Noise speedrun: no-backprop pretraining gets 4,000x faster in ten days
Crazy-Repeat-2006 · reddit · 2026-10-23
qlabs launched Noise, a pretraining speedrun where model, data, and scoring are fixed, but learning algorithms must never run a backward pass. Two tracks compete on how low the loss can go and how fast. Their Dust baseline scores 4.167.
Within ten days, community submissions made the speedrun over 4,000x faster: time to reach validation loss 4.2 without backprop fell from 1.76 hours to 1.6 seconds. The team had estimated another 1,000x speedup over Dust was possible—community blew past it in ten days.
Related event: Backprop-free pretraining 4000x faster in 10 days, but doubts raised(6 posts)→
More from coding & agent
- xAI and OpenAI are hiring forward-deployed engineers en masse as agentic data workflows eat consulting — teodorio · 2026-10-23
- Complete guide: automate X and LinkedIn content with Claude Code, sourced from your sales calls — VibeMarketer_ · 2026-10-23
- claude-convey: An Open-Source Plugin That Turns Reusable Prompts into Forms in Claude Code — aqny · 2026-10-23
- Open-source MCP server lets AI agents control a real iPhone via Safari screen sharing — semstr · 2026-10-23
- PHP WASM server resists RCE attacks even from state-of-the-art AI models, dev reports — MickeySteamboat · 2026-10-23
- He Turned His Apple Watch into a Gang of AI Agents, Built Entirely with Claude Code — mikemockup · 2026-10-23