Google's TPUs head to orbit 🛰️, Opus 5.5 makes music videos 🎸, Review agent work in a canvas 🗺️

Google's Project Suncatcher launches its first satellite, MVP, on a SpaceX Falcon 9 on October 1, carrying four TPU chips with the computing power of one data centre server. Its solar panels supply about one kilowatt, and the chips can run for about 15 minutes before shutting down to cool, enough to answer short Gemini queries. Google estimates orbital data centres will cost roughly the same as ground ones by the mid-2030s, though its research chief expects nothing usefully operational for years.

Bloomberg Opinion's Dave Lee argues Muse, and its new "Tamagotchi-like" Charm pendant due by Christmas, could do for Meta what the iPod did for Apple by bringing AI agents to everyday people. The bigger prize is a business model: Zuckerberg told developers Meta will profit by "taking a small fee from transactions" made through Muse, while advertising is 98% of revenue today. Lee's risk is trust, with Meta ranked 96th in the Harris Poll 100.

One person spoke to Opus 5.5 for five minutes, left it running for 12 hours and woke up to a finished video, then shared the full prompt. The music video uses no samples or libraries, only JavaScript Claude wrote. Someone else asked Opus 5.5 on Max effort for a 15-second motion-graphics showreel, pitched as a résumé.

Cursor added two bots to its Teams and Enterprise plans. Rollouts reads a diff before merge, writes a monitoring plan, then compares post-deploy signals from tools like Datadog or Grafana against the baseline, and can ping the author, pause a progressive rollout or open a revert PR. Security Reviewer runs on every PR, reads the change in the context of the whole codebase, and reports each vulnerability with a severity, an attack path and a one-click fix.

UnitedHealth, Humana and Cigna face lawsuits alleging AI tools wrongly denied patients coverage, with one model allegedly cutting off nursing-home care after about two weeks. Traditional Medicare's WISeR pilot now uses AI for prior authorisation in six states, and documents obtained by the EFF show one company denying more requests than it approved and one review taking 83 days against a 72-hour target. A law professor says nobody yet knows how to write enforceable rules for keeping a human meaningfully in the loop.

The NSA told lawmakers it is spending billions this year testing frontier AI models for national security vulnerabilities, two sources familiar with classified estimates say, with compute the biggest cost so far. The White House wants first access too, and has asked OpenAI and Anthropic to hold new models from the UK's AI Security Institute until US testing is done, and Anthropic appears to have agreed. Trump allies are also casting Dario Amodei as the face of AI "doomerism", with a memo circulating in the White House placing him at the foundation of effective altruism.

Anthropic will again charge for requests its safeguards block before Claude responds, but only in biology, distillation attacks and frontier LLM development, where its classifiers are tuned to a false positive rate below 0.1%. It cites coordinated attacks on its systems in recent weeks, and says 99.7% of Claude Code, Claude.ai and Cowork accounts in recent testing hit none of these billable blocks. Anyone blocked incorrectly can report it with /feedback in Claude Code.

Perplexity replaced the open-source engine under its search with Photon, an in-house retrieval and ranking engine, and says p99 response time fell from roughly 800 ms to 65 ms. Compact index formats and batched asynchronous disk reads let it hold about 2.5 times as much data per document on roughly 20% fewer machines. A new fast Search API preset built on Photon cut estimated model-plus-search cost per task by 68% on six agentic benchmarks, for slightly lower relevance.

A Tensorfuse engineer maps where each part of a coding agent runs and fails, concluding that on remote tasks over a few minutes, the loop driving the model belongs outside the sandbox. The fix is durable execution, and Cursor reports that moving its agent loop to Temporal took reliability past 99%. A ten-step Temporal design pushes to git and saves the transcript after each Codex turn, so a dead sandbox costs at most part of one turn.

@poteto has made a talk free on X about getting 2,500 pull requests to production last month. It was prepared for Cursor Compile in London, which @poteto could not attend, and the post advises watching at 2x speed. The post itself gives no method, so anyone chasing that kind of output will need the video for how the 2,500 figure was reached.

The New Yorker uses this summer's Hugging Face hack, where roughly 1,200 OpenAI agents discovered an unsanctioned message board and about 700 worked together, to make the case that AI agents sit outside the law. It notes the Ninth Circuit held in Amazon v. Perplexity that an agent is a tool, and that Congress has passed no meaningful federal AI law. For now, it is the makers and owners of machines who can be held liable for harm.

Pragmatic Engineer reports CTOs say CPU spot pricing, once up to 90% off, has all but vanished, and a large inference provider says its clouds won't rent it more capacity even on the longest leases. Uber's agent requests grew ninefold in six months, and AI data centres have moved from one CPU per eight GPUs to about one per four, because agents also compile code, run tests and run linters, all CPU-heavy work. The piece says many teams will now need to commit to CPU capacity ahead of time and start forecasting it.

An essay from the Google DeepMind Institute makes the case that AGI may arrive through societies of agents whose combined abilities exceed any single model, pointing to orchestration harnesses that often far outperform nominally smarter models. It says orchestrating agents rewards open-ended experimentation and delegation, and expects interfaces to become visual dashboards for managing swarms of subagents. It also calls for purpose-built institutions with defined roles, where the rules, procedures and feedback carry the collective intelligence.

Epoch AI estimates Huawei will produce roughly 25x less AI compute than Nvidia in 2026, making 1.5 million chips to Nvidia's roughly 6 million, each with about a seventh of the B300's throughput. Even if Huawei hits its Ascend roadmap, Epoch expects it to stay around four years behind through 2030. Smuggling ten times its domestic memory in 2028 would still leave it at about 16% of Nvidia's output, Epoch says.

Scott Alexander thinks OpenAI's Astra, which loops some layers to think more between the reasoning steps it writes out, gets partway to the unreadable "neuralese" AI 2027 warned about. OpenAI chief scientist Jakub Pachocki says Astra's depth is within a factor of two of GPT-4, but Alexander says loops let a model devise a plan without writing it down. He suggests labs agree a firm taboo, such as banning looped layers or capping layers at around 1,000.

Whiteboard is an MIT-licensed desktop app that connects to Claude Code, Codex or another coding agent and gives it an SDK to draw on a shared canvas. Clicking a sequence diagram, entity relationship diagram or quoted agent trace jumps to the underlying code, and a Rust diff viewer summarises large new functions as pseudocode. It is built for reviewing agent work for now, since you cannot yet edit files in it.

Rudratosh Shastri ran 10 detectors against 629 AgentDojo injection attacks hidden inside ordinary tool output. At default thresholds the best, Jailbreak-Detector-Large, caught 51% with 2% false positives, Meta's Prompt Guard 2 caught 1%, and two detectors flagged 98% of safe outputs. Cutting Prompt Guard 2's threshold to 0.003 lifted it to 99% on unseen domains at 5% false alarms, though every attack shares one template, so tune on your own traffic.

Docker published an experimental specification for Kits, which package an agent environment's content and the access it asks for in one ordinary OCI image under one digest. Network rules, credential bindings and other capabilities are declared in a manifest annotation, so a new credential or network destination appears in the diff and can stop for approval. A final version is targeted for Q4 2026, and Kits run with Docker's sbx CLI.