Well, neither Linux nor LLVM loudly proclaimed that they would be the next Internet or GUI. So I am inclined to believe that this will not be the case and the person doing the proclamation might be a little full of himself.

TinyGrad is GeoHot's system/compiler to map neural networks onto hardware. He consistently points out this one point: Because the exact number of cycles is know in advance, it can be scheduled, there's no need for branch prediction, or that type of thing in a CPU.

Essentially, he wants to be able to express programs, and even an operating system, as a directed acyclic graph of logical binary operations, so that you can have consistent and deterministic runtime behavior.

The bit about LLMs is a distraction, in my opinion.

> While there may be a legacy Linux running in a VM to manage all your cloud phoning spyware, the core functionality of the lifelike device is boot to neural network.

No, I do not think future devices will be "boot to neural network." Traditional algorithms still have a place. Your robot vacuum cleaner (his example) may still use A* to route plan, and Quicksort to display your cleanings in terms of most energy usage.

> Without CPUs, we can be freed from the tyranny of the halting problem.

Not sure what this means but I think it still makes sense to have a CPU directing things as in current architectures. You don't just have your neural engine, you also have your GPU, Audio system, input devices, etc. and those need a controller. Something needs to coordinate.

> Without CPUs, we can be freed from the tyranny of the halting problem.

Can someone please explain to me what this even means in this context?

Serious question.

He's got the kernel of a good idea. Deterministic data flows are a good thing. We keep almost getting there, with things like data flow architectures, FPGAs, etc. But there's always a premature optimization for the silicon, instead of the whole system. This leads to failure, over, and over.

He's wrong in the idea of using an LLM for general purpose compute. Using math instead of logic isn't a good thing for many use cases. You don't want a database, or an FFT in a Radar System to hallucinate, for example.

My personal focus is on homogeneous, clocked, bit level systolic arrays.[2] I'm starting to get the feeling the idea is really close to being a born secret[1] though, as it might enable anyone to really make high performance chips on any fab node.

[1] https://en.wikipedia.org/wiki/Born_secret

[2] https://github.com/mikewarot/Bitgrid

I don't know why I should switch from PyTorch to Tinygrad as a researcher and practitioner. In terms of kernel fusion, there is torch.compile. Not to say there is a large ecosystem behind PyTorch and almost every paper today is published with a PyTorch implementation. Probably what Tinygrad shines is bare-metal platforms?

I don’t understand the LLVM comparison. Is it somehow a compiler backend for conventional programming languages? Can you run C or Rust code?

> tinygrad has a hardware abstraction layer, a scheduler, and memory management. It's an operating system

Doesn't every ML framework have that?

I generally don't read anything by gh but I think he is cryptically just referring to something like XLA, whereby your NN architecture gets compiled straight to hardware, say to a custom asic, or to an FPGA bit stream, etc.

It's definitely going to happen but I don't think it will replace CPU's much like human brains can't quite replace CPU's and what they are optimised for.

Trying to make out that TinyGrad is leading the charge in this is quite self indulgent.

>Without CPUs, we can be freed from the tyranny of the halting problem.

In the same way that we can be freed of the tyranny of being able to write a for loop.

The only reason neural networks don't have control flow is because they are not very good. They are incredibly inefficient and the only way to properly solve that is to introduce control flow, for example: https://arxiv.org/abs/2311.10770

Great... Does this mean my pc will hallucinate kernel panics when it doesn't even have a kernel?

no it won't, because while hitting ioctls in python is cute

https://github.com/tinygrad/tinygrad/blob/master/extra/hip_g...

it is definitely not shippable

Isn’t this the guy who joined Twitter as an intern to “fix” search?

Yeah, good luck with that, lol