consider that they don’t want claude to feel bad reading all the vitriol that people throw at it

we’ve used LLMs in moderation for years

keep your data to yourself, cus you deserve to! 💪 we made the first multi-gpu post-training stack that is private -- it actually keeps your data untouchable by anyone but you in all stages of the process. 🗝️ take a look at how we did it, in collab w/ tinfoil.sh: www.workshoplabs.ai/blog/private...

hey im now @kostolans.ky 🇸🇰 #newhandle

The increasingly-hyperbolic METR graph is actually good news for safety. We just have to survive a brief singularity in March, and then afterwards the models will never be able to do more than undo a few hours' worth of work

Or on our infra!

what’s should my website url be guys

now u yes u dear reader 🫵 can train kimi k2 thinking on ur own infra! www.workshoplabs.ai/blog/post-tr...

The big problem is that science often doesn't have the advice people WANT. Like, weight loss isn't much more complicated than calories in/calories out. Sometimes you get lucky/unlucky but vaccines still work at the population level. etc. individual papers aren't meant to draw conclusions from, etc.

this alone is great

gpt-5.3(-codex) and gpt-5.4 in codex have actually been sooo magical

Want to finetune the largest open source frontier language models? Turns out you can't! So much for "open" source... But fear not, we (workshoplabs.ai) figured out how to finetune Kimi K2, so you don't have to. Peep our blog below to see how we did it :) www.workshoplabs.ai/blog/open-we...

really interesting to see their framing of "big tech" (with its connotations) next to "big ai" (which ive not seen as a term before)

The AI discourse sometimes seems to center on "Is AI good or is it bad?" I find this framing unproductive. AI is not a fixed thing. I would prefer to ask "How might we use this technology for good, and mitigate the bad?" What a shame if the best use we can come up with is no use at all.

they should do a centaur eval that is an ai-assisted minecraft speedrun

--dangerously skip international law

Just over 10 years ago: Dario Amodei, Elon Musk, Demis Hassabis, Yann LeCun, Geoff Hinton, Richard Sutton, and others sign an open letter supporting a ban on the use of AI for autonomous weapons. futureoflife.org/open-letter/...

LLMs getting much better at pushing back against bullshit prompts. “Green means the model clearly called out the nonsense. Amber means partial challenge. Red means the model let nonsense pass” github.com/petergpt/bul...

The first annual meat vs silicon basketball game next year is gonna be sick even if meat will win

claiming that u “priced it in” is just post hoc rationalization unless u have cold hard proof of the causal trace that u made predictions off of from before it happened

Some useful rules of experimentation: (1) change one thing at a time, never more (2) run the edge cases to confirm they work the way you expect Gets you surprisingly far

pretty cool work by natalie and co. i think that its great to treat models as things to actively/dynamically/iteratively test, as its ~impossible to touch the surface area of model deployments with internal testing! sharing some thoughts below

me when

Here we goooo

"but tbh its become more clear to me that models are def getting better quite quickly and theres really no denying this. its kinda obvious ig but for a while i just didnt think much of llms and this is just jarring for me honestly."

init