Media Zone | 2026-07-20

Monday's signal was the DAIR.AI weekly roundup, not the timeline. Routing meaningfulness and agent-failure-attribution dominated; Twitter's AI handles were quiet.

Today's signal

• Dominant story: DeepMind's "When Is Routing Meaningful?" reframes how routers should be judged.
• Pattern: agent evaluation shifting from pass/fail outcome to traceable trajectory (OAT, Failure-as-a-Process).
• Counter-signal: harness evolution "does not consistently outperform simple test-time scaling."
• Quiet area: no AI-handle Twitter signal, Reddit empty, Kurate unchanged (biomedical).

Routing, KV cache, compression, GPU

Routing gets both halves of its foundation

• DeepMind: a router can be accurate and still be a no-op (vacuous pool or unstable assignments).
• Two diagnostics: Hierarchic Social Entropy (pool diversity), perturbation robustness (stability).
• Fewer than 10 curated models recover most of a large pool's diversity.
• IBM: routing is system optimization, not classification. Caching beats sticker price (Sonnet $79 vs GPT-4.1 $155).
• The pair: DeepMind fixes the test, IBM fixes the objective. Neither router exists yet.

LLMs, agents, safety

Agent failure becomes traceable

• OAT (Microsoft + UW-Madison): label-free failure attribution from 100 success trajectories, 200-5000x faster than prompting.
• Failure-as-a-Process: 63,000 annotated steps; every failure has 3 timestamps (decisive error, point of no return, first observable).
• Harness Evolution Rethought: evolved harnesses (67.4) underperform plain parallel sampling (72.3) on Terminal-Bench 2.1.
• Two new surveys: self-improving agents, and metacognition in LLMs.