. @ollama is actually gollama.

softmax( (Q * K^T / sqrt(d_k)) ) * V is all you need.

and caught up with the sglang cache-aware routing design. this primarily routes based on perfix overlap, and each worker process the full range of text. i still think there should be a routing algorithm that squashes a sub range into each worker to maximize prefix overlap in a.

Spent a whole day reading transformer impl on Tinygrad and distributed LLM inference hmm.

BentoML has some good blogs

Things could have been worse.

me rn

is it fair to say @Etched is the opposite of @__tinygrad__ ?.

looking to implement a minimal / educational llm inference engine on top of tinygrad, but pounders the question: . what optimization should be implemented by hand, and which part will be supposed to be optimized by @__tinygrad__ itself?.

To really live, you must almost die.

Doing things produce information, just keep doing things.

Can it not be called "AskGPT"?. What else do people say beside asking questions?.

you can just ask things (since 2023).

In the Tesla app

Went into Tesla service center for a leaky tire replacement, tells me Tesla lowers prices for tires as "needs to be competitive".

yacine posts get infinitely better since unemployment.

Fate loves irony but hates hypocrisy.

RT @yacineMTB: kind of shocking. i actually did learn reinforcement learning in two weeks.

RT @elonmusk: @teslaownersSV @Gfilche @robotaxi Waiting on regulatory approvals, but probably in a month or two.