crumb ⏩⏩⏩⏩⏩⏩

crumb ⏩⏩⏩⏩⏩⏩

@crumb.bsky.social

robogirl in the nursing and architecture of diverse intelligences and of mind-affording substrates. https://crumb.offprint.app • https://hf.co/crumb • xe/xem/xer

634 followers679 following1.2K posts

crumb ⏩⏩⏩⏩⏩⏩ reposted

gerge
@gerge-lemons.bsky.social

Machine learning has a history of people trying ideas and them not working due an unrelated bug in their code (before coding agents). Just thinking about how much of science was like this throughout history. Leaning towards the opinion that we've actually missed lots of low hanging fruit this way

crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

very task dependent, it can engineer full forward+backward kernels but cant write an svg to a file. i added a little detection circuit that will gently nudge it about the format and will gradually ween it off

crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

I've never had trouble getting models to do something for me, maybe says something about abliterated model users' character/manner, or maybe just about the models that I use and my use cases. Who is really to say

crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

sorry ab cringe >> in dn, "rent lowering," i dont like hearing from folk doing seemingly every other thing than taking action to help prevent the future they apparently don't want. if you put the pressure on these things to happen behind closed doors there's no auditing at all.

View full thread
crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

sorry ab cringe >> in dn, "rent lowering," i dont like hearing from folk doing seemingly every other thing than taking action to help prevent the future they apparently don't want. if you put the pressure on these things to happen behind closed doors there's no auditing at all.

crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

it's super annoying to have to untangle their views that are entirely based on feeling/emotion when i want to be having serious conversations about the architecture and math and where this shit goes for real and not in theory

View full thread
crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

we didn't till either of the ~300 sqft patches of garden because we were basically lazy but to make up for it we built this raised bed and nice basket looking setups for the potatos that we can use next year alongside the gardens

crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

and built up a pretty herb garden and flowerbeds and beds for winecap mushrooms. i am not feeling so bad about what we have accomplished but it is not as much as i would have wanted to do

crumb ⏩⏩⏩⏩⏩⏩ reposted

Nathan Lambert
@natolambert.bsky.social

TLDR on the kimi k3.1 pro benchmarks (just released) is that it looks like they figured out their RL, where the initial release was a very impressive base model with lighter post-training. China is very much keeping pace with capability gains.

crumb ⏩⏩⏩⏩⏩⏩ reposted

Danny Page
@danny.page

Replying to @danny.page

the only moat that anthropic and openai have left seems to be sinophobia

crumb ⏩⏩⏩⏩⏩⏩
@crumb.bsky.social

am trying to train an LM made of boolean logic with a different kind of relaxation: RLing a "decision model" to act as a generative grammar conditioned on which neuron it's generating for. it's So Slow to train but a forward pass takes single Microseconds on a single cpu thread

crumb ⏩⏩⏩⏩⏩⏩ reposted

garrison
@garrison.corporate.fm

Replying to @crumb.bsky.social

> We offer no explanation as to why these architectures seem to work; we attribute their success, as all else, to divine benevolence. they're calling it the most ml paper of all time

crumb ⏩⏩⏩⏩⏩⏩ reposted

eliza🌻
@elizas.website

We take it for granted today, but a single 6W embedded NPU can do more tensor operations per second than a peasant in the 1400s could perform in his whole lifetime