Summary
Only the morning slot ran today, and it had no reposts, so the whole day's social signal comes from the Following feed. The strongest cluster is agent memory and harness evolution: Anthropic's "components of agent memory" post argues memory pays only when the agent writes, tests and keeps validated rules, and Evolver backs this with numbers (success up to 87.5%, tokens down 24%). The second cluster is the most useful for efficiency work. TRL shipped multi-teacher on-policy distillation, and a Salesforce paper prices it: self-distilling from the student's own RL checkpoint is 50x cheaper than GRPO. TCSExam is the standout single item, showing GPT-6 Astra proves 70% of research-level theorems in Lean but solves none of the open problems. The Jensen Huang affair was the loudest story and the lowest signal, with no artifacts behind it.
Posts
- Agent memory and harness evolution (cluster of 3) (@danizhu · @omarsar0 · Anthropic post · arXiv 2611.07595) [morning]. Memory helps only when the agent closes the loop and consolidates logs offline, and EverOS says its design follows the same idea. Evolver mines reusable procedures from past trajectories: success rises 83.1% to 87.5% and tokens fall 24% (memory summary, Evolver page).
- Distillation compute (cluster of 2) (@SergioPaniego · @rohanpaul_ai · arXiv 2611.08453) [morning]. TRL now merges several specialist teachers into one student trained on its own outputs. Salesforce finds outside-teacher distillation loses to GRPO per FLOP, but self-distillation is 50x cheaper and merging domain experts costs only 5-8% extra (summary).
- TCSExam (@rohanpaul_ai · arXiv 2611.07962) [morning]. GPT-6 Astra proves 70.0% of research-level CS theorems in Lean but resolves 0 of 30 open problems. Opus 5.5 leads the Lite subset at 72.3.
- Follow the Policy, Fail the Benchmark (@rohanpaul_ai · arXiv 2611.07042) [morning]. Fixing the scorer to follow the stated policy lifts LegalBench SARA acceptance by more than 80 points. Some "model failures" are benchmark failures.
- Open harness evals (@garrytan · repo) [morning]. Public evals of which harness finishes a task best for the least cost. Replies dispute the method, so treat early rankings as contested.
- NVIDIA Dynamo recipes for Korean open models (@NVIDIAAI) [morning]. Serving recipes for A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3, claiming up to 2x inference speed on B200.
- Jensen Huang affair (cluster of 3) (@dylan522p · @tatthang · @edzitron) [morning]. Reports link Huang to Alice Huang over alleged Chinese intelligence and chip-trafficking ties. Dylan Patel expects nothing beyond a congressional summons; see the digest for the board investigation.
- GitHub Copilot app sessions in VS Code (@github) [morning]. An agent session started in the Copilot app can be finished in the IDE with plan and context carried over. Product update.
- Academic Research Skills for Claude Code (@Fred834567) [morning]. Open-source skill pack for literature search, citation checks and peer review. Skip unless you want the workflow.
- Promo and off-topic [morning]. Grok Bot guides, "reverse engineer anything" repo lists, floating-solar H100 and NVIDIA AI Day Seoul event posts, a LeCun political retweet. Skip.