devansh retweetledi
devansh
1.4K posts
devansh
@devanshpandey
building aligned general learners. cofounder @si_pbc
San Francisco Katılım Mart 2020
652 Takip Edilen1.6K Takipçiler
"1/4 of the compute spent on the final model came from the base, the rest is from our training"
k2 and k2.5 (a continued pretrain of k2) each used 15T tokens
k2 is 32B active; by 6np, kimi did 5.76e+24 flops
if that's 1/4, cursor did 1.7e25 flops i.e. same as gpt-4 flops
Lee Robinson@leerob
Yep, Composer 2 started from an open-source base! We will do full pretraining in the future. Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different. And yes, we are following the license through our inference partner terms.
English
devansh retweetledi
devansh retweetledi
Everyone makes fun of Eight Sleep until it's 84º in SF and they don't have a liquid-cooled mattress. Suckers.
English
nothing gets you feeling the AI <> real world gap more acutely than filing your taxes
English
I just realized that an "order of magnitude estimate" is the same as having zero sig figs
English
devansh retweetledi
Introducing Trellis for Kimi K2 Thinking.
It's post-training code that's 50x faster than the best single-node open-source version and 2x cheaper than training APIs.
After safety testing, we're open-sourcing it, giving builders the best tools to customize a frontier model.
🧵
English
very cool work!!
Workshop Labs@WorkshopLabs
Open weights isn't open training. @AddieF38654 on our team wrote up her experience trying to post-train a 1T parameter MoE model using the existing open source infra. Let's find out how many monkey-patches it takes to post-train an open-weights model. A thread🧵
English
English
@Mjreard @ChanaMessinger the Amerification of oxford is in progress and its dread tides cannot be slowed
English
just witnessed not only public making out, but straight people publicly making out, at the local cafe/workspace/vintage clothes store/second hand bookshop/music venue. this place is cooked and it’s so over
English
devansh retweetledi
So many startups think their engineers are "cracked" but have no idea what that really means.
This team of 5 19yr olds built a 30 petabyte storage cluster in SF for ~$500k to get a 40x cheaper AWS S3 as a side quest to store 90M hours of video.
Now, that's cracked.
English
devansh retweetledi
PI's robot can now make a grilled cheese without burning it. It has thus passed the Alfred Test, a higher bar than the Turing Test, because I still cannot do that reliably.
Physical Intelligence@physical_int
We’ve developed a memory system for our models that provides both short-term visual memory and long-term semantic memory. Our approach allows us to train robots to perform long and complex tasks, like cleaning up a kitchen or preparing a grilled cheese sandwich from scratch 👇
English
devansh retweetledi
Standard Intelligence's @devanshpandey responds to @tszzl's tweet that "text is the universal interface," and explains why their new foundation model is trained on video:
"At some point in the arbitrarily long future, if we only use text models, we could force most things to be text. But I think there are just a lot of things that are much more native when done from a computer-use [perspective]."
"GUIs are designed for humans to use. We have this massive long tail of things on the internet that are entirely undoable by LLMs."
"For example, when I do ML engineering most of my time is spent doing the grunt work of engineering. It's a lot of looking at graphs, analyzing, and comparing loss curves. You can do this in text, but it's a much larger pain than doing it in the native interface."
"There's a reason humans don't interact with a computer purely through text, it would kind of suck."
roon@tszzl
text is the universal interface
English
devansh retweetledi
Happy Tuesday. Big show today:
- @patrickc (Stripe)
- @collision (Stripe)
- @bgurley (Benchmark)
- @ivanhzhao (Notion)
- @StefanoErmon (Inception AI)
- @thejamescad (Profound)
- @RuneKvist (AIUC)
- @reinerpope (MatX)
- @devanshpandey (SI)
See you on the stream.
TBPN@tbpn
English
Extremely impressive work, especially for a v1. I’m curious how v2 would be integrating test-time adaptive computation though. I know you guys were exploring latent recursion briefly (UTs styled). Are you still working on that or are you pivoting to more CoT based discrete approaches for the future?
English
there is much work to do. in the next few months we need to train 100x larger models on 200x as much data as we have done before, scale RL to have millions of agents learning things from the internet, and use the computer use prior to build the first models that can scalably learn. there's no team i'd rather be doing this with. if you're interested in being a part of SI's next chapter, i'd be very excited to chat.
Standard Intelligence@si_pbc
Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.
English
shout out to the trusty heap for storing all the video data needed to train FDM-1. tbh it might have more uptime than AWS in the past year 😛
Standard Intelligence@si_pbc
our S3 bill was way too expensive… so we built a 30 PB storage cluster in the heart of SF.
English
Computer use models shouldn't learn from screenshots.
We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.
GIF
English
devansh retweetledi
devansh retweetledi
feels like a pivotal moment for realtime
Standard Intelligence@si_pbc
Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.
English
Yep - OSworld and similar benchmarks tend to be particularly optimized for llm harnesses, instead of trying to capture real-world computer use.
Also the model isn't instruct-tuned yet etc, so lots to scale! The benchmark we're most excited about long-term is just our model being actually useful in the real world, and we can just test on stuff like CAD and e.g. universal tab model for computers for that.
English
This is super super cool. Love the idea of going straight to video instead of VLM + screenshot + a11y tree.
Curious if you happened to test it on OSWorld/WebArena, etc.? Those benchmarks might be shallow for this, as you could maybe just go straight to games/CAD and test there.
English