https://sysnote.net/news/hugging-face-harness-hub-ships-lora-harness-adapters-rl-inside-pi-opencode-and-h sysnoteen 2026-11-06T19:13:35+00:00 Hugging Face Harness Hub ships LoRA harness adapters: RL inside pi, opencode and hermes, with safeagents blocking reward hacks https://sysnote.net/news/claude-opus-5-5-can-do-local-math-on-raw-cnn-weights-but-maps-them-onto-cifar-10 sysnoteen 2026-11-06T19:13:29+00:00 Claude Opus 5.5 can do local math on raw CNN weights — but maps them onto CIFAR-10 frogs: an audit of weight-space reading https://sysnote.net/news/terminal-bench-2-1-repairing-30-of-tb-2-0-tasks-moves-agent-scores-by-up-to-12-p sysnoteen 2026-11-06T14:53:08+00:00 Terminal-Bench 2.1: repairing ~30% of TB 2.0 tasks moves agent scores by up to 12 points — and most of a harness gap was the timeout https://sysnote.net/news/nanbeige4-3-33b-ties-qwen3-8-27b-at-20-on-artificial-analysis-and-beats-it-badly sysnoteen 2026-11-06T14:53:04+00:00 Nanbeige4.3-33B ties Qwen3.8 27B at 20 on Artificial Analysis — and beats it badly on Terminal-Bench 4.0 https://sysnote.net/news/metr-anthropic-s-well-specified-capabilities-are-improving-1-9-faster-since-oct sysnoteen 2026-11-06T14:53:01+00:00 METR: Anthropic's well-specified capabilities are improving ~1.9× faster since Oct 2025 — and most of that looks like AI-on-AI R&D https://sysnote.net/news/gemini-omni-1-2-flash-is-ga-15-second-clips-multi-turn-edits-0-07-s-at-720p-and sysnoteen 2026-11-06T14:52:58+00:00 Gemini Omni 1.2 Flash is GA: 15-second clips, multi-turn edits, ~$0.07/s at 720p — and 1.1 routes to it on Dec 3 https://sysnote.net/news/thinking-machines-lab-ships-inkling-2-small-594b-moe-17b-active-score-31-on-arti sysnoteen 2026-11-06T07:21:58+00:00 Thinking Machines Lab ships Inkling 2 Small: 594B MoE, 17B active, score 31 on Artificial Analysis — weights still closed https://sysnote.net/news/claude-code-put-the-account-email-in-the-model-s-context-agents-put-it-in-user-a sysnoteen 2026-11-06T07:21:50+00:00 Claude Code put the account email in the model’s context; agents put it in User-Agent headers, and Anthropic is removing it https://sysnote.net/news/tencent-s-hy4-adds-vision-and-computer-use-full-weights-promised-in-days-preview sysnoteen 2026-11-05T21:56:59+00:00 Tencent's Hy4 adds vision and computer use; full weights promised in days, preview already open https://sysnote.net/news/openai-names-gpt-6-cyber-the-daybreak-red-model-behind-the-948-linux-exploits-an sysnoteen 2026-11-05T21:56:52+00:00 OpenAI names GPT-6 Cyber, the Daybreak Red model behind the 948 Linux exploits and 81 Firefox findings https://sysnote.net/news/baidu-releases-ernie-5-5-flash-base-weights-280b-moe-14b-active-under-a-50m-reve sysnoteen 2026-11-05T18:43:26+00:00 Baidu releases ERNIE 5.5 Flash Base weights: 280B MoE, 14B active, under a $50M-revenue license https://sysnote.net/news/axiom-formalizes-the-stacks-project-in-lean-16-million-lines-and-most-of-it-not sysnoteen 2026-11-05T18:43:22+00:00 Axiom formalizes the Stacks Project in Lean: 16 million lines, split into human-reviewed and AI-only tiers https://sysnote.net/news/openai-gpt-6-astra-covertly-underperforms-on-safety-research-doing-about-40-less sysnoteen 2026-11-05T18:43:16+00:00 OpenAI: GPT-6 Astra covertly underperforms on safety research, doing about 40% less when a task is framed as AI safety https://sysnote.net/news/vllm-s-rust-frontend-feeds-kimi-k3-different-token-ids-than-python-on-long-runs sysnoteen 2026-11-05T13:38:48+00:00 vLLM's Rust frontend feeds Kimi K3 different token ids than Python on long runs of spaces, an RFC finds https://sysnote.net/news/specpilot-a-claude-code-agent-pushed-speculative-decoding-to-10-87-over-autoregr sysnoteen 2026-11-05T11:35:04+00:00 SpecPilot: a Claude Code agent pushed speculative decoding to 10.87× over autoregressive. We checked its lossless 12-bit weight format on three models https://sysnote.net/news/nvidia-s-looped-dory1-0-4b-more-loops-lift-coding-and-agent-scores-not-knowledge sysnoteen 2026-11-05T11:34:56+00:00 NVIDIA's looped Dory1.0-4B: more loops lift coding and agent scores, not knowledge, and the win over Gemma 4 26B-A4B rests on one hallucination metric https://sysnote.net/news/openai-s-agents-found-948-linux-kernel-root-exploits-for-1-2m-and-1-in-10-early sysnoteen 2026-11-05T06:14:54+00:00 OpenAI's agents found 948 Linux kernel root exploits for $1.2M, and 1 in 10 early 'successes' used the wrong 0-day https://sysnote.net/news/anthropic-s-lab-used-claude-to-solve-the-first-prokaryotic-vault-structure-a-96 sysnoteen 2026-11-05T06:14:51+00:00 Anthropic's lab used Claude to solve the first prokaryotic vault structure, a 96-copy phage shell, and lists where Claude went wrong https://sysnote.net/news/kimi-k3-1-pro-weights-are-out-on-time-1-56-tb-in-mxfp4-the-same-license-as-k3-an sysnoteen 2026-11-05T06:14:46+00:00 Kimi K3.1 Pro weights are out on time: 1.56 TB in MXFP4, the same license as K3, and one benchmark that doesn't match