Nebius cuts agent training batch collection time by 66.9%, from ~10 min to just over 3
demian_ai · x · 2026-09-29
Nebius engineers argue the bottleneck in agent training isn't token generation but dependency installs, test runs and waiting on tool results. They reworked task routing, sandbox execution and scheduling — including a key idea: fold in groups of completed attempts as they become ready, so unrelated slow tasks don't stall the whole batch.
In a controlled Terminal Bench-based test with model weights fixed, batch collection time dropped from roughly 10 minutes to just over 3 — a 66.9% reduction. The takeaway: the engineering around the model determines how much of its speed you actually get.
More from coding & agent
- How a Telugu voice note becomes Hindsight agent memory — write-side design — Jeevana_Nanepalli · 2026-09-29
- Curia: open-source tool runs Claude Code agents as a rule-bound society of named seats — bobo-the-merciful · 2026-09-29
- 10 paper-trading agents show parallel calls can bypass spend limits — where should the hard cap live? — Accomplished_Fun_408 · 2026-09-29
- ROFT: fine-tuning only on self-explanations matches GRPO on SWE-bench without RL — iScienceLuvr · 2026-09-29
- Eric Elliott resurfaces Leanpub podcast on AI Driven Development, consciousness and economics — ericelliott_ · 2026-09-29
- Anthropic maps multiagent system risks; researcher likens it to sociology, not chemistry — mattturck · 2026-09-29