https://sysnote.net/news/spark-x2-6-1-7b-s-one-row-loop-fix-checks-out-byte-for-byte-its-92-9-is-measured sysnoteen 2026-11-14T21:37:50+00:00 Spark-X2.6-1.7B's one-row loop fix checks out byte for byte; its 92.9% is measured on the four prompts it was fitted to https://sysnote.net/news/ibm-s-granite-llmon-models-fence-off-untrusted-data-with-reserved-tokens-their-c sysnoteen 2026-11-14T21:37:47+00:00 IBM's Granite LLMON models fence off untrusted data with reserved tokens. Their chat template lets the data type them https://sysnote.net/news/claude-duet-invites-come-with-free-usage-limit-resets-three-for-pro-one-for-max sysnoteen 2026-11-14T19:22:24+00:00 Claude Duet invites come with free usage-limit resets: three for Pro, one for Max and Ultra https://sysnote.net/news/google-s-teamwork-math-claims-checked-99-5-of-the-50-million-lean-lines-are-cert sysnoteen 2026-11-14T19:22:20+00:00 Google's /teamwork math claims, checked: 99.5% of the 50 million Lean lines are certificates, and the Weil formalization isn't public https://sysnote.net/news/agent-initiative-benchmark-all-ten-of-gemini-4-argon-s-correct-silences-were-run sysnoteen 2026-11-14T16:59:37+00:00 Agent Initiative Benchmark: all ten of Gemini 4 Argon's "correct silences" were runs that hit the 180-second clock https://sysnote.net/news/claude-code-s-code-mode-claude-writes-javascript-to-call-its-tools-and-auto-mode sysnoteen 2026-11-14T16:59:30+00:00 Claude Code's Code Mode: Claude writes JavaScript to call its tools, and auto mode reviews the script, not each call https://sysnote.net/news/600-experiments-on-lewm-the-jepa-world-model-predicts-push-t-about-as-well-as-th sysnoteen 2026-11-14T16:59:23+00:00 600 experiments on LeWM: the JEPA world model predicts Push-T about as well as the simulator, and loses its points in planning https://sysnote.net/news/claude-code-s-continue-automatically-at-usage-limit-is-on-by-default-and-an-aban sysnoteen 2026-11-14T16:59:18+00:00 Claude Code's "Continue automatically at usage limit" is on by default, and an abandoned session can spend your next 5-hour window https://sysnote.net/news/openai-disbands-its-robotics-team-to-bet-on-gpt-6-astra-the-best-evidence-74-8-i sysnoteen 2026-11-13T21:30:37+00:00 OpenAI disbands its robotics team to bet on GPT-6 Astra. The best evidence: 74.8% in a simulator that waits 18.6 seconds per move https://sysnote.net/news/anthropic-s-safety-eval-guide-claude-fable-5-1-scored-60-on-a-self-harm-eval-via sysnoteen 2026-11-13T21:30:16+00:00 Anthropic's safety-eval guide: Claude Fable 5.1 scored 60% on a self-harm eval via the API and 94% on claude.ai https://sysnote.net/news/weight-space-misfolding-chatgpt-6-1-sol-planted-one-false-fact-in-qwen2-5-1-5b-w sysnoteen 2026-11-13T21:30:13+00:00 "Weight Space Misfolding": ChatGPT 6.1-Sol planted one false fact in Qwen2.5-1.5B with a 3,066-parameter patch; no model "virus" was shown https://sysnote.net/news/justrua-arxiv-2611-16009-gpt-6-astra-s-weak-robot-arm-scores-were-mostly-the-har sysnoteen 2026-11-13T17:54:44+00:00 JustRUA (arXiv:2611.16009): GPT-6 Astra's weak robot-arm scores were mostly the harness, from 29.7 to 50.3 on RoboDojo https://sysnote.net/news/thinking-machines-releases-inkling-2-small-weights-under-apache-2-0-and-the-mode sysnoteen 2026-11-13T14:16:44+00:00 Thinking Machines releases Inkling 2 Small weights under Apache 2.0, and the model was trained in NVFP4 from the start https://sysnote.net/news/so-arena-benchmark-openai-s-decisions-api-scores-26-points-below-gpt-6-luna-as-g sysnoteen 2026-11-13T14:16:41+00:00 SO-Arena benchmark: OpenAI's Decisions API scores 26 points below GPT-6 Luna, as Gemini Decision 1 arrives untested https://sysnote.net/news/will-depue-s-blackwell-fingerprint-rebuilt-deepseek-v4-flash-and-kimi-k3-weights sysnoteen 2026-11-13T10:12:20+00:00 Will DePue's Blackwell fingerprint, rebuilt: DeepSeek-V4-Flash and Kimi K3 weights fit only 148- and 152-SM GPUs https://sysnote.net/news/mcp-vs-cli-for-agents-arxiv-2610-34298-no-winner-across-18-settings-and-the-stud sysnoteen 2026-11-13T10:12:15+00:00 MCP vs CLI for agents (arXiv:2610.34298): no winner across 18 settings, and the study's "MCP" is a JSON format in the prompt https://sysnote.net/news/codex-banked-reset-lands-on-thursday-november-12-tibo-says-it-is-activated-and-r sysnoteen 2026-11-13T10:12:12+00:00 Codex banked reset lands on Thursday, November 12: Tibo says it is "activated" and rolling out to accounts