jellybean ❄️
@jdchawla29
Followers
783
Following
26K
Media
819
Statuses
6K
aspiring clown | watching patterns unfold into agi | 24
sf 🌉
Joined September 2017
even in ancient greece, shape rotators had more rizz than wordcels
59
214
4K
If all of GL_n was abelian, every symmetric matrix will be diagonal matrix. (A = Q D Q^T = Q Q^T D = D) Entire notion of correlation will disappear. Everything will be independent. There will be no 'us'. Relationship between you and I will mean nothing. Remember, you are
2
4
83
i do not “commit early and often”, i dont “utilize branches”, i dont have a “cicd pipeline”, once my imagined functionality seems to exist i deploy the local cloudflare environment STRAIGHT TO PROD, and if it breaks, I KILL MYSELF
4
4
31
i didn't even know HMMT was even something they used for llm evals until just now and it's already saturated.
We’ve released an early preview of Qwen3-Max-Thinking—an intermediate checkpoint still in training. Even at this stage, when augmented with tool use and scaled test-time compute, it achieves 100% on challenging reasoning benchmarks like AIME 2025 and HMMT. You can try the
1
0
5
We looked at OSWorld, a popular evaluation of AI computer use capabilities. Our findings: tasks are simple, many don't require GUIs, and success often hinges on interpreting ambiguous instructions. The benchmark is also not stable over time. See thread for details!
9
13
154
—"I can't believe bf16 numerics have held back RL progress for so long" —"I disagree, the reward collapse results seem too pessimistic. Are you going to convert your models to fp16 AMP to check if your runs improve?" —"No, are you?" —"No."
11
17
406
FP16 can have a smaller training-inference gap compared to BFloat16, thus fits better for RL. Even the difference between RL algorithms vanishes once FP16 is adopted. Surprising!
31
126
1K
please bro you just need a foob.json in repo root I swear. please bro just a yorp.yaml this is the last one just a snorp.json and snorp.lock bro please
2
3
36
in another life i would love to get stuck in an elevator with you
0
5
48
Working on llm RL is one of the most intellectually satisfying things I ever done, both from a system and ml perspective
16
13
410
spend enough time in the space and you'll eventually end up writing an agent framework and a training library
2
0
18