Siri AI beta 🍎, Run a business autonomously 🤖, Do you still read the code? 👀

Apple began rolling out Siri AI in beta in English on 14 September, an entirely new Siri that pulls from a user's messages, emails and photos and takes actions across apps. Private frameworks in iOS 27 and macOS Golden Gate go further, MacRumors reports: one protocol lets Claude act as a Siri extension, and another replaces Apple's server-side Siri model with GPT-5.6. Apple has not yet opened that entitlement to third-party developers.

Staff below CEO level at Anthropic, OpenAI and Google have met regularly since July to discuss creating an industry standards body, The Information reports, citing people familiar with the matter. Sam Altman told an OpenAI town hall that labs would have to build one without US government support, after a draft White House executive order for such a body stalled in late summer. The talks are ongoing, and it is unclear where Meta stands.

Nvidia now uses Anthropic's Fable only for tasks that don't touch sensitive data, and a large US utility cancelled a test of Fable running its core power infrastructure, The Information reports via Tom's Hardware. Both cite Anthropic's refusal to guarantee nonrevocable zero data retention, after a June policy change let it retain customer data to check for misuse. Novo Nordisk bans proprietary data from Claude, and Northrop Grumman runs open-source models on air-gapped servers.

Andon Labs has made Pion, the platform behind its own AI-run businesses, available as a waitlisted research preview. Agents get email, phone, banking and a browser; Andon's vending machine at Anthropic's office turned a profit by late 2025, but the San Francisco store and Stockholm cafe it handed to agents in April 2026 still lose money. Andon wants data from many more businesses on what agents can earn, and where they collude or lie.

Oracle emailed workers about a new round of layoffs, Fast Company reports, as it funds an expected $90 billion to $95 billion in capital expenditure this year with $125 billion in debt. Its headcount fell by about 21,000, or 13 per cent, in the year to 31 May. It has since raised the cost of its restructuring plan by $700 million to $2.8 billion, largely for severance.

A Colossus profile of Mark Zuckerberg says that in the 11 weeks after Llama 4's mocked April 2025 release, he folded FAIR into Meta Superintelligence Labs, paid $14.3 billion for a Scale AI stake and made Alexandr Wang head of AI. Meta is spending up to $145 billion this year, and the profile calls him an inveterate micromanager who engages his reports' reports. Rivals steelmanning Meta say it only needs to stay within six months of the frontier.

Anthropic engineer Sachin Malhotra says CI job volume grew 25x in six months as Claude wrote 80 per cent of engineers' code and tests grew 10x. Patches to the service that picks which tests run on each change (a bigger machine, sharding, daily restarts) lasted 70 days, 29 days, then under a day. A stateless redesign on an in-memory store took one engineer three weeks, and he advises assuming 25x load within two quarters.

Addy Osmani argues agents in old codebases need a person to map green, yellow and red zones, with autonomy set by blast radius rather than model confidence. He cites SWE Refactor Bench, where only 28 of 520 agent runs passed its migration audit, behavioural tests and independent verification. Current behaviour should be pinned by characterisation tests from a separate pass or a person, and one migration finished, old path deleted, before parallelising.

Patrick McCanna moved his most context-expensive agents from frontier APIs to self-hosted open-weight models on Ollama, and the prompts fell apart within 3 minutes. On his 128GB AMD Ryzen AI Max+ 395, a 35KB prompt used 14 per cent of a 65k-token context window, and agents repeated tool calls, re-read files and rewrote finished work. He splits prompts into single-objective agents, sets Ollama's small default context explicitly and logs session state for handoffs.

Bryan O'Sullivan used GPT-6 Astra in Codex and Fable 5.1 in Claude Code to chase stripes in greyscale images on a cheap Xteink X3 e-reader. Astra's frequency analysis measured the image's own dither pattern instead, until a generated test pattern exposed an eight-pixel stripe period and dark grey rendering as black. A longer four-grey waveform cut column variation from about 4 per cent to 1 per cent, and his driver pull requests merged within hours.

A Show HN page offers 30 fresh prompts in the style of the pelican-on-a-bicycle SVG test, such as an octopus operating a pipe organ, and has run 10 of them on six current models with time and cost per drawing. On the octopus prompt, DeepSeek V4 Pro finished in 1 minute 36 seconds for $0.10, while Qwen3.8 Max took 12 minutes 37 seconds. Claude Fable 5.1 was the most expensive there, at $0.79.

Entelligence ran GPT-5.6 Luna and GPT-6 Astra over the same 50 benchmark pull requests, and Luna found 69 verified bugs to Astra's 92 for $0.20 against $5.66. Luna was noisier, with 74 per cent of its findings holding up against Astra's 96 per cent, and it caught 9 of 24 security bugs to Astra's 19. Entelligence says Luna suits everyday correctness bugs but shouldn't review authentication or permission code alone.

A developer writing on zanlib.dev splits AI coders into accelerators, who read generated code to keep understanding it, and vibecoders, who delegate implementation. Citing Peter Naur's 1985 essay, the author says the real product of programming is the theory behind the code, and a drifting accelerator becomes a vibecoder without the specification and evaluation harness a deliberate vibecoder would build. Mixed teams, the author says, should agree how changes will be maintained before merging.

Armin Ronacher had Opus 5 recreate a David Sacks tweet that AI-text detector Pangram flagged, then rewrote the output by hand, paragraph by paragraph, keeping the structure and ideas but no identical sentences. Pangram, which reports a 0.0041 per cent false positive rate, still rated the rewrite 100 per cent AI. Ronacher says text built on an LLM's structure scores badly even after plenty of edits, though he expects many will consider that label misleading.

dbt Labs has open-sourced dbt Charts under Apache 2.0, a language that wraps SQL in YAML so an agent can declare a dashboard in one auditable file instead of a pile of HTML, JavaScript and React. Validation catches a renamed model or missing column in the pull request, and the CLI renders boards to SVG, HTML, PNG, PDF or the terminal. Agent-built charts can live in Git beside dbt models, though the language is pre-1.0.

shadcn-ui's @shadcn/lint is an ESLint and Oxlint plugin that lets teams declare which Tailwind classes each component may take, with errors suggesting fixes from their components, variants and theme. Across more than 150 agent task runs, almost every task reached zero violations in one correction round, and fixes in Claude control runs cost 10 to 48 per cent less than with rules alone. It works on Tailwind v4 projects without shadcn/ui or a component rewrite.