power and get 80 % of the performance. This is a known fact about GPUs, but like, that's one of my design constraints. So when you start to add all these design constraints, good luck building a tiny box yourself.

**[0:35:10] George Hotz:** Obviously, it can be done, but you need something that has actually quite a bit of scale and resources to do it.

**[0:35:15] Alessio Fanelli:** And you see like the, under the desk, it's like one of the main use cases, kind of like individual developer use or...

**[0:35:21] George Hotz:** Yeah, what I also see is more of a like an AI hub for your home, right? As we start to get like home robotics kind of stuff, you don't want to put the inference on the robot, but you also don't want to put the inference on the cloud. Well, you don't want to put it on the robot because, okay, it's $1 ,500. watts, tiny box. You'll put batteries, not charge them. Bad idea. Just wireless. Wireless is 0.5 milliseconds.

**[0:35:43] George Hotz:** This is super fast. You don't want to go to the cloud for two reasons. One, the cloud's far away. Okay, it's not that far away. You can kind of address this. But two, cloud's also mad expensive. Cloud GPUs are way more expensive than running that GPU at your house. At least any rates you're going to get. Maybe if you commit to buy, well, yeah, I'm going to buy 10 ,000 GPUs for three years, then maybe the cloud will give you a good rate.

**[0:36:06] George Hotz:** But like, you want to buy one GPU in the cloud? I mean, okay, you can go to like VAST, but like if you're going to an Azure AWS, oh, that's expensive. This is like a personal data center

**[0:36:14] swyx:** instead of a cloud data center.

**[0:36:16] George Hotz:** We like the term compute cluster so we can use NVIDIA GPUs.

**[0:36:20] swyx:** Data centers may be a little bit dated. It's a compute cluster, which is totally legal under the CUDA license agreement. You talk a lot about the PCIe connection. Do you think there's any fat there to trim? What do you mean? You're limited by bandwidth.

**[0:36:32] George Hotz:** Okay, for some things, yes. So bandwidth is roughly 10x less than what you can get with NV-linked A100s. NV-linked A100s are going to have, and then you can even get like full fabric and NVIDIA really pushes on that stuff, 600 gigabytes per second. And PCIe 4, you're going to get 60. So you're getting 10x less. That said, why do you need the bandwidth? And the answer is you need it for training huge models. If you're training on a tiny box, your limit's going to be about 7 billion.

**[0:37:00] George Hotz:** If you're training on big stuff, your limit's going to be like 70 billion, right? Okay, you can hack it to get a bit higher. You can hack it like GPT hacked it to get a bit higher. But like that 65 billion in Lama, like there's a reason they chose 65 billion, right? And that's what can reasonably fit model parallel on a GPU, right? So yes, you are going to end up training models. The cap's going to be like 7 billion.

**[0:37:20] George Hotz:** But I actually heard this on your podcast. I don't think that the best chatbot models are going to be the big ones. I think the best chatbot models are going to be the ones where you had a thousand training runs instead of one. And I don't think that the interconnect bandwidth is going to matter that much. So what are we optimizing for instead of compute optimal? What do you mean compute optimal?

**[0:37:36] swyx:** You're talking about this, the Lama style models where you train for like 200 X. You train longer, yeah. Yeah, yeah.

**[0:37:42] George Hotz:** So, okay, you can always make your model better by doing one of two things, right? And a comma, we just have a strict limit on it. You can always make your model better by training longer and you can always make your model better by making it bigger. But these aren't the interesting ones, right? Particularly the making it bigger because training it longer, fine. You know, you're getting a better set of weights the inference is the same.

**[0:37:59] George Hotz:** The inference is the same whether I trained it for a day or a week. Okay, if it's 1 billion versus 10 billion, well, I tend to make my inference too, right? So I think that these big models are kind of, sure, they're great if you're research labs and you're trying to like max out this hypothetical thing.

**[0:38:13] swyx:** Which you can talk about later.

**[0:38:14] George Hotz:** Yeah, yeah, yeah. But if you're like a startup or you're like an individual or you're trying to deploy this to the edge anywhere, you don't need that many weights. You actually don't want that many weights.

**[0:38:24] swyx:** Optimizing for inference rather than capabilities during benchmarks.

**[0:38:28] George Hotz:** Yes. And I think the inference thing, right? There's going to be so much more. Right now, the ratio between like training and inference on clouds, I think it's only still like, I think it's like two or three X, right? It's two or three X more inference, which doesn't make any sense. There should be way more inference. There should be 10 to 100 X more inference in the world than training. But then also like, what is training, right?

**[0:38:47] George Hotz:** You start to see these things like LoRa, like it's kind of blurring the lines between inference and training. And I think that that blurred line is actually really good. I'd like to see much more like on-device training or on-device fine-tuning of the final layer. we're pushing toward this stuff at Comma, right? Like, why am I shipping a fixed model? I totally want this model to fine tune based on like how, you know, your left tire is flat, right?

**[0:39:07] George Hotz:** Every time you cut the same turn because your left tire is flat, well, I should learn that, right?

**[0:39:11] swyx:** So would Comma pursue parameter efficient fine tuning?

**[0:39:15] George Hotz:** Yeah, we're looking into stuff like that. I mean, Comma is already very parameter efficient because we have to like run this thing in a car and you have to like cool it and power it.

**[0:39:21] Alessio Fanelli:** And so this kind of like intelligence cluster you have in your home, you see when the person is using third-party model, they load them locally and kind of do the final fine tuning. It kind of stays within the box.

**[0:39:32] George Hotz:** I think that that's one version of it for the privacy conscious. I also see a world where you can have your tiny box in its down cycles. Mine, Flopcoin, right? You know, not all, turns out not all crypto is a scam. There's one way to tell if crypto is a scam. If they're selling the coin before they make the product, it's a scam. If they have the product and then they sell the coin, it's maybe not a scam, right?

**[0:39:53] George Hotz:** So yeah, my thought is like each tiny box would let you, would have a private key on it. And you have to do it this way. You can't just let anyone join because of civil attacks, right? There's a real problem of like, how do I ensure your data is correct? And the way that I ensure your data is correct on the tiny net is if you ever send wrong data, you're banned from the net of life.

**[0:40:08] George Hotz:** Your $15 ,000 hardware box is banned. So, you know, don't cheat. Obviously, if it messes up, we'll forgive you. Somebody's going to try to jailbreak your devices. There's no jailbreak. There's no jailbreak. It's just a different network. There's just a private key on each device, right? Like if you buy a tiny box from the tiny corp, I give you a private key, it's in my backend server, right? You want to hack my server, that's illegal. Anything you want to do on the device, the device is yours.

**[0:40:28] George Hotz:** My server's mine, right? Have you looked into like federated training at all? Okay, there's orders of magnitude. Federated training, you mean like over the cloud and stuff?

**[0:40:37] swyx:** Over the internet? Over the internet, but also distributed on a bunch of devices, right?

**[0:40:41] George Hotz:** Yeah. Which some people are. I'm very bearish on this stuff. Yeah. Because your interconnect bandwidth, right? So, okay. At the high end, you have your interconnect bandwidth of NVLink, which is 600 gigabytes per second, right? The tiny box has 60 gigabytes per second. And then your internet has 125 megabytes per second. Not gigabits, 125 megabytes. So, okay, that's how many orders of magnitude we're talking here? Like from 60 down to 125? Like, all right, that's over 100x.

**[0:41:09] George Hotz:** That's 400x, right? So what you can do is inference, right? For inference, you don't care. For inference, there's so little bandwidth at the top and the bottom of the model that, yeah, you can do federated inference, right? And that's kind of what I'm talking about. There's also interesting things to push into. Like you're like, but okay, what if you want to run closed source models? This stuff gets kind of interesting, like using TPMs on the boxes and stuff.

**[0:41:31] George Hotz:** But then someone might jailbreak my device. So, you know, maybe we don't try to do that.

**[0:41:34] Alessio Fanelli:** Yeah. What's like the enterprise use case? Do you see companies buying a bunch of these and like stacking them together?

**[0:41:39] George Hotz:** The tiny box is like the first version of what we're building. But what I really want to do is be on the absolute edge of flops per dollar and flops per watt. These are the two numbers that matter. So the enterprise use case is you want to train, like Comma. So Comma just built out a new compute cluster. It's about a person and a half.

**[0:41:56] swyx:** A person being 20 petaflops.

**[0:41:58] George Hotz:** A person is 20 petaflops. It's about 30 petaflops. We built out a little compute cluster. And we paid double what you theoretically could per flop. You theoretically could pay half per flop if you designed a bunch of custom stuff. And yeah, I mean, I could see that being, you know, TinyCorp, when Comma's going to be the first customer, I'm going to build a box for Comma. And then I'm going to show off the box I built for Comma and be like, okay, like, do you want to build, I sell $250 ,000 training computers?

**[0:42:24] George Hotz:** Or how much is one H100 box? It's 400 grand? Okay, I'll build you a 400 grand training computer and it'll be 10x better than that H100 box. Again, not for every use case. For some, you need the interconnect bandwidth. But for 90 % of most companies model training use cases, the tiny box will be 5x faster for the same price.

**[0:42:41] Alessio Fanelli:** You mentioned the person of compute. How do we build a human for $20 million?

**[0:42:46] George Hotz:** Well, it's a lot cheaper now. So like I said, Comma spent about half a million on our person and a half.

**[0:42:53] Alessio Fanelli:** What are some of the numbers people should think of when they compare compute to people? So GPT-4 was 100 % years of training. That's more like on the timescale. 20 petaflops is one person. I think right now the math was that for the price of the most expensive thing we build, which is the International Space Station, we could build one Tampa. Yeah, one Tampa of compute. It's the ultimate currency of measurement.

**[0:43:20] George Hotz:** Yeah, yeah, we could build. So like the biggest training clusters today, I know less about how GPT-4 was trained. I know some rough numbers on the weights and stuff, but Lama - 1 trillion parameters? Well, okay, so GPT-4 is 220 billion in each head, and then it's an eight-way mixture model. So mixture models are what you do when you're out of ideas. So, you know, it's a mixture model. They just train the same model eight times, and then they have some little trick.

**[0:43:42] George Hotz:** They actually do 16 inferences, but no, it's not like - So the multimodality is just a vision model kind of glommed on. I mean, the multimodality is like obvious what it is too. You just put the vision model in the same token space as your language model. Oh, did people think it was something else? No, no, the mixture has nothing to do with the vision or language aspect of it. It just has to do with, well, okay, we can't really make models bigger than 220 billion parameters.

**[0:44:04] George Hotz:** We want it to be better. Well, how can we make it better? Well, we can train it longer. And okay, we've actually already maxed that out. Getting diminishing returns there. Okay. Mixture of experts. Yeah, a mixture of experts. We'll train eight of them, right? All right. So, you know, the real truth is whenever a start, whenever a company is secretive, it's because they're hiding something that's not that cool. And people have this wrong idea over and over again that they think they're hiding it because it's really cool.

**[0:44:28] George Hotz:** It must be amazing. It's 1 trillion parameters. No, it's a little bigger than GPT-3. And they did an eight-way mixture of experts. Like, all right, dude, anyone can spend eight times the money and get that. Coming back to what I think is actually going to happen is, yeah, people are going to train smaller models for longer and fine-tune them and find all these tricks. OpenAI used to publish stuff on this, you know, when they would publish stuff about how much better the training has gotten given the same, holding compute constant.

**[0:44:55] George Hotz:** And it's gotten a lot better, right? Then compare like batch norm to no batch norm.

**[0:45:00] swyx:** Because you're finding algorithms like flash attention.

**[0:45:02] George Hotz:** Yeah, well, flash attention, yeah. And flash attention is the same compute. Flash attention is an interesting fact where it's actually the identical compute. It's just a more efficient way to do the compute. But I'm even talking about like, look at the new embeddings people are using. They used to use these boring old embeddings. Now Llama uses that complex one. And now there's Alibi. I'm not up to date on all the latest stuff. But those tricks give you so much.

**[0:45:23] swyx:** There's been a whole round trip with positional embeddings. I don't know if you've seen this discussion.

**[0:45:27] George Hotz:** I haven't followed exactly. I mean, you quickly run into the obvious problem with positional embeddings, which is you have to invalidate your KV cache if you run off the context. So that's why I think these new ones, they're playing with them. But I'm not an expert on the latest up-to-date

**[0:45:41] Alessio Fanelli:** language model stuff. What are some of the things, I mean, that people are getting wrong? So back to autonomous driving, there was like the whole like LiDAR versus vision thing, you know, it's like people don't get into accidents because they cannot see well, they get into accidents because they get distracted and all these things. Do you see similarities today on like the path to AGI?

**[0:45:59] George Hotz:** Nothing I say about this is ever going to compete with how Rich Sutton stated it. Rich Sutton, the writer of reinforcement learning, the bitter lesson. Nothing I say is ever going to compete with. The bitter lesson's way better than any way I'm going to phrase this. Just go read that. And then like, I'm sorry, it's bitter, but you actually just have to believe it. Like over and over again, people make this mistake. They're like, oh, we're going to hand engineer this thing.

**[0:46:20] swyx:** We're going to hand, no, like stop wasting time. I mean, OpenAI is not taking the bitter lesson. They were leaders in deep learning for a long, long, long time. Well, OpenAI was the absolute

**[0:46:28] George Hotz:** leader to the thesis that compute is all you need, right? And there's a question of how long this thesis is going to continue for. It's a cool thesis. And look, I think I would be lying along with everybody else. I was into language models way back in the day for the Hutter Prize. I got into AI through the Hutter Prize. 2014, I'm trying to build compressive models of Wikipedia. And I'm like, okay, why is this so hard?

**[0:46:49] George Hotz:** What this is, is a language model. And I'm playing with these Bayesian things. And I'm just like, oh, but I get it. I have two data points and they're almost the same, but how do I measure that almost? I just wrapped my head around this. And this was around the time Carpathia released the first RNN that generated the Shakespeare stuff. And I'm like, okay, I get it. It's neural networks that are compressors. Now, you can't actually win the Hutter Prize with these things because the Hutter Prize is MDL.

**[0:47:16] George Hotz:** It's the size of the model plus the size of the encodings, embeddings. So, yeah, you can't. I mean, probably now you can because it's gotten so good. But, yeah, back in the day, you kind of couldn't. So I was like, okay, cool. This is what it is. I kind of get it. I didn't expect that it would continue to work this well. I thought there'd be real limits to how good autocomplete could get. That's fancy autocomplete.

**[0:47:38] George Hotz:** But yeah, it works well. So like, yeah, what is OpenAI getting wrong? Technically, not that much. I don't know. Like if I was a researcher, why would I go work there?

**[0:47:48] swyx:** Yes. So why is OpenAI like the Miami Heat?

**[0:47:52] George Hotz:** No, look, I don't, I don't, this is my technical stuff. I don't really want to harp on this, but like why go work at OpenAI when you could go work at Facebook as a researcher? Open AI can keep ideologues who believe ideological stuff, and Facebook can keep every researcher who's like, dude, I just want to build AI and publish it.

**[0:48:08] Alessio Fanelli:** Yeah. Any other thoughts? Tiny Corp? Bounties?

**[0:48:11] George Hotz:** You know, I've been thinking a lot about what it means to hire in today's world. Okay, look, I'm a believer that machines are going to replace everything in about 20 years. So, okay. What is that thing that people can still do that computers can't? And this is a narrowing list. But like, you know, back in the day, like, imagine I was starting your company in 1960. Oh, we're going to have to hire a whole bunch of calculators in the basement to do all the, you know, math to support the calculus.

**[0:48:36] George Hotz:** Dude, have you heard about computers? Why don't we just buy a few of those? Oh, wow, man. Right. So like, I feel like that's kind of happening again. And I'm thinking about, I will post in my Discord. I'll be like, okay, who wants to like, okay, I just changed my unary ops used to be log and exp in like E. I changed them to be log2 and exp2 because hardware has log2 and exp2 accelerators. Yeah, and of course you can just change a base.

**[0:49:00] George Hotz:** It's one multiply to get it back to E. But like I made the primitives log2 and exp2, right? I just posted in the Discord. I'm like, could someone put this pull request up? And someone eventually did and I merged it. But I'm like, this is almost to the level where models can do it. We're almost to the point where I can say that to a model and the model can do it. Have you tried? Yeah. I don't know.

**[0:49:20] George Hotz:** I think autocomplete went further than I thought it would. but I'm also relatively unimpressed with these chatbots. The problem is if your loss function is categorical cross entropy on the internet, your responses will always be mid. Yes. Mode collapse is what I call it. I don't know. I'm not even talking about mode collapse. You're actually trying to predict these. Like, look, I rap. I'm a hobbyist rapper. And like, when I try to get these things to write rap, the rap sound like the kind of raps you read in the YouTube comments.

**[0:49:45] George Hotz:** Nursery school. Yeah, it's like, all right, great. You rhyme box with fox. Sick rhyme, bro. You know, and Drake is rhyming, give it up for me with napkins and cutlery, right? Like, all right, come on.

**[0:49:55] swyx:** He's got like this thing about orange. Orange is famous that you can't write.

**[0:49:59] George Hotz:** Yeah Yeah. But now, of course, you know, four-inch screws and orange juice is in GPT's training corpus. Yeah, so I think it went further than like everyone kind of thought it would. But the thing that I really want to see is like somebody put 10 LLMs in a room and have them discuss the answer before they give it to me. You can actually do this, right? And I think the coding things have to be the same way.

**[0:50:18] George Hotz:** There is no coder alive, no matter how good you are, that sits down, well, I'm going to start at cell A1 and type my program, and then I'm going to press run and it's going to work. No one programs like that. So why do we expect the models to, right? So there's a lot that still needs to be done. But at the TinyCorp, I want to be on the cutting edge of this too. I want to be like program generation.

**[0:50:39] George Hotz:** I mean, what is TinyGrad? It's a compiler. It generates programs. Generate the fastest program that meets the spec, right? Why am I not just having ML do that? So, you know, it's kind of a, you have to exist fluidly with the machines. And I've come around on a lot of stuff. I'm like, wait, TinyCrop should be a remote company. I can't do this in person. Really? Yeah. Like, comma makes sense to be in person. Like, comma, sure.

**[0:51:02] George Hotz:** Yeah, we're getting off in San Diego. But that was a six-year-old company, right? And it works. And it works for a certain type of people and a certain type of culture. But what's going to be different this time? Okay, remote. But now it's remote. And now I'm getting these people who apply. And I'm like, I literally have a thousand applications. I'm not calling you to do a technical screen. I can't really tell anything from a technical screen.

**[0:51:20] George Hotz:** What am I going to do? Make a code on a whiteboard? Like bring up a shared notebook document so we could, oh, like that's not going to work. Okay. So then I'm moved to the next thing. We do this at comma with good success programming challenges. I've also found them to be like completely non-predictive. I found one thing to actually be predictive and it's, wait a second, just write code in tiny grad. It's open source, right?

**[0:51:40] George Hotz:** And yeah. So, you know, I'm talking to a few people who've been contributing and like contribute or you know the job's not for you but you can do it remote and it's like it's a chill job like you're not you're like oh yeah well i work for the tiny corp like well you're writing mit licensed software like you see what it's doing right like we'll just i think think of it maybe more of like a stipend than a salary and then also some equity like if you know

**[0:52:00] Alessio Fanelli:** i get rich we all get rich how do you think about agents and kind of like thinking of them as people versus like job to be done sean built this thing called small developer it's in the same vein like

**[0:52:10] swyx:** The human in the loop with the language model and just iterating while you write code. I think that's absolutely where it goes.

**[0:52:17] Alessio Fanelli:** And there's like a, it's not like one thing. It's like there's small interpreter. There's like small debugger. It's kind of like all these different jobs to be done. It's a small world. Yeah, I know this is like the small podcast. It's like small AI meets tiny verb. So we're all in the same wavelength. How do you think about that? Do you think people will have a human-like interaction where it's like, oh, this is like the AI developer or like is it I'm the human being supercharged by the AI tools?

**[0:52:41] George Hotz:** Oh, I think it's, yeah, much more like I'm the human supercharged by the AI tools. I think that like coding is tool complete, right? Like driving is not tool complete. We hire people to drive who are like below the API line, right? There's an API line in the world, right? Love that. Yeah, there's an API line in the world. And like, you can think like Uber is a really clear example, right? There's the people below the API line and the people above the API line.

**[0:53:00] George Hotz:** And the way you can tell if you're below or above, by the way, is your manager a computer, right? Who's the manager of the Uber driver? Well, computer, right?

**[0:53:06] swyx:** Do the machine tell you what to do or do you tell machines what to do? Exactly.

**[0:53:08] George Hotz:** Exactly. So coding is too complete, right? Coding is too complete. Coding is above the API line. So it will always be tools supercharging your coding workflow. And it will never be you performing some like task like, okay, well, I can do everything except for actually starting a Docker container. Like it just doesn't make any sense, right? Yeah. So it will always be sort of tools. And look, we see the same stuff with all the, like people are like, stable diffusion is going to replace artists.

**[0:53:36] George Hotz:** or whatever. It's like, dude, like... It's going to create new artists. Did Photoshop replace artists? Like, what are you talking about? Right? Like, you know, a real artist's finger paint. They can't use brushes. Brushes are, you know, brushes are going to replace all the... Okay. Like, I just can't. Like, it's all just tools and the tools are going to get better and better and better. And then eventually, yes, the tools are going to replace us. But, you know, that's still 20 years away.

**[0:53:59] George Hotz:** So, you know, I've got a company to run in the meantime.

**[0:54:01] swyx:** So I've written about the APL line before and I think that's from Venkatesh. I don't know if you've...

**[0:54:06] George Hotz:** I definitely took it from someone. It's definitely not mine.

**[0:54:08] swyx:** It's VGR. But I also have a speculated higher line than that, which is the Kanban board. Like who tells the programmers what to do, right? So are you above or below the Kanban board? Has that evolved your management thinking?

**[0:54:22] George Hotz:** Yeah, like that's sort of what I mean. Like I'm just going to describe the pull request in two sentences and then like, yeah.

**[0:54:28] swyx:** So you are running the Kanban board or the bounties?

**[0:54:30] George Hotz:** Yes, the bounties are the Kanban board, exactly. And that is kind of the high level. and then like, yeah, we'll get AIs to fill in some and we'll get people to fill in others. And that's also what it means to be like full-time at TinyCorp, right? Would you start, and I wrote this up pretty concretely. I'm like, okay, step one is you do bounties for the company. Step two is you propose bounties for the company, right?

**[0:54:50] George Hotz:** You don't obviously pay them, we pay them, but you propose them. And I'm like, yeah, that's a good bounty. That like helps with the main workflow of the company. And step three is you get hired full-time, you get equity, we all, you know, maybe get rich. What else are you designing differently about the employee experience? You know, some people really like to like, like keep a separation, right? Some people really like to keep a separation between like employees and management or customers and employees, like a comma.

**[0:55:14] George Hotz:** You know, the reason I do the DevKit thing, it's like, dude, you buy a comma thing, you're an employee of the company. Like you're just part of the company. It's all the same thing. There's no like secrets. There's no dividing lines. There's no like, it's all a spectrum for like, you know, down here at the spectrum, like you pay. And then up here at the spectrum, you get paid. You understand this is the same spectrum of college, right?

**[0:55:32] George Hotz:** Like for undergrad, you pay, and then you get up here to like, you know, I'm doing a PhD program, you get paid. Okay, well, cool. Welcome to the, you know.

**[0:55:40] Alessio Fanelli:** What about comma bodies? You mentioned a lot of this stuff is clearly virtual, but then there's below the API line you actually need.

**[0:55:47] swyx:** Wait, this is a thing that's been announced? Comma bodies? We sell them. Oh, shit.

**[0:55:51] George Hotz:** You can buy them. There's a thousand bucks on our website.

**[0:55:53] swyx:** Oh, okay. No, no, no. I'm thinking about like what Tesla announced with like the humanoid robots. It's the same thing. Yeah, yeah, yeah.

**[0:55:58] George Hotz:** Except, of course, we made the comma version of it. Tesla uses 20 actuators. We use two. right? Like how do you, how do you build the simplest possible thing that can like turn the robotics problem into entirely a software problem? So right now it is literally just a comma three on a pole with two wheels. Um, it balances, keeps the comma three up there. And like, there's so much you could do with that already, right? Like this should replace, how many security guards could this replace, right?

**[0:56:25] George Hotz:** If this thing could just competently wander around a space and take pictures and, you know, focus in on things, send you a text message when someone's trying to break into your building, you know, like, like this could already do so much, of course, but the software's not there yet. Right? So how do we turn robotics into a thing where it's very clearly a software problem? You know, the people don't accept that self-driving cars are a software problem.

**[0:56:46] George Hotz:** Like, I don't, I don't know what to tell you, man. Like literally just watch the video yourself and then drive with a joystick, right? Can you drive? And we've actually done this test. you've actually done this test where you've had someone okay you just watch this video and here's a joystick and you gotta drive the car and of course they can drive the car it takes a little bit of practice to get used to the joystick but the problem is all the model right so it can now make the model better

**[0:57:07] swyx:** our second most popular episode ever was about segment anything coming out of Facebook which as far as I understand the state of the art in computer vision what are you hoping for there that you need for comma

**[0:57:17] George Hotz:** I haven't used segment anything like the large YOLOs or not I've used like large YOLOs and I'm super impressed by them yeah I got to check out segment anything. I don't think it's a distinct problem, right? Okay, here's something that I'm interested in. All right, we have great LLMs. We have great text-to-speech models. And we have great speech-to-text models. Okay, so why can I not talk to an LLM? Like, I can have a normal conversation with it.

**[0:57:39] swyx:** You can with the latency of like two seconds every time.

**[0:57:41] George Hotz:** Right. And then it feels so unnatural. It's this like staccato. Like, I don't like the RLHF models. I don't like the tuned versions of them. You take on the personality of a customer support agent, right? Like, oh, come on. I like Lama more than Chachibiti. Chachibiti's personality just grated on me. Whereas Lama, cool. I write a little bit of pretext paragraph. I can put you in any scenario I want. That's interesting to me. I think there is really no distinction between computer vision and language and any of this stuff.

**[0:58:15] George Hotz:** It's all eventually going to be fused into one massive. To say computer vision is solved, well, it doesn't make any sense because what's the output of a computer vision model. Segmentation? Like, what a weird task. Who cares? OCR. Who cares? I don't care if you can segment which pixels make up that laptop. I care if you can pick it up.

**[0:58:32] Alessio Fanelli:** And you're going to have the local cluster. You're going to have the body.

**[0:58:36] George Hotz:** Yeah. Yeah. I think that's kind of where that goes.

**[0:58:39] swyx:** Maybe we can paint the future of like, the year is 2050. You've achieved all you wanted at TinyCorp. What is the AI-enabled future like?

**[0:58:48] George Hotz:** Well, Tiny Corp's the second company. Kama was the first. Kama builds the hardware infrastructure. Tiny Corp builds the software infrastructure. The third company is the first one that's going to build a real product. And that product is AI Girlfriend. No, I'm dead serious. This is the dream product. This is the absolute dream product. Girlfriend is just the like - Stand-in. Well, no, it's not a stand-in. No No. I actually mean it, right?

**[0:59:11] George Hotz:** So I've been wanting to merge with a machine ever since I was mad little. How do I merge with a machine, right? And you can look at maybe the Elon style way of thinking about is Neuralink. I'm like, I don't think we need any of this. You ever do some of your friends maybe to get into relationships, and you start thinking of them and their partner as the same person? You start thinking of them as one person?

**[0:59:34] George Hotz:** I mean, they are kind of merged. Humans can just kind of do this. It's so cool. It's this ability that we already have. So I don't need to put electrodes in my brain to merge with a machine. I need an AI girlfriend. Right. So that's what I mean. Like this is, this is the third product. This is the third company. And yeah, in 2050, I mean, like, ah, it's so hard. I just like, maybe I can imagine like 2035.

**[0:59:59] George Hotz:** I don't even know 2050, but like, yeah, 2035, like, yeah, that'd be really great.

**[1:00:03] swyx:** In terms of merging, like, isn't it, shouldn't you work on brain upload rather than, yeah,

**[1:00:07] George Hotz:** girlfriend? Brain upload, right. I don't need brain upload either. Like there's, there's thousands of hours of me on YouTube, right? Yes. If you might, how much of my brain's already uploaded. That's only the stuff that you voice. Yeah, it's not that different. It's not that different, right? You really think a model with, you know, an exaflop of compute couldn't extract everything that's really going on in my brain? I'm a pretty open person, right? Like, I'm not running a complex filter.

**[1:00:31] George Hotz:** Humans can't run that complex of a filter. Like, humans just can't. Like, this is actually a cool quirk of biology. It's like, well, humans, like, can't

**[1:00:38] Alessio Fanelli:** lie that well. So is it good or bad to put all of your stream of consciousness out there? I mean,

**[1:00:43] George Hotz:** I think it's good. I mean, he's streaming every day. I want to live forever.

**[1:00:49] swyx:** We said off mic that we may be the first immortals, right?

**[1:00:52] George Hotz:** Yeah. This is how you live forever. It's a question of, okay, how many weights do I have? Okay, let's say I have 1 trillion weights. We're talking about 100 terabytes here. But it's not really 100 terabytes, right? Because it's column graph complexity. How much redundancy is there in those weights? So maximally compressed, how big is the weight file for my brain? Quantize it, whatever you want. quantization is a poor man's compression. I think we're only talking really here about like maybe a couple gigabytes, right?

**[1:01:20] George Hotz:** And then if you have like a couple gigabytes of true information of yourself up there, cool man. Like what does it mean for me to live forever? Like that's me.

**[1:01:28] Alessio Fanelli:** No, I think that's good. And I think there's a bit of like a professionalization of social media where like a lot of people only have what's like PC out there, you know? And I feel like you're going to get, going back to the ChatGPT thing, right? you're going to train a model on everything that's public about a lot of people.

**[1:01:44] George Hotz:** Then no one's going to run their model and they're going to die. Don't let you see on social media.

**[1:01:50] swyx:** We're moving on to what would normally be called the lightning round, but just general takes because you're a generally interesting person with many other interests. What does the goddess of everything else mean to you?

**[1:01:58] George Hotz:** Oh, it means that AI is not really going to kill us. Really? Of course. Tell us more. Lex asked me this, like, is AI going to kill us all? And I was quick to say yes, but I don't actually really believe it. I think there's a decent chance that AI kills 95 % of us. Okay.

**[1:02:12] Alessio Fanelli:** But they saw on your Twitch streams that you're with them. So they're not going to...

**[1:02:16] George Hotz:** No, I don't think... I actually... I don't also think it's AI. Like, I think the AI alignment problem is so misstated. I think it's actually not a question of whether the computer is aligned with the company who owns the computer. It's a question of whether that company is aligned with you or that government's aligned with you. And the answer is no. And that's how you end up dead. So what the guidance of everything else means to me is like, the complexity will continue.

**[1:02:35] George Hotz:** Paper clippers don't exist. There are forces. The paper clipper is cancer. The paper clipper is really just a perfect form of cancer. And the goddess of everything else says, yeah, but cancer doesn't win.

**[1:02:48] swyx:** It's a beautiful story for those who haven't heard it. And you read it out and I listened to it.

**[1:02:52] Alessio Fanelli:** Yeah. What are you grateful for today?

**[1:02:55] George Hotz:** Oh, man. I mean, it's all just like, I haven't, I haven't thinking about this stuff forever. Like that it's actually like happening and it's happening in an accessible way too. I guess that's what I'm really grateful for. It's not like AI is not some Manhattan Project style. You don't know anything about it. Closed doors. I'll fight really hard to keep it that way. I'm grateful for just how much is released out there and how much I can just learn and stay up to date.

**[1:03:21] George Hotz:** And I'm grateful to the true fabric of reality that I didn't need differential equations to understand it. I don't need some like, there's a limit to my math abilities. I can do most undergrad math, but I took some grad math classes and okay, now we're getting to the end of what I can do. And it's just the actual end of what I can do. I'm limited by my brain. But, you know, ML stuff? You need high school math.

**[1:03:45] George Hotz:** You know what I mean? When I learned to apply a matrix, seventh grade, it's all easy.

**[1:03:49] swyx:** You need more electrical engineering than you need high school math, really.

**[1:03:52] George Hotz:** Yeah. Well, you need electrical engineering to build the machines. But even that, these machines are simpler than the machines that have existed before. The compute stack looks really nice. So, you know, yeah. I'm grateful that it's all happening and I get to understand it.

**[1:04:04] Alessio Fanelli:** to be here. John Carmack mentioned there's about six insights we have left. Do you have an intuition for what some of the paths people should be taking? Obviously, you're working on one. What are some of the other branches of the tree that people should go under?

**[1:04:17] George Hotz:** I don't think I'm working on one of the six insights. I don't think TinyGrid's any one of the six insights. Something I really like that Elon does, and I try to be inspired by it, is look at the boring tunnel machine and ask how you can build a 10x cheaper one. All right. Look at the rocket. How can I build a 10x cheaper one? All right. Look at the electric car and say, how can I build a 10x cheaper or cheaper or can go further or whatever, whatever, whatever, right?

**[1:04:39] George Hotz:** And you just do the straight up physics math, right? I'm trying to do the same thing with ML frameworks, right? And in doing so, making sure that this stuff remains accessible. You could imagine a world where if Google TPUs were actually the ultimate, if Google TPUs were actually the best training things. I mean, actually, I'm kind of grateful for NVIDIA, right? Because if Google TPUs were the ultimate, now you have this huge closed source compiler in between XLA and the hardware.

**[1:05:04] George Hotz:** And yeah, that's just a really bad thing. So I mean, something that is somewhat upsetting about the TinyCorp is that it is trying to prevent downside. But it's not all trying to prevent downside. Like we're also building computers and we're gonna build some awesome, powerful, cheap computers along the way. So no, I'm not really working directly on any of the six tricks. I also think the six tricks are kind of gonna be like luck. I think it's just gonna be like, you know, please tell me more about what covariate shift is and how that inspired you to come up with batch normalization.

**[1:05:31] George Hotz:** Please tell me more about why it's a transformer and it has a query, a key, and a value, right? Like Schmidt-Huber described it better in fast weights. You know, like, I mean, my theory about why transformers work have nothing to do with this attention mechanism and just the fact that, like, it's semi-weight sharing, right? Like, because the weight matrix is being generated on the fly, you can, like, compress the weight matrix, right? Like, this is what that, there's an operation in the transformer, which, and by the way, this is like Qualcomm's S &P-E can't run transformers for this reason.

**[1:05:59] George Hotz:** so most matrix multiplies in neural networks are weights times values right whereas um you know when you get to the the the outer product in uh transformers well it's weight times weight it's a it's values times values yeah right um so snpe like doesn't even support that operation right so it's like that operation that gives the transformer its power it has nothing to do with the fact that it's attention right and this is a funny like but that is one of the six tricks right?

**[1:06:23] George Hotz:** Batch, like these norms are a trick. Transformers are a trick. Okay. Six more.

**[1:06:29] swyx:** So you talk about attention as weight compression.

**[1:06:33] George Hotz:** Compression is not exactly the right word. What I mean is that the weights can change dynamically based on the context. So there was this thing in Pack 8 in the Hutter Prize that I absolutely loved, and I've never seen it again in neural networks, and it's a really good trick. Okay. Imagine you have 256 weight sets for a layer, right? And then you choose which of the weight sets you're loading in based on some context. And that context can come from another neural net.

**[1:06:54] George Hotz:** So I have another neural net which projects 256 wide, one hot, do a softmax, predict it, and then I actually load the weights in. And I can do this operation at both test time and train time. I can do this operation at both training and inference, and I load in the weights given the context. That is what transformers do. But transformers, instead of having 256 discrete ones, it's actually just that, but continuous. Which is funny that that was in language models.

**[1:07:17] George Hotz:** And I just like, when I understood that about transformers, I'm like, oh, this is a

**[1:07:20] Alessio Fanelli:** real trick and why are they using the word attention yeah and today is actually the anniversary of attention is all you need what oh yeah that's a day today six years ago six years six years

**[1:07:30] George Hotz:** change the world wow well there's one of your envelope tricks right and you can easily write it on an envelope you know think about how you write out that how many times have you written that because it's not in any libraries because it's like all used a little differently each time yeah if you just write out that exact same you know you've name checked uh elon a few times yeah

**[1:07:46] swyx:** I think about both of you as systems thinkers. Input, output, something in between. Sure. What's different about your style versus his? Elon's fundamental science for the world is physics.

**[1:07:55] George Hotz:** Mine is information theory. But you do a lot of physics as well. I mean, like you base it on. And Elon does a lot of information theory as well too. But the difference maybe is expressed in what your ambitions are, right? Elon's ambitions may be like… Go to Mars. Go to Mars, right? Go to Mars is the ultimate modernist physics ambition, right? It's a physics, but I'm getting to Mars, right? What are electric cars? It's a physics problem, right?

**[1:08:19] George Hotz:** Okay, now he's like pushing on the autonomy stuff and you push a little on information theory, but fundamentally his dreams are physics-based dreams. My dreams are information-based dreams. I want to live forever in virtual reality with my AI girlfriend. Those are the aspirations of someone who accepts information theory as a core science. So I think that's the main difference between me and him. He has physics-based aspirations and I have information-based aspirations.

**[1:08:39] swyx:** Mark Andreessen, he is a, hi Mark, he's a listener. He is a big proponent of effective accelerationism. You've been a bit more critical. Why do you say that EAC is not taken seriously by its adherents?

**[1:08:50] George Hotz:** Oh, well, only the left takes ideology seriously. It's just like a fact, right? Is the right more cynical? Is that what it is? I don't know. It's like the left actually manages to get energy around the ideologies, right? Look, here you have two effective altruists named Sam going in front of Congress. Only one of them is in jail. You know, it's interesting. They're both calling for regulation in their respective spaces, right?

**[1:09:11] swyx:** So SBF is definitely like kind of a wolf in sheep's clothing. kind of, right? Like, he only adopted EAC or EA to market.

**[1:09:19] George Hotz:** Sam Altman is a genuinely good guy who is not interested in power seeking for himself. All right. We don't have to go. Fair enough, fair enough. But no, EAC is not like, like you are not serious, right? Mark Andreessen, I like Mark Andreessen, but it's like someone who's like 2019 whose like eyes were opened about like the political world being not exact. You mean all the people on the news were lying to me? Yeah, bro, they were lying to you.

**[1:09:43] George Hotz:** Like, okay, we all figured this out five years ago. Now, what are you going to do about it? I'm going to complain about it on Twitter. Great. And that's what the yak is.

**[1:09:51] Alessio Fanelli:** Last and maybe most important, why was Avatar 2 bad?

**[1:09:55] George Hotz:** I have a whole, you can go on my blog. I rewrote the script of Avatar 2. I wrote a script that actually might make you feel something for the characters. I killed Jake Sully in the first scene. Like, you had to. Do you really think his second story arc topped his first one? No, of course not. You had to kill the guy and make the movie about the brothers, right? And just that alone and realizing that.

**[1:10:14] George Hotz:** You could have kept the Titanic scene. I didn't even take it out. I left your Titanic scene, James Cameron. But I wrote you a story. So you're just...

**[1:10:23] swyx:** He needs ships to sink in water.

**[1:10:25] George Hotz:** Look, it's a great scene. But the movie was just like... The romance... I never remember.

**[1:10:30] swyx:** Great CGI let down by the writing, maybe.

**[1:10:32] George Hotz:** It's a beautiful world. And that's why I care so much. You don't hear me ranting about Pirates of the Caribbean 2 being a terrible story. Because come on, what do you expect, man? And Johnny Depp's like, wow, I had a movie that made me rich. I love this.

**[1:10:45] Alessio Fanelli:** But this goes back to the midpoint. I think you wrote, it feels like ChatGPT wrote the movie. And that's my worry a little bit. It's kind of converging towards that.

**[1:10:54] swyx:** Moloch wrote the movie. Sorry, I didn't want to interrupt you.

**[1:10:59] George Hotz:** No, I closed a pull request two days ago. I was like, was this written by ChatGPT? And I just closed it. You know what? I honestly feel bad if you were a human who wrote this. Incapable of being more patex. But if I have a classifier running in my head that asks, is this an AI or is this a human? The only way to deal with all this, it's like the worst possible. People are like, how are you mad about these chatbots?

**[1:11:25] George Hotz:** You're not mad about Tesla. I don't want to buy a Tesla. I don't have to buy a Tesla. And it won't really impact my life negatively. But if I don't want to use a chatbot, it's still going to impact my life negatively. of the amount of like personalized spam that now makes me spend more cycles on my classifier to tell if it's spam or not, because you can now use AIs and generate this so cheaply. Like, no, I mean, we have to move to a model where everything's just a dollar, right?

**[1:11:46] George Hotz:** Like you want to send me an email, it's a dollar. Like you guys wouldn't have cared. None of my friends would care. No one would care except the spammers, right? Like we just got to move to those sort of models.

**[1:11:54] Alessio Fanelli:** Awesome. One last message you want everyone to remember.

**[1:11:58] George Hotz:** Go try TinyGrad. I hope that we're a serious competitor to what's out there. And then I want to take it all the way. We'll start with just building something for GPUs and then we'll start building chips and then we'll start building fabs and then we'll start building silicon mines and then we'll have the first self-reproducing robot using... Yeah, okay.

**[1:12:17] Alessio Fanelli:** All right, George. Thank you so much for coming on. You did a big inspiration. Thank you. Thanks. Thank you.
