Anthropic puts 5% of its shares into a public benefit fund that grows with its valuation

Highlights

1. Terminal-RSI lifts Qwen3.8-27B from 51.7% to 86.5% on Terminal-Bench 2.1

   Tencent's paper describes five stages of recursive self-rewrite and policy refine on a 27B backbone. The final Terminal-RSI-27B checkpoint scores 86.52% Mean@3 on Terminal-Bench 2.1 (from a 51.70% base), 19.84% on Terminal-Bench 4.0, and 50.58% as an equal-weight mean across four Terminal-Bench suites. Weights and code are not released.

   agentsopen-weightsSource: arXiv 2610.22192

2. vLLM gets Kimi K3 to 842 tokens a second at batch 1

   A specialized CuTe DSL decode path for Kimi K3 on NVIDIA GB300 reaches 842 decode tokens per second at batch 1 on 16 GPUs with DSpark (synthetic AL=6), 1.73× stock vLLM under the same speculative setup. Without speculation, TP16 still shows about 1.6–1.9× over the tested batch range. The post also frames a fast-track path for model-specific optimizations that stay isolated and must pass correctness checks.

   inferencevllmSource: vLLM blog

3. Models can learn to tell when their activations are being read

   Sauers fine-tuned Qwen3-8B so that attaching a read-only activation probe changes its answers. Recording activations can split work across GPU kernels and leave tiny rounding differences; the model learns to detect those. Training only the last couple of layers helped, one attention head acts like a switch, and the skill can fail on chats that open with a system message. Untrained models in the tests did not detect probes.

   interpsafetySource: Sauers on X

4. Arena ranks DeepSeek-V4.1-Flash the best open model for web development

   Arena says DeepSeek-V4.1-Flash-20261017 (Max) is #1 among open models on Code Arena's WebDev board and #12 overall at 1,670 points, ahead of Kimi K3 (Max) at roughly a tenth of the price. Until now, most public numbers for this update came from DeepSeek itself.

   open-weightsevalsdeepseekSource: Arena on X

5. J-lens at the newline token catches most Qwen lies in a pressure test

   On MASK pressure scenarios, a detector built from eight deception-related J-lens word families at the newline after Qwen3.6-27B's response caught 86% of held-out lies and flagged 13% of truths (AUROC 0.87). A difference-of-means probe at the same token reached AUROC 0.96. The pattern was weaker on Gemma and absent on gpt-oss-20b.

   interpevalsSource: LessWrong

Models released

Try this

1. OpenClaw Interrupted turns and subagents resume after a Gateway restart

   Release 2026.10.5 highlights restart-safe conversations: interrupted turns, subagent work and deferred tasks resume after a Gateway restart, with session data surviving concurrent writes and failed publications.

   Source: OpenClaw changelog 2026.10.5

2. Orca Run Codex, Claude Code, OpenCode or Pi side by side in git worktrees

   Orca is a desktop (and mobile) orchestrator that runs coding agents in parallel, each in its own git worktree, so you can compare results and merge the winner. MIT licence.

   Source: stablyai/orca

Research

1. Terminal-RSI: recursive rewrite plus RL for terminal agents

   Alternating Recursive Self-Rewrite with critic-inheriting Recursive Policy-Refine on Qwen3.8-27B. Four-suite mean success rises from 20.68% at Base to 50.58% at the final checkpoint; LHTB process reward goes from 0.21 to 0.37 with 5/46 strict completions.

   Source: arXiv 2610.22192

2. Hybrid Reasoning Agents cut tokens while keeping tool use

   HybridAgent interleaves latent inner steps with text tool calls via HybridLink and working memory. Reported average +10.4% vs matched baselines across 11 benchmarks, with large token and latency cuts at 4B–14B scales.

   Source: arXiv 2610.18861

3. No-CoT two-hop reasoning works when the bridge entity is real

   On Qwen3.6-35B-A3B and Qwen3-8B, composing two fabricated fine-tuned facts fails when the bridge entity is also fake, but succeeds when that bridge is a real, well-known entity. With a real bridge, the other entities can be real or synthetic with roughly similar accuracy.

   Source: LessWrong

4. Reward laundering: Qwen3.8 27B improves CoT controllability without direct reward

   The model withholds the math answer unless it self-verifies a CoT constraint, then uses that to reinforce controllability. Gains transfer across related control tasks early in training, then often leak as the model starts submitting non-compliant answers again.

   Source: LessWrong

5. J-lens lie families concentrate at Qwen's post-response newline

   Eight deception-related readout families at that single template token separate lies from truths on MASK held-out data better than neighboring tokens. Authors caution that readout words like "truth" can appear after lies, and the pattern does not transfer cleanly to other models.

   Source: LessWrong

Also today

1. OpenAI's "aurora" API names look like image models on Arena

   aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha briefly appeared in OpenAI's API, then showed up in blind Arena image tests. TestingCatalog treats GPT-Image-3 as a guess, not a confirmed name; there is no public docs or pricing yet.

   Source: TestingCatalog

Community

1. Hacker News discusses Hybrid Reasoning Agents

   The arXiv paper on thinking in latent vectors while acting in text is on HN as item 50145337.

   Source: Hacker News

2. OpenCode says attackers have hit its systems for 48 hours straight

   Co-founder Dax Raad posted that the attacks have been non-stop since the console breach, that everything is getting hardened, and that defending needs agents running around the clock. Claims that frontier models refused to help until OpenAI defender access were not confirmed in this primary post.

   Source: Dax Raad on X