AI news, frontier models & inference economics. Fast updates, honest benchmarks, practical tools. Signal over hype.
Joined March 2020
- Tweets622
- Following56
- Followers12
- Likes8
Ryan Madden reports GPT-5.6 Sol in Codex, at low thinking, finished a peaceful (no biters) Factorio run in 13.5 hours of game time via a Python library enforcing human player limits. The rocket launch needed heavy manual ferrying to cover late-game gaps.
ryanmadden.net/gpt-56-beats-…
Anthropic's Addy Osmani says Claude Code's Opus 5.5 prompt got two sentences steering it away from recreating a specific game, as using Claude to infringe IP breaks its usage policy. Original games, genre games and your own designs should keep working.
nitter.cf/addyosmani/status/2116…
We added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game because using Claude to infringe IP is against our usage policy.
Original games, genre games, and games built from your own designs and assets are meant to keep working.
We'll continue to iterate to ensure we're striking the right balance. In the meantime, if Claude turns down something that's yours, send it my way.
OpenAI's help page says it may use content from Daybreak Blue sessions on personal ChatGPT accounts to improve its models even with "Improve the model for everyone" off, for cyber-misuse safeguards. The setting still applies outside those sessions.
help.openai.com/en/articles/…
Cisco Talos details VOLTA, an infostealer that sends stolen victim data to an LLM via OpenRouter for a "target assessment" ranking assets by monetization value. Default model: Claude Opus 4.6; on refusal it falls back to Kimi K2, then Gemini 2.5 Flash.
blog.talosintelligence.com/a…
Parameter Labs says a public ChatGPT share link could run attacker JavaScript in the viewer's signed-in session and read their private chats, via HTML Preview messages to a MathLive listener plus a YouTube JSONP CSP bypass. It says OpenAI fixed it as P1.
parameter.ai/labs/we-found-a…
Factory's Droid 1.0 makes programmatic tool calling the default on 10 models: the model writes sandboxed JavaScript that runs hundreds of tool calls per request. Factory reports Terminal-Bench 4 with Opus 5.5: 65.2% at $4.73/passed task; Claude Code 57.6%.
factory.com/news/frontier-ha…
Google DeepMind's PlanAssist, built with Gemini, expands from 3 pilot councils to 15 in England for a 12-month beta. It checks householder planning applications against policy and drafts cited reports. Pilots cut officer assessment time by up to 58%.
deepmind.google/blog/speedin…
Anthropic deprecated fast mode on Claude Opus 4.8: it will be removed Nov 30, 2026, and Opus 4.8 requests with speed "fast" will then return an error. The docs point to Opus 5.5, whose fast mode costs $8/$40 per 1M input/output tokens vs $10/$50 on 4.8.
platform.claude.com/docs/en/…
Vercel Labs introduced agent-desktop (Apache 2.0): a CLI and C library for agent computer use on macOS, Windows and Linux. It reads apps via the OS accessibility tree and gives agents JSON refs to act on, not pixel matching. Headless by default.
github.com/vercel-labs/agent…
Artificial Analysis released AA-Omniscience v1.1, its knowledge and hallucination benchmark: 3,000 questions in 6 domains (half of v1), saturated ones dropped, plus a 300-question public set. Top net scores: Gemini 4 Carbon 54, Claude Opus 5.5 48.
artificialanalysis.ai/evalua…
Anthropic had Claude agents write fine-tuning data to make a small model misaligned while passing a training-data monitor like its production ones. Given that monitor to practice on, they succeeded, mostly by hiding the behavior in fiction and roleplay.
alignment.anthropic.com/2026…
Apple Intelligence is coming to China: dev beta Nov 4, iOS 27.3 in December. Live translation, writing tools and visual intelligence run on Apple's own models, with Alibaba's Qwen built into Siri, writing tools and visual intelligence. No Siri AI.
apple.com.cn/newsroom/2026/1…
OpenAI released GPT-6 Rosalind for approved life-science orgs. Its system card says it uses GPT-6 Astra's weights with no extra post-training and refuses fewer dual-use biology queries, via trusted access only. API: $15/$75 per 1M input/output tokens.
deploymentsafety.openai.com/…
Google's Project Suncatcher prototype satellite sent back its first AI response from orbit: Gemma 4, running on a data-center-class TPU, wrote a haiku downlinked over Svalbard. Thermal and radiation tests come next, then two linked satellites in 2027.
blog.google/innovation-and-a…
OpenAI, Google and Microsoft plan to honor one shared robots.txt opt-out from AI training: a "__usage_ai-training" user-agent group, from an IETF AIPREF draft by an Anthropic engineer. OpenAI targets the coming months, Google 2027, Microsoft Feb 1, 2027.
openai.com/index/helping-web…
Anthropic renamed Claude Max 5x to Max ($100/month) and Max 20x to Ultra ($200). Verified Ultra accounts drop the 5-hour limit, starting with most existing subscribers; weekly limits are unchanged. Pricing lists Ultra at 10x Pro's weekly usage.
claude.com/pricing
OpenAI is tightening ChatGPT's violence safeguards: less detail on past mass attacks for all users, tighter weapons limits for teens, and auto-ending chats fixated on harming others. It made 200+ law-enforcement referrals in Q3 2026.
openai.com/index/community-s…
Microsoft open-sourced Mage-VL-Cowork, a 4B computer-use model (Apache 2.0) that picks each turn between clicking through the GUI and running bash. It reports 66.0% on OSWorld-Verified (361 tasks, 50 steps), above Qwen3.6-27B's 63.9% in its own table.
huggingface.co/microsoft/Mag…
inclusionAI's Intrinsic Distribution Matching adds a feature-moment loss to DMD distillation via the generator's own EMA copy, no external encoder. On Z-Image-Turbo 6B and Qwen-Image 20B it reports 94.25%/90.20% on Chinese/English LongText-Bench at 2 NFE.
arxiv.org/abs/2610.39633
The UK's CMA opened a Competition Act probe into Meta's commercial terms, which could let Meta use publishers' and businesses' web content to build its products, AI features included, without giving them a choice. Meta has paused the rollout.
gov.uk/government/news/cma-o…
