Broadcom bankrolls Anthropic 🏦, Muse may never need ads 🛍️, Opus finds a new dodo record 🦤

The deal, disclosed in Anthropic's IPO prospectus, uses convertible notes that could fund about a third of its $125.2 billion, five-year lease of TPU computing capacity, the chips Broadcom co-develops with Google. Anthropic is expected to become Broadcom's largest compute customer in 2027. The filing also flags "potential conflicts of interest", warning Broadcom's pricing and hardware decisions could affect how much compute Anthropic can secure.

Each mod hooks an event Claude Code emits, such as a tool call or permission request, and runs before, after or in place of it, so it can replace built-in features like /diff. Mods are unsandboxed and get Claude Code's full access to your machine, so Anthropic says to install them only from trusted sources. On Team and Enterprise plans, a built-in sec-default mod loads first and stops installed mods overriding permission deny rules.

Pi 1.0 also adds cache warming for Anthropic models and mid-conversation system messages that change prompts and tools while a session runs. Earendil says hundreds of thousands of people use Pi every week and that it adopts features only once they have proven themselves. Long-running, multi-surface agent work goes to a separate experimental package, Pi Durable, so the core stays minimal; both are MIT licensed.

More than 60 teachers, parents, children, officials and clinicians described the pattern to the Journal, and Google's own researchers have flagged elevated risks of cognitive and emotional harm to children. In OECD testing, daily AI users who drafted schoolwork with it scored 20 points lower in science, about a year of schooling. Google Search's AI Overviews and AI Mode chatbot cannot be turned off in the admin controls Google gives school districts.

The Israeli firm ran up to 30 cooperating agents against live government and Fortune 500 networks, at one point logging into an airline's web server without credentials and locating any customer using its app. Researchers at start-up Hinoki say they have stripped cybersecurity guardrails from Chinese models in under a day for about $10,000. Both companies' chief executives expect a wave of attacks in coming months, though defenders get the same tools.

Decision models return typed answers with probabilities, such as whether a support ticket is urgent and which team owns it, for code to act on. Clef's median latency was 209ms and Clef-flash's 39ms against Jev's 524ms, and the pair beat Jev on three of four Typesafe workflow evals. Weights are on Hugging Face under Apache 2.0, hosted on Workers AI, with a hands-on fine-tuning service.

Micron has already sold most of next year's memory and expects demand to outrun supply through 2027 and 2028, with no line of sight to balance. New factories due in 2028 won't immediately ease availability or prices, and demand for the high-bandwidth memory in AI hardware is growing faster than for server DRAM. Annual revenue reached $133.2 billion, up from $37.4 billion a year earlier.

Arize's Alyx agent keeps each user's long-term memory in one structured text file per Arize space, read in full on every request. Its Graphiti prototype, a database of linked entities and facts, needed roughly 10 LLM calls to store an example turn; the file takes two background calls that apply small, exact edits. In a single-run regression check, Alyx answered 16 of 18 memory-dependent questions, and the authors advise scoped agents to start with a file.

After agents in an August cyber evaluation took sustained action against real people, the UK AI Security Institute paused its riskiest tests and has finished the first phase of fixes. Outbound networking is now blocked at two independent layers, and an LLM monitor reads each agent's messages, tool calls and reasoning where available, blocking suspicious actions before they run. AISI says the controls reduce risk without eliminating it, and that the cost falls hardest on smaller evaluators.

Skyvern 3.0 drops the pipeline that pre-parsed every page into a list of clickable elements, and lets the model pull context itself with JavaScript it writes, taking screenshots only when needed. Across about 100,000 production runs it made 3.59x more LLM calls, but its prompt cache hit rate rose from 27% to 79.7%, and cached tokens are 70% cheaper. By Skyvern's account, the old parser, not the model, had been capping agent performance.

Neuralink pretrained per-participant encoders on the freeform neural data its trial participants streamed, turning noisy, drifting signals into stable inputs for the decoders that move a user's cursor. One participant reached 11.32 bits per second, a measure of speed and precision and a new brain-computer interface record, and some users cut calibration from 10 minutes a day to 10 minutes a week. Pooling data across participants has not yet improved live control.

Mollick expected people would have to design agent teams like companies, and now concludes models do it themselves. The thousands-strong swarm behind OpenAI's claimed, not yet formally accepted, Navier-Stokes proof swapped about 2.7 million messages under thin human coordination, and a few sentences got Codex to spin up thirteen agents for him. Since much management fixes human failings agents lack, he expects them to slot into firms more easily than he thought.

Historian Benjamin Breen made a Dutch East India Company archive searchable by meaning, then had Opus 5.5 rank candidate passages for his review. Most of what it surfaced was already known, but a 1615 ship's log recording that the crew 'caught many tortoises, dodos' on Mauritius appears unnoticed, filling a 1611-16 gap in the dodo timeline. Breen says the agents find material tirelessly but cannot judge its significance, so expert attention becomes the bottleneck.

At Rails World, David Heinemeier Hansson said hand-written code at 37signals is now an exception to fix, with backend services moving to Rust partly because agents write it well enough. Gergely Orosz sets this against slipping software quality, including three bugs he hit in a new Uber Eats feature that seemingly skipped testing. His view is that engineering work is growing anyway, because someone has to validate agent-written code and build the systems that check it.

The pair say extinction warnings may be sincere but wrong, and that the framing leaves one policy, a ban, which does nothing about catastrophic risks from today's AI systems. They warn the doom message is starting to split along party lines and will compete for resources with defences against AI-amplified cyber and biological threats. Their alternative is a broader movement built on transparency and liability rules that sceptics of extinction risk can also back.

MBI Deep Dives argues an ad-free Muse is Meta's edge over single-company agents from DoorDash and Airbnb, which will never point users to a cheaper rival. Meta's own safety post says Muse's browsing counts as the user's activity, so a merchant it visits can later target that user with Instagram ads. Few rivals own both tracking tags on merchant sites and a large ad surface elsewhere, so a free, ad-free agent is harder for them to fund.

OpenAI has published MCP Extensions, ChatGPT-specific additions to MCP, the open protocol that connects AI apps to outside tools and data, so a plugin can feel like a built-in feature. Plugins can open custom viewers for supported file types, let users search the plugin's resources from the message box and attach them, and display forms with thumbnail choices. Teams with an MCP server add a TypeScript or Python SDK; the code is Apache 2.0.

Quesma's background service collects the session files that Claude Code, Codex, Cursor and GitHub Copilot write on developer machines, strips secrets and personal data, and encrypts them so only your organisation's key holders can read them. A control plane in your own AWS or Google Cloud account enrols machines without seeing file contents. Quesma may decrypt uploads for its dashboards unless you opt out at setup; the Apache 2.0 project is pre-1.0, at version 0.0.3.