Ali Naeimi

AI Research Engineer | Protein and Genome LM | Perfetto enjoyer | Distributed Pretraining optimization

[Ali Naeimi @Ali_NT99]

@aether_oracle @teortaxesTex GPT models perform really poorly in BSBench so you kinda answered yourself

[Ali Naeimi @Ali_NT99]

@mattpocockuk opencode go provider, using oc-go-cc to route to claude code

[Ali Naeimi @Ali_NT99]

If you're on Grace Hopper/Blackwell and not fully utilizing the C2C BW for targetted offloading, you're ngmi...

[Ali Naeimi @Ali_NT99]

@kalomaze the small gap between DSV4 flash and pro really makes you wonder what went wrong with the DSV4 pro training... is it just undercooked or is the arch inherently doesn't scale well...

[Ali Naeimi @Ali_NT99]

@kalomaze I'm curious about gpt 5.2... I think 5.2 was a bigger model than 5.4... I always felt like it had more knowledge in nicher areas

[Ali Naeimi @Ali_NT99]

@teortaxesTex @spicey_lemonade or a catastrophic forgetting issue?

[Ali Naeimi @Ali_NT99]

@llmdevguy kimi k2.7 should be compared to composer 2.5 since they share a base model and differ in post-training... I'm skeptical that 2.7 can leapfrog composer 2.5...

[Ali Naeimi @Ali_NT99]

@antirez JSC Jupiter says hi

[Ali Naeimi @Ali_NT99]

@elliotarledge what model are you running? q3.6 27b?

[Ali Naeimi @Ali_NT99]

@SzymonOzog_ haha :)) I have to admit my kernels were optimized on sm120 and were performing at 25% SOL on sm90, but opus 4.7 was only able to push it to around 40% before... Mythos reached around 70%

[Ali Naeimi @Ali_NT99]

@paradite_ Bold strategy to stand out from other anthropic applicants... let's see if it works out for him.

[Ali Naeimi @Ali_NT99]

@dogacel0 It's clear that Mythos was RLd extensively on kernel generation using various new DSLs even... I made it go over my triton kernel and suggested rewriting them in CuteDSL which was super interesting as other LLMs usually don't even know about CuteDSL. then it made them 3x faster

[Ali Naeimi @Ali_NT99]

@juliarturc @atmoio I think the same goal of "finding meaning" can be achieved by going down at least one level of abstraction, without the need to change careers.

[Ali Naeimi @Ali_NT99]

@Memetic_Theory If they can get Karpathy, they can get anyone.

[Ali Naeimi @Ali_NT99]

@xeophon Crazy that anthropic researchers can't come up with a better optimizer than Muon

[Ali Naeimi @Ali_NT99]

@cognition Good stuff! hopefully the team will be fast in covering other models... would be nice to see DSV4, GLM5.1, Qwen3.7 Max, Composer 2.5, Gemini Flash 3.5, M3 etc. and also the SOTA local models like Gemma 4 31B and Qwen 3.6 27B

[Ali Naeimi @Ali_NT99]

@osieberling Congrats on the release! So this is basically canon_layers with dynamic kernel size right?

[Ali Naeimi @Ali_NT99]

@thegenioo @datacurve The only model that sort of approaches 5.5 or 4.8 for my tasks (mostly ML and perf engineering) has been Qwen 3.7 Max, but it's still too expensive and slow. 5.5 remains undefeated, can't wait for 5.6

[Ali Naeimi @Ali_NT99]

@htihle Any plans to test qwen 3.7 max? It’s been by far the best chinese model for ML tasks in my experience. Although it’s very expensive per task even compared to 5.5 or 4.8