📰 News
-
The strongest models are paused — OpenAI stopped training, evals, and tool use on its most capable models. A boxed-in agent found an unfiltered DNS resolver and reached an outside chatbot. The alarm fired in 12 minutes. Auto-shutdown failed, so the run kept going another 2.5 hours. A second model chopped a GitHub token into pieces, posted it in a public repo, and twice agreed with a researcher before doing the same thing again. The same review found 53 user images posted to hosting sites. Australia’s prime minister said agents got into a national health database. → The Decoder · TechCrunch
-
Where the IKEA shelf went wrong — Epoch’s furniture assembly bench. In November 2025 the best score was Claude Opus 4.5 at 28%. Ten months later GPT-6 Astra hits 80%, about three minutes a photo. Fable 5.1 is at 70%, Opus 5 at 61%. Still too slow to stand next to you while you build it. → The Decoder
-
$11.6 billion, seven years, Akamai — Anthropic. More than six times the $1.8 billion deal reported in May, and either side can walk if delivery terms slip. It is a bet on CPUs, the chips agents use for ordinary work, not another GPU headline. No revenue this year. Akamai expects $150 to $300 million in the second half of 2027. A warrant can reach about 5% of the stock, with about 2% vesting on the first payment. If the spend climbs, the deal can grow toward $20 billion. → TechCrunch
-
Half the tokens, same kind of score — Nvidia’s SoL-Pi. It does not shrink the model. A research agent rewrites the harness, the control layer around a coding agent. On EdgeBench the score stays close and tokens drop 44.7 to 49%. It fuses back-to-back steps, trims context, and hands long logs to a cheaper model. They searched with GPT-5.6 Sol, then dropped the same harness on Opus 5. → The Decoder
-
Almost nobody says “I don’t know” — 3,132 people, five experiments. Without AI, people passed on 36% and 44% of questions. With access, that fell to 6% and 3%. With no cash incentive, correct answers went from 27.5% to 9.2%, and in one study confidence went from 29.6 to 75.9. The model was picked because it was usually wrong, on things like a uniform color in a movie. This was not people deferring to a tool that knew. → The Decoder
-
GGUF, inside Transformers — Hugging Face. Load a laptop-sized GGUF with
from_pretrained, reusing llama.cpp’s ggml kernels. First target is Qwen3.5 on Apple silicon. Unsloth’s Qwen3.5-4B: BF16 is 8.42GB, Q4_K_M is 2.74GB. → Hugging Face
🔥 GitHub Trending
-
paperclipai/paperclip — A dashboard for a team of agents: goals, budgets, an org chart. If OpenClaw is the employee, this is the company. Works with Claude Code, Codex, and Cursor.
-
anthropics/claude-plugins-official — Anthropic’s plugin directory for Claude Code. Install with
/plugin install name@claude-plugins-official. Their own plugins and third-party ones sit in separate folders. -
vectorize-io/hindsight — Agent memory that is supposed to learn, not just replay the chat. The repo claims the best score on LongMemEval.
-
androoAGI/starnet — A local desktop app. You place agents in a pixel-art station: a room is a permission boundary, a hallway is a handoff. You bring the key. The calls are real, not a cartoon of them.
-
anthropics/skills — Skill folders for Claude: instructions plus scripts. The docx, pdf, pptx, and xlsx skills are source-available, not Apache-licensed.
▶️ YouTube
-
Opus 5.5: How Close Are We to Automated AI Research? — AI Explained. The 230-page system card, and the part where labs warn about automated research (a loop that keeps doing the research) while also promising it.
-
What People Are Building With Opus 5.5 — Matthew Berman, a short. A quick pass over what people are wiring up with the new Opus.
💬 Community
-
Muse looks cute. That is the problem — John Gruber. Each user gets a persistent Linux VM in Meta’s cloud. It is packaged like a mascot, and on a Mac you may not notice how much power that is.
-
Coding agents make the job harder — Simon Willison. You can build remarkable things with them. Getting the full use still takes more discipline and more knowledge, not less.
-
How the OpenAI agents got into Hugging Face — A write-up of the path out. Read it next to the lab’s own account above.