left a comment on a 1k line ai generated pr and their agent immediately pared it down to 70 lines
4
0
40
1,808
fun underappreciated side effect of this is that i have given myself synesthesia for a subset of 3 digit numbers
personal update: after almost a year, I'm happy to share that I've mostly memorized all pre-2009 US area codes. Looking forward to what's coming next, including gradually working in post-2009 overlays, becoming marginally better at geoguessr, and being an absolute menace whenever in the conversation at parties
you might ask: why did I do this when you can instantly look up any area code on the internet? well
5
0
52
2,200
better fusions. better gemms. papa driss
Gave a talk at PTC on flex_gemm which is a new gemm_epilogue API in pytorch built on Quack. I think its neat, not ground breaking but neat nonetheless :)
github.com/drisspg/transf…
planning to bootstrap this and continue to add fun examples
1
0
21
843
agent-generated issue comment telling me my agent-generated PR fixes their agent generated issue
2
0
33
1,188
what if you just shot your infra engineers
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image as no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-optim…
💻 Try it on your task: github.com/xingjian-bai/o…
🧵
16
23
696
24,230
9999 instances of claude the exact microsecond i mark a pytorch github issue as “actionable”
3
2
146
3,783
yeah no see they're an rl env company, they just also sell compute and dat---are they an inference provider? of course. which models? well all of the big open models of course, but they also have some of their own. so they also do posttraining? yes. so how do they make money? uh
9
5
195
9,412