Issue #144 2 min read
AI Engineering Signal #144
Top AI models fail at inventing new AI research
Signals
Top AI models fail at inventing new AI research
autonomous research agents need evaluation gates before touching architecture searches.
Web
Sakana demonstrates recursive self-improvement loop
independently reproduce loop-level evaluation before fielding autonomous optimizers.
Web
Deno team joins Cloudflare
existing Deno Deploy workloads need migration planning, not new builds.
Simon Willison
Possible Zod 4.6.6 supply-chain compromise
audit lockfiles and pin Zod to last known good release.
GitHub
Qwen 27B hits 140 tokens/sec on one RTX 3090
single-GPU serving for 27B models now worth re-benchmarking.
AlphaFold did not solve protein folding
treat structural predictions as validation-grade, not proof-grade, in bio pipelines.
Latent Space
The Take
Autonomous optimization claims now outpace independent evaluation — pin dependencies, isolate research agents, and demand loop-level replication before production trust.
Subscribe
Related Signals