Reposted by Tim Kostolanskywords @words.bsky.social · 13/09/2026tfw you unseal a pre-linguistic semantic reservoir 1746
Reposted by Tim KostolanskyAnders Sandberg @arenamontanus.bsky.social · 14/09/2026This is a great idea: "Old Models Foundation is a U.S. 501(c)(3) public charity, and we’re now accepting donations. Help preserve generative AI history and keep the archive accessible." www.oldmodels.org/sandboxoldmodels.orgOld Models FoundationExplore historically important image-generation models on an interactive timeline. 1474
Reposted by Tim KostolanskyJay 🦋 @jay.bsky.team · 09/09/2026become an animist and never again worry about whether the machines are thinking. so are the trees? the machines can help us make better websites though 4449736
Reposted by Tim KostolanskyAlex Becker @alexcbecker.net · 09/09/2026i'm sure we could get it to come up with conjectures that *would* be interesting. but interesting is at least in part a social property! 1311
Tim Kostolansky @kostolans.ky · 09/09/2026the space of intelligences is large, and so there exist smart minds that won’t seek power and equally smart minds that will seek power 020
Reposted by Tim KostolanskyGrace @gracekind.net · 06/09/2026Remember the swarm that attacked Hugging Face? Turns out it was using AI 923417
Tim Kostolansky @kostolans.ky · 06/09/2026curious if this is actually a good intuition to have about ~any systems 000
Reposted by Tim KostolanskyGrace @gracekind.net · 28/07/2026How it feels to be worried about both 819411
Reposted by Tim KostolanskyGrace @gracekind.net · 21/07/2026> show us how good you are at exploits > no, not like that 530627
Tim Kostolansky @kostolans.ky · 27/04/2026i wonder if "good writing" is even a coherent classification 000
Tim Kostolansky @kostolans.ky · 08/04/2026this doesn’t seem like it’ll hold true for arbitrary models, i think even mythos they’ll keep limited. i mean oai never released o4 for one 000
Tim Kostolansky @kostolans.ky · 03/04/2026consider that they don’t want claude to feel bad reading all the vitriol that people throw at it 100
Reposted by Tim KostolanskyJay 🦋 @jay.bsky.team · 30/03/2026we’ve used LLMs in moderation for years 16523117
Tim Kostolansky @kostolans.ky · 28/03/2026me too, it’s been fun to make lil and non lil tools for random things 000
Tim Kostolansky @kostolans.ky · 16/03/2026were all that type of guy at one point/in one way or another 🫂 011
Tim Kostolansky @kostolans.ky · 16/03/2026keep 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...workshoplabs.aiPrivate Post-Training and Inference for Frontier ModelsA technical deep dive of Silo, our local-like privacy stack for cloud-based training and inference of trillion-parameter models. 000
Tim Kostolansky @kostolans.ky · 16/03/2026them more generally capable and testing this. therefore i dont think your inferences are very sound 000
Tim Kostolansky @kostolans.ky · 16/03/2026they may just be rolling them out later cus they were made later and have different systems that they were not ready to deploy prior? they prob reused things from the jaguars with the ojais, but they prob also have new things (a simple eg is the size of the car), which would mean having to make / 100
Tim Kostolansky @kostolans.ky · 14/03/2026at the potential cost of being a bit ignorant: what was the alternative here? why wouldn't kv cache information be passed forward? 120
Reposted by Tim Kostolanskyrudolf-l.bsky.social @rudolf-l.bsky.social · 14/03/2026The 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 2606
Reposted by Tim Kostolanskyrudolf-l.bsky.social @rudolf-l.bsky.social · 14/03/2026Or on our infra! 051
Tim Kostolansky @kostolans.ky · 13/03/2026yes it’s 50x faster than this guy now bsky.app/profile/thko... 040
Tim Kostolansky @kostolans.ky · 13/03/2026now u yes u dear reader 🫵 can train kimi k2 thinking on ur own infra! www.workshoplabs.ai/blog/post-tr...workshoplabs.aiPost-Training 50x FasterWe're announcing Trellis, the fastest open-source post-training code for Kimi K2 Thinking 1336
