SPascareli13

That's the smallest dan luu post I've seen.

There's a new SOTA model every few months, are you supposed to buy a new chip every new release?

My thought was that ASICs turned out not to be worth it for crypto mining because consumer hardware evolved fast enough to do it, while also being cheaper and having some resale value, while ASICs are useless besides mining and have no resale value.

So I'm extrapolating this same idea to LLM inference.

Yet it doesn't seem that ASICs will have any particular advantage over consumer hardware since AI is very memory heavy, which is (right now) expensive no matter how you package it. And the compute is just simple matrix multiplication, which is almost entirely what GPUs were meant to do anyway.

Like how crypto used ASICS but then didn't because the scaling of consumer hardware made it obsolete?

But that assumes I have some switch case somewhere right? If I passed an array of "orders" to a downstream function, it knows that what it has is orders, not something else, so it doesn't need to check for anything, and in the case I checked for errors upstream (when I try to allocate a new order) all downstream functions know that no invalid order can be passed, in which case you only have one check at creation time.

That's why I haven't fully understood yet how working like this is simpler.

I never fully understood how to work with this "reserved" values being valid instead of errors when you can't allocate more memory, in either case you still need to check if you have a real entity or a reserved/error, right? Does it really make things simpler?

Amazing writing, very accessible too.

So batching requests is always something I think should increase performance by a lot, but most server implementations make this pretty difficult, but the thing I struggle the most to understand is how to keep the latency down if you have multiple clients request all batched together? The total amount of latency for all clients is always the latency for the slowest.

You need some JS in the middle but yes, you write your image to a buffer and share it with JS, then in the JS you put it in the canvas.

Just trying to learn C again, making things from scratch in a multiplatform way, interfacing with X11 on Linux and wasm on the browser.

It's been fun dealing with memory and C's weird design in this age of agentic coding.

Interesting aside, BYD (Chinese EV automaker) has took over production in the old Ford factory in Camaçari.

If the model is trained to be a interpreter, then that means that the loss should reach 0 for it to be fully trained?

Also, if it's execution is purely deterministic, you probably don't need non linearity in the layers, right?

My current stance with reviewing code is: It's not ok to make another human review the code you made with AI, if you used AI then you're the reviewer, so unless you come to me with a well defined question or decision to make, just merge it and take responsibility.

Obviously that could only work in a high trust environment, that why open source suffers so much with AI submissions.

Freeciv was what brought me too the civ world, I'm sure this project will be the same for many children of this generation.

I really dislike this idea of testing in go: only ever use an interface, never the real implementation + mockgen the mocks based on this interface + use the mocks to assert that a function is called, with exactly this parameters and in this exact order.

I find this types of tests incredibly coupled with the implementation, since any chance require you to chance your interfaces + mocks + tests, also very brittle and many times it ends up not even testing the thing that actually matters.

I try to make integration test whenever possible now, even if they are costly I find that the flexibility of being able to change my implementation and not break a thousand tests for no reason much better to work with.