xjdr
@_xjdr
ptx enjoyer
Writing jitted jax code is like playing Dark Souls but in python
before i invest as an LP i'd love to know what your firm's DD processes and policies are? ummm, some twitter trolls tell us if the vibes are fire or cap
i believe that B200s are widely available despite export restrictions. i very much do not believe GB200 NVL 72s are available.
training futuristic superintelligence using HPC management software written 2002
with the exception of like 20 people, i simply do not believe that all of a sudden all of you are working on "continuous learning" and "novel agent environments" . smh
👀👀
Apparently Dion is now being worked on for Torch Titan: github.com/pytorch/torcht… :-)
this actually made me lol but - when it comes to quantization, the comparison isn't apples-to-apples with GPUs. we have an approach called truepoint that uses mixed-precision storage but maintains mathematically lossless accumulation in HW during compute. diff architectures,…
this actually made me lol but - when it comes to quantization, the comparison isn't apples-to-apples with GPUs. we have an approach called truepoint that uses mixed-precision storage but maintains mathematically lossless accumulation in HW during compute. diff architectures,…
What are the current best practices (repos?) for using Megatron-Core for large scale training? Trying to repro something and any time saved beating my head against the wall would be greatly appreciated.
i feel the same
It’s truly a privilege to be able to wake up every morning, see where the latest intelligence frontier is, and help push it a little further.
i feel like this is the proper framing for the upside take from todays OAI announcement. "This new approach has yielded impressive improvements wrt the IMO problem set and is likely to further generalize which is very exciting." i endorse this take
this IMO gold will fly past us as quickly as the turing test did soon normies will say “duh of course they’re good at math, they’re computers” but the RL breakthroughs the team made to solve math (congrats!!) will likely generalize to environments with much higher direct value
This is still the announcement I am most excited about in the past week
Engineers spend 70% of their time understanding code, not writing it. That’s why we built Asimov at @reflection_ai. The best-in-class code research agent, built for teams and organizations.
oooo formalization and verifiers about to become so hot right now (for the unwashed masses). looking forward to the all the unhinged takes (from the unwashed masses)
I've been really excited about this team and this launch for a while now. Really looking forward to getting my hands on it!
Engineers spend 70% of their time understanding code, not writing it. That’s why we built Asimov at @reflection_ai. The best-in-class code research agent, built for teams and organizations.