$ cat briefs/daily/2026-10-13.md
2026-10-13
October 13, 2026 (Tue)
3 stories · 2 paper picks · 3 new pages
Top Stories
1. Anthropic tests replication beyond datacenter hardware
A 9B model sets up inference on 8 of 9 hardware types. In a separate, simplified network experiment, GLM-5.2 reaches 8/8 nodes in 4 of 5 trials. The controlling model runs externally; the target serves a smaller model. Anthropic explicitly says it has no full rogue replicator.
Why it matters: hardware fit and reliable replication remain separate requirements; the experiment measures progress without demonstrating an uncontrolled deployment. → AI-Enabled Cyberattacks (source)
2. Distillation defenses now have measured effects and customer costs
Anthropic reports reasoning-token extraction falling by 10× to 100× on its platform after classifier deployment. Controls include preserved context, transcript-access restrictions and account enforcement. Benign requests can be blocked and selected blocked requests are charged; sustained sophisticated circumvention remains untested.
Why it matters: capability protection changes the API contract for legitimate developers too. → Adversarial Distillation (source)
3. GLM-Legal specializes the model and its execution harness together
Z.ai reports 19.2% all-pass on Harvey LAB, 12.5 percentage points above its base model. Legal mid-training feeds RL and harness refinement using practitioner feedback. AutoClaw access is forthcoming; price, license and release date remain unspecified.
Why it matters: a domain release can train the whole workflow, while its score still depends on a particular evaluation. → GLM-Legal (source)
Paper Picks
Stateless Language Agents — persist evidence outside conversation history. Fresh role-specific contexts and an Advisor directing Workers reach the strongest kernel baseline's final performance with over 84% fewer tokens in the authors' tests. Read it for long-run research design; token savings do not establish universal cost savings. → Stateless Language Agents — durable research state outside conversations (source)
SWE-Duel — generate the task during evaluation. Agents alternate authoring repairable defects and reviewing them. Across 1,500 turns, the best solver and best task author differ. Read it for separating those capabilities; participants are model-harness pairs. → SWE-Duel — coding agents generate and solve each other's tests (source)
Watch
- HSS: human participants score 93.1% on intuitive visual reasoning, versus GPT-6-astra's 53.6% at maximum reasoning effort. The task-specific gap persists with visual manipulation tools. → Eval Harness Configuration (source)
New in Wiki
- GLM-Legal, Stateless Language Agents — durable research state outside conversations and SWE-Duel — coding agents generate and solve each other's tests — specifications, methods and evidence limits. (model) (research) (evaluation)
Updates
- ZCode 3.15.1: command confirmation, browser permissions and per-subagent models. → Z.ai (source)
- Decision 3.0 on Core ML: device-specific latency measurements distinguish public JevBench accuracy from sealed leaderboard scores. → Model Routing (source)