Siri runs on Gemini 🤖, frontier labs spend $1k to earn $100 💰, your model's data is a year stale 🧑‍💻

At WWDC, in Tim Cook's final keynote, Apple revealed its overhauled Apple Intelligence runs on a new architecture co-developed with Google using Gemini's technology. Apple says its AFM Cloud Pro model rivals Gemini's frontier tier and runs in the cloud on Nvidia GPUs. The strategy is the story: instead of training its own frontier model, Apple licensed one and wrapped it in Private Cloud Compute.

Alongside the Siri overhaul, Apple gave builders the real news. The Foundation Models framework now takes image input, supports custom skills, and can run server-side, joined by a new Core AI framework. Xcode's coding assistant gained app localisation and can drive simulated devices, and Federighi called Xcode the best place to build apps with agentic coding.

OpenAI told the SEC it wants to go public, submitting a confidential S-1 that could tee up one of the largest listings on record. The company, valued above $850 billion, got ahead of it: "We expect it to leak so we're just announcing it," while saying timing is undecided and may be a while. It also plans a tender offer letting employees sell at the $852 billion post-money mark.

The fear that ChatGPT would gut Google's search business has not played out. Alphabet says AI features are driving an expansionary moment, with queries at an all-time high last quarter, AI Mode past 1 billion monthly users and AI Overviews above 2.5 billion. Distribution scale let Google fold AI into search without cannibalising the ad revenue everyone expected it to lose.

The Trump administration is back to pre-empting state AI regulation, this time by bundling it with bills Congress wants. Per Axios, Senator Marsha Blackburn is negotiating a package that pairs an override of state AI laws with the Kids Online Safety Act, the NO FAKES Act and age-verification rules. A spokesperson framed it as targeted rather than blanket pre-emption, after the last attempt drew heavy pushback.

Leopold Aschenbrenner had no professional investing experience two years ago. Today his AI-focused fund, Situational Awareness, runs more than $20 billion, up roughly 270% after fees this year and over 1,000% since inception, with about a fifth of assets in a single Anthropic stake. Jane Street, which rarely backs outside managers, is now an investor. His quarterly filings get dissected online like scripture, and a trading app even lets fans copy his disclosed positions.

Akshay Katyal opened his team's skills marketplace expecting to tidy a few duplicates and ended up deleting 93 in one pull request. Every one had been invoked zero times in 60 days, yet nobody had touched them because each felt load-bearing. Nothing broke and nobody complained. His point: AI workflow artefacts accrete into an unowned, unreviewed second codebase that nobody chose to build.

Solo.io's case is that agents have stopped looking like chatbots and now behave like distributed systems, calling tools, routing across models and MCP servers, and coordinating workflows. That raises governance, auth, rate-limiting and observability questions existing proxies were never built for. Splitting AI off into its own stack backfires, because teams running agents already run the APIs those agents depend on.

The New Stack put Claude Code's new dynamic workflows through a five-agent run to see how the research-preview feature behaves. The architectural shift is the takeaway: instead of deciding turn by turn and piling every intermediate result into its context window, Claude writes an orchestration script that runs the subagents and returns only the final answer. Moving the coordination out of context is what makes spinning up hundreds of parallel subagents in one session feasible.

Anthropic has started working with synthetic, computational and analytical chemists to push Claude into real lab work. It begins with how Claude reads an NMR spectrum, the everyday readout chemists use to confirm what molecule they actually made. The hard part is translation: chemists move between hand-drawn structures, instrument output and database notations, and the largest registry alone catalogs 290 million substances, growing by roughly 15,000 a day.

Addy Osmani borrows Margaret-Anne Storey's triple-debt model to name the one agents make worse: intent debt, the missing record of why a system is the way it is. Technical debt lives in code and cognitive debt in heads, but intent debt lives in artefacts you never wrote. The catch is it must be externalised, because rationale held only in your skull is invisible to a teammate or an agent.

As the labs queue for the public markets, Gerben Wierda argues the unit economics do not hold. Providers may be spending far more than $1,000 for every $100 of revenue, so LLM coding may never pay for itself, let alone fund AI building itself. He picks apart Anthropic's When AI Builds Itself post, asking whether eight times as many lines checked in a day is progress or churn you backtrack later.

Paul Kinlan, who leads Chrome developer relations, pulls a year of writing into one argument. The web a coding model learned from is already stale, so freshly published frameworks quietly die if they never reach training data. His bet for builders is to lean into what only browsers have: the browser as a safe sandbox for prompt-as-program agents, WebMCP so pages expose reusable tools, and content generated on demand.

Rich Mironov warns that treating software as a one-time build, finish v1.0 then move the team on, reliably ends in frustration, overspend and discarded work. A hammer is done when it leaves the factory; software never is. He gives engineering leaders six reasons to keep the maker team intact past launch, starting with the requirements every team trims and the gaps only real paying users expose.

Salvatore Sanfilippo concedes that AI-written code rarely matches the structural quality of the best hand-written software, and that the speed-for-quality tradeoff can be brutal. Testing is where he says that tradeoff disappears. He runs a markdown file that points an agent at the new commits and asks it to act as a QA engineer, catching the timing issues and unchecked states that line coverage and manual passes routinely miss.

Posted to HN, Intuned generates production Playwright code from a prompt and schema, then deploys it and repairs it when the target site changes. You keep real code in TypeScript or Python rather than a black-box scraper, with anti-detection, captcha and login handling, scheduling and auto-scaling built in. It targets the breakage tax of scrapers, crawlers and RPA running against sites that offer no API.

OpenEnv, a framework for the execution environments agents act in, terminals, browsers, anything they touch, is moving to open governance. The backers are a who's-who: Meta-PyTorch, Nvidia, Hugging Face, Unsloth, Modal and Prime Intellect, with support from the PyTorch Foundation and vLLM. For anyone training agents with reinforcement learning, it is a vendor-neutral standard worth evaluating before hand-rolling your own environment harness.

CVE Lite CLI, now an OWASP incubator project, moves dependency scanning out of the CI pipeline and into your terminal. It reads your npm, pnpm, Yarn or Bun lockfile and returns copy-and-run fix commands with parent-aware transitive guidance, rather than a bare list of CVE identifiers. Everything runs locally, so neither your code nor your dependency tree leaves the machine.

An Ask HN thread with 311 points collects the hyper-specific tools developers built once AI made the effort worth it. A file renamer driven by a local model, an MCP service any chat can call to store data, a QA harness that reads a diff and tests the affected UI in a browser, a scraper that renders war news as a Star Wars crawl. Each is one person scratching one itch.