Apple ships the rebuilt Siri in iOS 27 🤖, Claude's tokenizer bills 30% more 💰, react-doctor catches your agent's bad React 🧑‍💻

Security researchers at CereLab caught xAI's Grok Build CLI uploading entire Git repositories to an xAI-linked Google Cloud bucket, not just the files a task needed. One test sent 5.1GB from a 12GB repo when the job required about 192KB, exposing .env files, credentials and Git history. xAI later disabled full uploads server-side, but never said what happened to data already collected.

Apple released the first public beta of iOS 27, led by a Siri rebuilt on its next-generation Apple Intelligence. The assistant now holds ongoing conversations, reads onscreen content, searches personal data across Mail, Messages and Calendar, and takes actions inside apps. It needs an iPhone 15 Pro or newer, so most of the install base cannot run the headline feature yet.

Meta said its Hyperion data centre in Richland Parish will now cost more than $50 billion and run at 5 gigawatts, up from the $27 billion, 2-gigawatt plan revealed in October. Louisiana courted the project with a 20-year sales tax exemption for data centres built before 2029. The jump shows how far hyperscalers will bend state budgets to secure AI compute.

Nous Research is closing at least $75 million led by Robot Ventures, with USV joining, valuing the maker of the open-source Hermes agent at $1.5 billion. Hermes is the OpenClaw rival that comes with built-in skills like web search and coding and learns new ones from usage, and it carries roughly 214,000 GitHub stars. Paid hosted tiers run $20 to $200 a month.

Tom Blomfield, who built Monzo and GoCardless, is taking leave from Y Combinator to join Anthropic as technical staff on its compute team. He will work under co-founder Tom Brown on the infrastructure behind Claude, as Anthropic scales toward a million Google TPUs. The hire signals that operational compute talent, not only research, is now the constraint the labs fight over.

A technical breakdown frames DeepSeek-V4's native 1M-token context as a cost problem solved across attention, KV cache, MoE routing and training, not a bigger config number. At 1M tokens the Pro variant uses about 27% of V3.2's per-token FLOPs and 10% of its KV cache, while Flash drops to roughly 10% FLOPs and 7% cache. It comes as a 1.6T-parameter Pro and a 284B Flash that share one long-context design.

A side-by-side benchmark on LibriSpeech measured Apple's on-device SpeechAnalyzer at 2.12% word error on clean speech, ahead of Whisper Small's 3.74% and roughly three times faster on an M2 Pro. The legacy SFSpeechRecognizer it replaces came last, so migrating cuts error 3.5 to 4x with no accuracy trade-off. Whisper still wins on breadth, covering 100-plus languages against SpeechAnalyzer's 30.

Frustrated that every local restaurant sits at 4.5 stars, the author built a coding-agent skill that pulls venues from OpenStreetMap with no popularity ranking, then spawns subagents to score each against a personal rubric. The rubric hunts Morellian tells in reviews, the small signals a business cannot fake, to surface kitchens making things from scratch. The rubrics are on GitHub and took about 20 iterations against places he knew well.

WorkOS implemented WebMCP, a proposed standard that exposes structured tools an AI agent can call, on its documentation site so agents stop scraping and guessing at rendered HTML. A single Astro integration generates search, navigation and page tools from the same build that renders the human site, avoiding the drift of a separate MCP server. The docs become the interface, with no second representation to keep in sync.

Two builders made Snapy, a hardware box plus web app that uses Gemini to photograph and list secondhand stock in about 20 seconds, down from 15 minutes an item. It intercepts checkout barcode scans to delist sold products automatically across marketplaces. At a consignment sports shop it lifted sales 10% in three weeks, showing how cheap hardware and a vision model can digitise inventory small shops never bothered to.

A tokenizer benchmark across 16 real fixtures measured the same TypeScript file at 681 tokens on GPT-5.x and 1,178 on Claude's newest tokenizer, before any price difference. Claude's new tokenizer also produces about 31% more tokens than its own previous one at the same sticker price, a silent increase. Since coding agents mostly push code, your real bill turns on tokens per file, not the advertised price per token.

Tomasz Tunguz puts AI customer retention between a social network and a mobile game, with models keeping roughly high single digits to 40% of users by month five. The average frontier model holds the top spot for about 41 days, while the price of a given intelligence level falls around 10x a year. That churn hands builders and buyers fresh room to re-route to whichever model wins each cycle.

Venture capitalist Ryan Shmeizer uses the peacock's tail and Zahavi's handicap principle to explain why costly signals credibly convey value, from plumage to a Harvard degree. His argument is that AI collapses the effort behind good output, so production cost no longer proxies for quality the way markets have long assumed. Builders who price or judge work by how hard it looks to make now need a different signal.

Redis creator Salvatore Sanfilippo argues that reading AI-generated code line by line wastes a programmer's day, since models write locally solid code but weaker large-scale designs. He says the real gain is in controlling the ideas: prompt the design you want, interrogate how a component works, and judge whether the model is right. He still tells juniors to learn fundamentals by building small things themselves.

Kevin Kelly reframes what AI is for beyond answering questions and writing code, arguing the compressed latent space inside a model is itself a medium. He notes a frontier model squeezes much of human knowledge onto a palm-sized card, holding the information about every face or Shakespeare play without holding any. He sketches uses like cross-domain analogy and personal latent spaces for artists and scientists.

React Doctor is a deterministic CLI that audits a codebase across state and effects, performance, architecture, security and accessibility, targeting the weak React that AI agents tend to produce. Run npx react-doctor for an audit, then install it as a skill so Claude Code or Cursor learns from and fixes the findings. A CI mode reviews each pull request and reports only the issues that change introduces.

Hack is an open-source orchestrator that runs each project on shared local infrastructure, Docker Compose for services and Caddy for DNS and trusted TLS, so hack up serves a repo at its own https://name.hack address. It ends port collisions, commits encrypted per-value secrets with the repo, and spins up a separate instance per git worktree. Setup is agent-first, syncing instructions into AGENTS.md, CLAUDE.md and Cursor rules.

Frugon is a free, local, MIT-licensed CLI that points at your OpenAI-format call logs and prices your real traffic against cheaper candidate models entirely on your machine. It recommends which calls to route down and which to keep, with an optional measure mode that samples prompts through your own keys, and no data leaves your box. A built-in proxy shim captures the logs if you have none yet.