Summary
The morning slot carried no reposts or AI-handle tweets, so this synthesis reads the morning capture of the Following feed. The strongest technical signal is a cluster of three posts on agent memory and harness evolution: Anthropic's "components of agent memory" post was echoed by a memory startup, and omarsar0 pointed at trajectory-driven harness evolution. A second cluster of two posts covers distillation: TRL's multi-teacher on-policy distillation release and the Salesforce paper pricing distillation against RL. Garry Tan started publishing open harness evals, which drew pushback. The rest of the feed was dominated by the Jensen Huang affair, which was loud and carried no artifacts. Standalone items worth a look are NVIDIA's Dynamo recipes for Korean open models and a Tsinghua/Moonshot result that GPT-6 Astra proves 70% of research-level CS theorems in Lean but fails a harder subset.
Posts
- Agent memory and harness evolution (cluster of 3) (@danizhu, @omarsar0, Anthropic post). Anthropic's post argues memory helps only when the model closes the loop: write a theory, test it, store validated rules. Offline "dreaming" consolidates raw logs between sessions. EverOS's founder says their Reflection step follows the same principle. omarsar0 highlights Evolver (arXiv 2611.07595), which stores reusable procedural "Gene" assets mined from past trajectories for Claude Code and Codex: success rises from 83.13% to 87.50% and tokens fall 24.24%. See the memory summary and the EnvOpt and Evolver page.
- Distillation compute (cluster of 2) (@SergioPaniego, @rohanpaul_ai, arXiv 2611.08453). Hugging Face TRL now supports multi-teacher on-policy distillation (MOPD), which merges several specialist teachers into one student trained on its own outputs. Salesforce's paper explains when this pays: distilling from an outside teacher loses to GRPO (group-based RL) per FLOP, but distilling from the student's own RL checkpoint is 50x cheaper, and merging per-domain RL experts costs only 5-8% extra compute. See the summary.
- Open harness evals (@garrytan, repo). Tan is running public evals of which agent harness completes a task best for the least cost on frontier models. Many replies dispute the method, so treat early rankings as contested.
- NVIDIA Dynamo recipes for Korean open models (@NVIDIAAI). Deployment recipes for SK Telecom's A.X K2, LG's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif-3, claiming up to 2x inference speed on B200.
- TCSExam (@rohanpaul_ai, arXiv 2611.07962). Tsinghua and Moonshot AI find GPT-6 Astra proves 70.0% of research-level computer science theorems in Lean, yet resolves 0 of 30 open problems in a harder split. Opus 5.5 leads the Lite subset at 72.3.
- Follow the Policy, Fail the Benchmark (@rohanpaul_ai, arXiv 2611.07042). An audit of policy-based decision benchmarks. Fixing the scorer to follow the stated policy raises acceptance on LegalBench SARA by more than 80 points, so some "model failures" are benchmark failures.
- Jensen Huang affair (cluster of 3) (@dylan522p, @tatthang, @edzitron). Reports tie Huang to Alice Huang, alleged to have links to Chinese intelligence and chip trafficking. Dylan Patel predicts nothing happens to Huang in six months beyond a congressional summons. The rest is snark. See the digest Industry Pulse for the board investigation.
- GitHub Copilot app sessions in VS Code (@github). An agent session started in the Copilot app can be finished in the IDE with task, plan and conversation carried over. Product update.
- Academic Research Skills for Claude Code (@Fred834567). An open-source skill pack covering literature search, gap finding, citation checks, peer review and formatting. Skip unless you want the workflow.
- Skip: Grok Bot guide promos, "reverse engineer anything" repo lists, floating-solar H100 and NVIDIA AI Day Seoul event posts, a Yann LeCun political retweet.