Jonas

ai/ml, mechinterp, diffusion, llms, compilers, robotics, space

[Jonas @LoosJonas]

@JankDankins_ @liminal_bardo They can (at least sometimes) https://t.co/JlzqCKKy4a

[Jonas @LoosJonas]

@Briankeating I mean if that question is our problem and "AI can solve our problems" ...

[Jonas @LoosJonas]

@pors it's in fact ~2 orders of magnitude slower than standard NN in my experiments (and I'm having trouble replicating IBNNs advantages)

[Jonas @LoosJonas]

@pors looks interesting, but also incredibly slow

[Jonas @LoosJonas]

@IterIntellectus Some aspects are quite uncontroversial & universal although often fuzzy, e.g. "lying is usually bad", "being helpful tends to be good", "don't eradicate humanity", etc. - so I think that at least some progress is possible

[Jonas @LoosJonas]

@VictorTaelin Maybe use AI for the first 95% just to see if sth is possible and then reimplement from scratch?

[Jonas @LoosJonas]

I was thinking a bit about a new lock design that prevents picking by locking the pins in place before validation

[Jonas @LoosJonas]

@che_shr_cat Interesting method! Just wondering - are there also cases where it's strictly worse than standard AdamW? Or is it always at least similarly effective?

[Jonas @LoosJonas]

@ebarenholtz If the changes the layers introduce in the residual stream meaningfully correspond to geometric operations in the manifolds visualized in these fancy animations - then "they think in shapes" doesn't really sound like the worst description, or?

[Jonas @LoosJonas]

@mariusmosbach @johnhewtt (semi-implicitly, their KL penalty apparently does smth similar - https://t.co/UoCQhvbCV0)

[Jonas @LoosJonas]

@mariusmosbach @johnhewtt My first idea was to add this as an extra regularization loss, so I was positively surprised seeing that it apparently already learns this implicitly

[Jonas @LoosJonas]

@tszzl Having them on a separate machine with remote control is such a blessing, except for all the anthropic outages

[Jonas @LoosJonas]

@giffmana Emergent Goblification

[Jonas @LoosJonas]

@bojie_li Apparently only ~1% improvement could be achieved by increasing question catalogue size, so it seems already good.

Nice that the data is public!

[Jonas @LoosJonas]

@bojie_li Awesome work! Did you check how size prediction accuracy scales with benchmark size? Would it make sense to double the number of questions to get better predictions, or is it saturated already?

[Jonas @LoosJonas]

@sytelus > Con: we will never know what models are thinking.

I guess it would be relatively easy to build a translator. Will not be perfect, but even todays CoTs aren't necessarily 100% faithful.

[Jonas @LoosJonas]

@kalomaze could it be that it's pretrained with all data in chat format and only this template?

[Jonas @LoosJonas]

@VictorTaelin Nice! Would probably be helpful if it included more hard problems, because if already current models score ~90%, distingushing future models seems noisy

[Jonas @LoosJonas]

@marvinsxtr @gabrieldernbach @TUBerlin @bifoldberlin @AIgnostics @ChariteBerlin nice project page and awesome to see code/model/data being public!