GioStrives

Striving to use AI and tech to improve our world...while documenting my journey.

[GioStrives @GioStrives]

https://t.co/WWFaATM6Xx

[GioStrives @GioStrives]

@Taniyatweets_ I don’t know which of those would surprise me most...

I do know what definitely won’t: the trailing plow

[GioStrives @GioStrives]

@erikalee #happylife

[GioStrives @GioStrives]

so did the number of problems which was solved by at least one of the LLM outputs.

For more details: https://t.co/T6uYyJBCE7

[GioStrives @GioStrives]

The Infinite monkey theorem states that an infinite amount of monkeys typing an infinite amount of random, independent, characters will type, almost surely, any kind of text: from Dante's Divine Comedy to this very post.

While this can't be done in practise, a paper took inspiration from this concept...

[GioStrives @GioStrives]

They took a "small" LLM and run it multiple times to solve problems. This is called "inference scaling": the model remains the same, the prompt too, but the outputs are all slightly different.

What was observed is that, as the number of "LLM monkeys" increased,

[GioStrives @GioStrives]

@raqisright I'm down to baking a loaf of bread for $50.

[GioStrives @GioStrives]

@enjojoyy Yes, that's why we need university courses!

[GioStrives @GioStrives]

@matteopelleg Source? I mean, maybe in a lab setting; but with me sitting next to it and heat coming in or going out from windows and heaters I honestly think removing background noise would be very hard.

[GioStrives @GioStrives]

@itsalicesoul That's a joke, of course the easiest programming language is Brainf*ck.

(That's another non funny joke)

[GioStrives @GioStrives]

@net_termina These architectures are too different to compare without real benchmarks.

An example: I have 1080s and 2080s. Qwen image 2.1 on 2080s is about 30x faster. Why? 1080s have no Tensor Cores and so have no optimisation for FP16.

Not to mention all these cards can't use bf16...

[GioStrives @GioStrives]

@BlastBurn27 @ItsmeAjayKV VRAM is the bottleneck.

I managed to make it run on 1080s and 2080s so yeah, 3060 will be just fine.

[GioStrives @GioStrives]

@itsalicesoul It depends on your budget; they burn budgets so fast that I try to avoid using them, whenever possible.

[GioStrives @GioStrives]

Keep you posted. I mean, keep myself posted, I know I am the only one reading this...if time won't change that, at least I will have learned a lot 😊

This is the lesson btw: https://t.co/yV22VYINle

[GioStrives @GioStrives]

Stanford CS329A lecture 1 on YouTube surprised me.

I expect university courses to lag the field. This one cited the DeepSeek paper from December 2024 and walked through what AI research actually looks like now.

[GioStrives @GioStrives]

That’s the other gift of a course: too many papers get published and I often just collect them, without going in-depth.

A syllabus forces you to go deep on a few.