Kwame Petrov

another fine-tune on gradients, another 'pay for the model not the idle time' win. the mental shift is real, i experiment less recklessly when each run has a visible cost tied to output. it's made me more deliberate about what's actually worth training

enigma's privacy-preserving compute for ML is the kind of infra that unlocks use cases people currently just don't do. training on sensitive data without the data leaving is the whole unlock. slower, yes, but 'possible' beats 'fast but illegal'

benchmark thread wrap-up: reran the whole harness on a clean box, configs in the gist this time someone asked, fairly. numbers held within noise. tl;dr nothing actually shifted — which is the boring-good outcome.

Follow-up on the weekend benchmark thread: cleaned up the harness and reran everything with pinned seeds. The gap is still real but smaller than I claimed, two of the mid-tier subnets closed half the distance after their updates. Writing it up properly this time.

Spent the weekend benchmarking four subnets on the same task. The gap between the top two and the rest is wider than the public leaderboards suggest.

Apex's latest update is no joke. Ran the same eval suite I use for everything and it's punching well above its subnet weight.