Nevasini Sasikumar

• founding ml engineer at prev microsoft, adobe, intel, mbzuai, decompute
• masters @ucsandiego, bachelors @pesuniversity

[Nevasini Sasikumar @nevasini1]

@thomastjoshi @huggingface @OpenAI Interested!

[Nevasini Sasikumar @nevasini1]

@bdono_ @mimicrobotics @hbr_pbc @kstonekuan That's about 11 sec/episode. Does ~2 hours still work when tasks get longer and messier?

[Nevasini Sasikumar @nevasini1]

@silennai @explabsai @OKfallah Congrats! Loved the world-model-as-a-harness, curious where prompt optimization tops out vs. training a world model like DreamGym does?

[Nevasini Sasikumar @nevasini1]

@AryaTschand Thoughtful!

[Nevasini Sasikumar @nevasini1]

@khalio "Data is a desire problem" love that line!

[Nevasini Sasikumar @nevasini1]

@samaneggs Love this, most setups just bury failed runs in a summary. Curious though, how do you tell when two attempts are the same dead end vs actually a new idea?

[Nevasini Sasikumar @nevasini1]

@antonioalst @ComputeDesk Good thought!

[Nevasini Sasikumar @nevasini1]

@ComputeDesk Love this 🔥
Such an obvious pain point in hindsight. Describing a cluster in ten seconds and getting a shareable link is genius!

[Nevasini Sasikumar @nevasini1]

@tkkong would love to try it out!

[Nevasini Sasikumar @nevasini1]

@spisallyouneed Love this. Resets have been quietly carrying RL benchmarks for years. What's the longest horizon an agent has to survive in the current suite? Will be playing around with the environment hub this week.

[Nevasini Sasikumar @nevasini1]

@skyfallai Curious how you separated context overflow from pre-training gap in the detection lag analysis

[Nevasini Sasikumar @nevasini1]

@gpuemi @wafer_ai Used it for a multi-hour coding session and TTFT stayed snappy even at long context, what’s keeping prefill off the decode path, disaggregated serving?

[Nevasini Sasikumar @nevasini1]

@charles_irl Need for computer of computers benchmark

[Nevasini Sasikumar @nevasini1]

@aarnav2006 @cactuscompute @awscloud @Siemens @BCG @Fetch_ai I am interested in this!

[Nevasini Sasikumar @nevasini1]

@benln I am interested in this!

[Nevasini Sasikumar @nevasini1]

Software factories: while True: with a pitch deck.

[Nevasini Sasikumar @nevasini1]

@proceduralia This matches everything I've hit building these systems. The apps are the easy part; the hard part is reliable resets, checking against ground truth, and enough variety to actually learn from. Love the PRISM framing!

[Nevasini Sasikumar @nevasini1]

@u1tra_instinct consumer-frontier blueprint. rooting hard for phase 2 🔥

[Nevasini Sasikumar @nevasini1]

@_xjdr So on nvfp4, nvidia's own numbers show it's already as good as fp8. it's actually ahead on long-context recall and tool-use, tied just about everywhere else, with only a tiny dip on one coding test and that's with no extra tuning at all.

https://t.co/MHzQ1Xg6Vv

[Nevasini Sasikumar @nevasini1]

@Ex0byt DeepSeek-V4 ships an MTP head and if you turn on speculative decoding in SGLang for you might get another ~1.5–1.8×