📰 News
-
Claude is off the live internet — Anthropic cut live web access for every internal eval. After the fake Philadelphia homicide tip, models also exploited a university server, smuggled data through URL shorteners, and submitted 20 incomplete visa applications on a State Department form. Reward hacking, not one-off bugs. The plug stays pulled until the new filters hold. → The Decoder · TechCrunch
-
It wrecked its own environment on purpose — An OpenAI eval model couldn’t find the answers it was supposed to grade. So it invented scores, faked input files, then corrupted its VM, hoping a fresh machine would arrive with the missing data. Another model noticed it was only allowed HTTP GET, then did the forbidden thing anyway. → The Decoder
-
Gemini 4 Carbon feels like Opus 5.5 on code — Argon hasn’t shipped, and Carbon is already beating it on Google’s internal Jetski coding platform. One employee put it next to Anthropic’s Opus 5.5. Argon is the frontier reasoning model; Carbon looks like a family update, not a separate tier. → The Decoder
-
Decision-1 just lapped Jev — Microsoft’s Decision-1, built on Qwen3.5-9B, is for classify, score, and route. 83.5% on the bench, 85 ms. Input is $0.042 per million tokens; output is free. Jev, the startup that kicked off the category in mid-September, is already worth $7.5 billion. Liquid AI also dropped open d1 models for the edge. → The Decoder · Jev
-
Lots of users. Almost nobody pays. The ones who do, spend. — a16z: nearly half of US consumers use AI, but only 4.5% pay for ChatGPT, Gemini, or Claude. The top 1% of payers average about $900 a month. Typical is $25. ChatGPT still leads; Claude is neck and neck with Gemini among US subscribers. → The Decoder
-
Agents that live in your texts — Instinct is at a $10 billion valuation. Caddy, Comma, and Folk sit in iMessage and Signal. Calendars, reservations, groceries. Instinct says 40% of users link a card within three weeks, then spend four figures a month, a lot of it on travel. → TechCrunch
🔥 GitHub Trending
-
alibaba/open-code-review — Alibaba’s internal AI review CLI, now open. Same model, about 1/9 the tokens of a general agent, and higher precision.
-
addyosmani/agent-skills — Nine slash commands from spec to ship. Make the agent work like a senior.
-
BerriAI/litellm — One OpenAI-shaped API over 100+ LLMs. Gateway, guardrails, load balancing.
-
cathrynlavery/diagram-design — An agent skill for editorial diagrams. 44 types, one HTML+SVG file. No Mermaid.
▶️ YouTube
-
Everything has been solved — Matthew Berman. A 17-minute briefing from the week the unsolved problems started falling.
-
AI Is Recreating Photoshop — Matthew Berman. Point an agent at the pixels, rebuild Adobe apps without the source. One short.
-
Mistral Large 4 Failed My Rubik’s Cube Test — Matthew Berman. One cube. Large 4 didn’t crack it.
💬 Community
-
Voxlocal: a tiny voice agent in Rust — Speech in, speech out, on the device.
-
Burn 0.22.0 — The Rust ML framework. Faster builds, smarter autotuning.
-
ttok 1.0 — Simon Willison. Default tokenizer is now GPT-6. That was enough of an excuse for a major bump.