Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design
SonglinYang4 · x · 2026-11-11
Interactive & Active AI ran an auto-research experiment: Claude Opus 5.5 designed and screened Transformer architectures on its own under NextTransformer Arena v1 rules, with humans only approving plans and budgets.
Setup
- Data, training length (46,875 updates), and per-token compute (FLOPs within ±1%) locked; only architecture could change
- Each candidate trained just 2,000 updates (4% of a full run); a full run costs 1,400+ GPU-hours
- Only training loss and dev BPB available to the search
Key findings
- Rankings after 2,000 steps tracked full runs closely (Spearman ρ = 0.92), validated on 11 architectures with known results
- Across three rounds and 12 candidates, Opus found AttnRes-Canon: Full AttnRes combined with Canon layers at four positions
- After full training (49.152B tokens): CORE score 35.99, +2.0 over the starting point Block AttnRes, validation BPB 0.0024 lower—first place on both published leaderboard metrics
The author also speculates: if a 2k-step sample can predict a 50B-token run, does this kind of auto-research prediction itself follow a scaling law?
More from coding & agent
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11
- 'Adam can now do CAE': AI agent showcase in engineering simulation draws buzz — burhop · 2026-11-11
- Let your agent scan DB analytics to find missing indexes — DanielLockyer · 2026-11-11