Post

daniel holmgren šŸ« 
@dholms.at

that basically explains why AI stuff blows itā€™s like auto-tune for ideas, like yeah you hit the note but you lose the subtlety, the richness, the soul

59 reposts273 likes

daniel holmgren šŸ« 
@dholms.at

ofc you can correct this by saying ā€œdonā€™t over pick 7 or under pick it. really try for a uniform distributionā€ but itā€™s an uphill battle and this is a tiny isolated example. obvs way harder to head off with anything of actual complexity

pottfrillan
@pottfrillan.bsky.social

Source code exists to communicate with humans as well as computers and I think that AI generated code will lose out in that regard too, not because of "richness of the soul" but for similar reasons to why AI isn't good at getting a point across in text

daniel holmgren šŸ« 
@dholms.at

that's true. banging out dumb code is so easy with these things. but it was an uphill battle to write the reference implementation of our new protocol using them. as you point out, readability matters quite a bit for reference impls

š€¹š€©š€–š€™ Mina
@covering.space

Are you thinking of random.org perhaps? For what it's worth, for cryptographic purposes when randomness better than a pseudo-random number generator is needed they'll combine things like the precise timing of mouse movements with a complicated algorithm and basically get true random numbers

random.orgRANDOM.ORG - True Random Number ServiceRANDOM.ORG offers true random numbers to anyone on the Internet. The randomness comes from atmospheric noise, which for many purposes is better than the pseudo-random number algorithms typically used ...
Neon Genderless Evangelion
@tommyblacksox.bsky.social

I personally get hung up on absolutes all the time, so no, nothing can be random in the sense that it will be influenced by *something,* like dice rolling differently in the wind. I think hooking a *truly random* input to a computer would degrade the signal quality by interpreting it

J Dub
@johnwalters.org

Isn't this true of most, if not all ML models? We all want to identify the signal in a sea of noise, but you do risk missing out on some interesting noise.

Rob Ramos
@rob.tempomusic.fan

ā€œAutotune for ideasā€ might be the best way Iā€™ve seen LLMs be described. Because autotune can be used to creative derivative works or can be used as a method of expression.

Kevin Kuhl
@kevinrkuhl.com

It's kind of ridiculous that there's no interface for temperature or top-p in most LLM interfaces. Everything gets pushed behind layers of prompts and makes it really hard to use as a tool that you have control over.

Etienne
@etienneburdet.com

Shameless plug but that's the same idea here. They really don't seem to be able to test random things like we do (at inference time, this exactly what they do during training!). For now... mu.social/profile/etie...

Etienne@etienneburdet.com

First, Opus would probably still be burning tokens in loops without me. I would still be writing backup, rollbacks and tooling without it. Also feels LLMs have an extremly "adult" way of "thinking"ā€”answering?ā€”, where I can turn on the six years old in me any time šŸ˜