Nick Levine

New work with @AlecRad and @DavidDuvenaud: Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text. Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread:

on a research cluster I use you can clearly see utilization spike ahead of conference paper submission deadlines. If a futures market takes off get ready for conference cycle seasonality quant systems (although probably a subtle pattern / drop in the bucket in the grand scheme of things)

What is the best argument for the "compute" market will evolve in such a way that there will be an actively traded market for capacity, rather than just long-term bilateral contracts between inference providers and GPU owners? (Assuming those remain distinct categories)

@TheStalwart @evanjconrad

@evanjconrad got it - thanks!

oh yeah, we removed them because people kept getting confused our order book supports pricing all of these differently: - 100 nodes colocated for 3months starting in 1 week - 100 nodes non-colocated for 3 months starting in 1 week - 100 nodes starting whenever, for 3 months min - 100 nodes starting RIGHT NOW, for 3 months - 100 nodes starting RIGHT NOW, for 1 year sometimes it was flat in early SFC because a seller isn't being non-optimal and the book wasn't liquid enough for someone to take advantage of the arb

we build gpu clusters sometimes we take over other people's clusters like a property manager (just like other clouds do) we let you buy large, long-term contracts that let you sublease to let you sublease, we built an order book we built our cloud around the order book we are not a brokerage, we are not a gpu reseller

@Sirupsen @turbopuffer don't have a use for it yet, but may soon

@status_effects @turbopuffer are you puffin' btw?

turbopuffer tagline should be 'puff around and find out' @turbopuffer

@_chenglou ❤️

We’ll look back at Talkie as one of the newer pieces of AI-native art, alongside HP Belanciaga, golden gate Claude, spiral, pit and the rest

@gwyntel @xlr8harder awesome! we have a discord, wanna dm me and i'll send an invite?

@xlr8harder If we can make a talkie user group, I'd love to join. I have about 15k rows of synthetic data generated from TalkieLM.

I wanted to play with the Talkie 1930 models, but they weren't packaged in a convenient transformers format, so I had codex convert them. They can also now be used with vllm transformers backend. Here they are, in case it's useful to anyone else: huggingface.co/collections/xl…

Now I can do what I wanted to do: play with the base model using few shot chat prompts, a la URIAL style

@TheStalwart 🎵 walking around with your head in the clouds 🎵

@arthurpostingg @MattZeitlin just applied for a delay repay :)

@status_effects @MattZeitlin words typed by a man who has never had the displeasure of taking an avanti west coast service (39.9% percent “on time” between 2024-2025)

Was spending some time on Google Maps and all the non London cities you’ve heard of in England are really close to each other by US standards

talkie is a goblin skeptic.

This one had me on the edge of my seat

@AndyMasley biang biang noodles at xian impression. falafel sandwich at mr falafel if in shepherds bush for some reason

Would appreciate recs for things to do in London, especially where the good vegan food is

@eggsyntax @AlecRad @DavidDuvenaud ❤️ let us know if you end up doing any research with talkie and how we can make it as useful as possible!

@status_effects @AlecRad @DavidDuvenaud (I say thanks especially because I can imagine just how much tedious work must have been involved in this particular project, and you conscientiously slogging through that for months is a real gift to the research community 🙏)

New work with @AlecRad and @DavidDuvenaud: Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text. Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread:

@thsottiaux @sama @gao_zibo Exciting stuff. Are remote ssh sessions on the list?

@gao_zibo All of this and more is coming

codex mac app is winning SO HARD. just need: - native editor - iOS app - full browser - openclaw then it might be the home default app

This, is an excellent finding

@bo_austin_ banger

I love the idea of framing ‘the public learning about historical events not previously discussed’ as ‘effectively invented, in the public imagination’. it’s honestly so fascinating. it’s sort of a new type of being stupid.