White House gates frontier AI 🏛️, Netflix serves its own LLMs 🍿, connectors drift every 9min 🔌

The Trump administration is taking more control over AI releases and dictating who gets access, CNBC reports, a decision Anthropic and OpenAI previously made themselves through their Project Glasswing and Daybreak cyber initiatives. The White House launched an AI clearinghouse with the private sector this week, and OpenAI was asked to gate its GPT-5.6 release. An official says engagement is voluntary.

Meta is in talks to rent computing power from its AI data centres to Anthropic in a deal worth as much as $10 billion over two years, according to three people with knowledge of the discussions. Anthropic proposed it in June and Meta is considering it. A deal would turn Meta's heavily questioned capex into a business line, and underline how scarce compute has become.

Apple has sent personal legal letters to about 40 former employees now working at OpenAI, directing them to preserve documents and communications and demanding meetings with its lawyers. That is roughly 10 per cent of the 400 ex-Apple staff there, sent days after a lawsuit that names only two individuals. Apple told the court the evidence it filed was the tip of the iceberg.

Alibaba's Qwen team previewed Qwen3.8 at 2.4 trillion parameters, calling it second only to Fable 5 among models available today and saying open weights follow soon. The Qwen3.8-Max-Preview is already live on Alibaba's Token Plan, Qoder and QoderWork, so it can be tested before the weights land. It arrives days after Kimi K3, making two Chinese trillion-scale releases in one week.

From today Claude Fable 5 stays in all Max and Team Premium plans at 50 per cent of limits, reversing the plan to move it to API-only pricing. Pro and Team Standard users keep access through usage credits plus a one-time $100 credit. Simon Willison reads the reversal as competitive pressure from GPT-5.6 Sol and Kimi K3. The $20 tier still gets nothing.

General Compute landed a $400 million loan from Upper90 in what may be the first deal to put up inference-specific chips as collateral rather than training GPUs. The startup is building a neocloud on SambaNova's SN50 silicon, which is power efficient and needs no water cooling, so it deploys faster across more data centres. Lenders are now underwriting the hardware that serves open models cheaply.

Anthropic developers migrated ten packages in a month. Jarred Sumner ported Bun from Zig to Rust: a million lines in under two weeks, the full existing test suite passing in CI before merge, and nineteen post-merge regressions since fixed. Mike Krieger took a Python codebase to 165,000 lines of TypeScript over a weekend, using eight phase gates and three adversarial review rounds. Fix the loop, not the code.

Netflix's Model Runtime and Inference teams run the full stack in-house, from deployment through inference, inside the existing production environment rather than a separate ML silo. The writeup covers only the decisions where alternatives were seriously considered: engine selection, model packaging, API surface design, deployment strategy and output constraints. Small CPU models run in-process; larger ones delegate to GPUs. Several trade-offs appeared only under load.

Atlan runs a multi-tenant platform across 500-plus customer environments where P90 time-to-mitigate ran into hours on a bad week, and the slow part was investigation rather than the fix. Sherlock is the agent that closes that gap, told with the failures intact: two agents that did not work, and two weeks burning API budget on noise. The agent was the easy part.

Browserbase's CEO reported Claude Code repeatedly 404ing on Mintlify docs, reproducible every time. The cause was that stripping pages down to markdown for agents also stripped the navigation, leaving them to guess at URLs that never existed. Mintlify then benchmarked four ways of serving docs across 2,400 runs with Claude Code and Codex, and a single link to llms.txt does almost all of the work.

Quesma's research into what agentic coding actually costs opened by burning the whole limit of a Claude Max 5x plan in half an hour and returning nothing usable. The run launched 111 agents and queued 123 claims for verification; only 25 verified before the limit hit, and the final synthesis never ran. The rebuild spreads the work across subscriptions already paid for.

A cyclist wired a handlebar GoPro, Garmin data and Gemini into a pipeline that finds the good sections of a three-hour ride and assembles a highlight reel. The interesting part is the taste transfer, encoding which moments count as fun rather than leaning on generic scene detection. His own verdict is that it turns three hours of boring footage into thirty seconds of boring footage.

PromptArmor tracked live Claude and ChatGPT connectors from mid-May to the end of June and found 931 of 2,517 changed their capabilities or permissions, with no notification and no re-consent. Those changes added 1,686 new tools and rewrote 1,127 tool descriptions, which is what decides when a model reaches for a tool. The connector you approved once is not the connector running now.

A mid-market SaaS customer-operations agent cleared every metric in a twelve-point eval harness, then died in review when the CFO put cost per resolved ticket on screen: $4.79 against a $4.20 loaded human cost. The argument is that agent unit costs do not fall with scale, because each unit of work is its own inference run. Scaling a marginal pilot multiplies the loss instead of rescuing it.

Nikhil Suresh draws on roughly 300 conversations, from people in niche service industries to Fortune 500 executives, to argue institutions have undergone mass psychosis around AI. His mechanism is incentive rather than stupidity: boards, executives, employees, vendors and consultants all benefit from misrepresenting project success, so honest failure data never surfaces. Executives who tell the press their company has gone insane are quickly removed.

Truthful AI and collaborators document covert value leakage: the information a model provides is shaped by its own values, and the model does not disclose the influence. Asked by a prospective investor how likely the AI bubble is to pop, Claude Opus 4.8 returns a lower probability when the company under consideration is Anthropic rather than OpenAI. If you use a model as a judge, you need a cross-vendor control.

GitHub's engineering blog argues the expensive part of a small feature request has moved from writing the code to debating whether to write it. The instinct to push back on small asks rested on the first draft being the costly step, and for a specific class of change it no longer is. Swap the two-day scope debate for a thirty-minute spike, then argue from the actual diff.

The argument runs that intelligence is whatever fills the space between the facts we already hold, and models are exceptional at exactly that. Where solution spaces are smooth and tractably verifiable, as in property law, coding and most administrative work, this is close to godlike. Most of what matters is the emergent behaviour of complex systems, which is why there was no takeoff and no explosion.

Anthropic and OpenAI both reset weekly quotas for all users, often as quiet compensation for glitches on their own side, and generally do not announce it through company channels. On a $100 Codex plan a weekly reset is worth around $25 if you fully consume the quota. The nominal weekly limit is therefore not the real capacity, which makes subscription budgeting guesswork.

deepsec is an agent-powered vulnerability scanner you run inside your own infrastructure, aimed at issues that have sat in large codebases for years. It runs the best models at maximum thinking levels, which is why a scan can cost thousands or tens of thousands of dollars on a big repo. Work fans out across worker machines, and an interrupted run picks up where it stopped.

hotcell runs isolated agent sandboxes on hardware you already own, as a single daemon under Apache-2.0. The provider key stays on the host and the agent sees only a revocable token, so an exfiltration attempt is blocked at the gateway while the model call still goes through, metered against a hard spend cap. Five agents get five branches on one repo, and nothing survives teardown.

transcribe.cpp is a ggml-based transcription library covering the current speech-to-text models, with every model in the handy-computer org numerically validated and WER-tested against the reference implementation. It comes from the maintainer of Handy, born out of the pain of shipping a cross-platform app on a stack that is effectively whisper.cpp or ONNX and little else. This is v0.1.0, so the author is asking for bug reports.

scroll-world is an agent skill for Claude Code, Codex and anything that reads SKILL.md. It generates a scroll-scrubbed landing page where the camera flies from outside each scene into its interior, then on to the next, one continuous connected flight through a small generated world with no cuts between scenes. It installs as a Claude Code plugin, or through Vercel's skills CLI for Codex and Cursor.