A top SQL benchmark got 52% wrong 🥇 Anthropic's agents run lab robots 🔬 Hugging Face's $399 robot duck 🦆

Meta is one of Anthropic's largest customers even as Mark Zuckerberg publicly attacked leading labs for consolidating power. At one point this year Meta internally projected spending up to $10 billion annually on Anthropic's models, after engineers took to Claude Code so heavily they competed on internal token leaderboards, a practice called tokenmaxxing. Meta still spends hundreds of millions a month while shifting engineers onto its own Muse Code.

Anthropic opened a research preview of the Model Hardware Standard, a specification letting AI agents operate lab and factory instruments in parallel. A standardised driver cuts integration from weeks to hours: Carnegie Mellon ran dose-response experiments about three times faster, and a QuEra agent built a controller recovering its laser lock 99.3 per cent of the time unaided. MHS is model-agnostic and speaks MCP, so existing agent harnesses work.

Google's Gemini Omni 1.1 Flash is now production-ready for generative video on the Gemini API. Scene extension now reads up to 10 seconds of prior context rather than the final second, clips grow in 10-second steps to a cumulative 40 seconds, and 360p previews render up to 60 per cent faster at a third of 720p's cost. Builders can draft cheaply and reserve 4K renders for final cuts.

Nvidia has paused deals in the AI Compute Partnership, a July initiative offering cloud providers credit support in exchange for half of the revenue earned above an agreed hourly rate. Its quarterly filing discloses $36 billion of commitments under agreements typically running six years, and staff raised antitrust concerns about telling providers whom they could rent to. Anyone pricing GPU capacity now faces unsettled financing terms.

OpenAI has published a four-part call to action on cyber defence, warning that a limited window exists before AI-enabled attacks become far more widespread and sophisticated. It assigns duties to four audiences: every organisation, cybersecurity and technology partners, governments and frontier AI companies, naming longstanding bugs, excessive permissions, misconfigurations and weak authentication as the accumulated exposure. The bar it demands covers what you buy, build and deploy, AI-generated code included.

Hugging Face put the Microduck on sale on Thursday, a 25-centimetre-tall duck robot at $399 that arrives before Christmas. It waddles, crouches, rights itself after falling and lifts up to 800 grams with its beak, sensing via a camera, lidar and two IMUs; the SDK, simulation and full reinforcement-learning stack are on GitHub. Behaviours train in simulation and deploy straight to the robot, cheap enough for real physical-AI experiments.

Transfyr came out of stealth this week with $25 million in seed funding to attack the hidden factors that make experiments succeed or fail. In its Cambridge lab, scientists wear headband cameras, three more watch each work station, and they narrate into microphones rather than keep notebooks, while the model writes its own commentary, down to pipetting counts. Tacit lab technique is becoming training data at millisecond resolution.

A study ran Claude Code and Codex through 288 runs on 17 real tasks and found CLAUDE.md and AGENTS.md context files made no clear difference to correctness. They did change behaviour: warned that a repository's test suite was slow, Claude ran more targeted tests and wasted less time. He now runs Claude Code's /doctor every few weeks and keeps these files to what a model cannot infer from code.

Chroma's Foundation memory layer runs a swarm of agents editing one wiki, where a batch lasts minutes and its read set is discovered by searching. Its Fission protocol is two-phase locking with wound-wait deadlock prevention and no rollback: a wounded transaction keeps the pages it already wrote and retries on its original timestamp. Aborting costs minutes and tokens of re-reasoning, so the prototype trades whole-wiki atomicity for keeping that work.

An engineer at exe.dev spent six months under one rule, never finishing a stuck agent's code himself but fixing whatever the agent was missing instead. Work moved onto disposable Linux VMs, about 20 at peak, with agents running unsupervised under botd, a manager he vibe-coded and never read. That manager collapsed under its own weight this month, which is his argument for settling architecture before the agent writes code.

An audit of 2,500 examples from BIRD Train, a text-to-SQL training set, put the gold reference query wrong in 52.1 per cent of cases. LLM and human review produced BIRD-Platinum, and fine-tuning Kimi-K2.6 on it with reinforcement learning from verifiable rewards hit 88.55 per cent on Arcwise-Plat-SQL against 86.75 for GPT-5.6 Sol Ultra. Human parity of 92.96 per cent needed 16 samples and majority voting, at $0.56 per task.

a16z argues that pricing an application in tokens imports the model provider's cost structure and anchors it to a falling unit. In its survey of 50 technical AI buyers, 27 preferred credits tied to recognisable work and 14 preferred tokens; Clay's 2026 memo says it mispriced Pro credits in 2022 and ran that segment at a loss for years. Price the highest layer of value you can measure and defend.

OpenAI's case study puts Asana's Enzyme removal at $12,000 of model and infrastructure cost against a roughly $6 million, five-year alternative. The Pragmatic Engineer reports the five years was when an opportunistic migration would realistically have finished, and the $6 million was a back-of-envelope per-file estimate. AI makes these migrations practical, Airbnb moved 3,500 test files in six weeks, but the saving is measured against work nobody was doing.

AI is remarkably good at translating an idea into instructions, but no AI reliably answers which instructions should exist, so the bottleneck moves from writing speed to how much complexity an organisation can understand. A 3,000-line pull request of generated code is still a 3,000-line addition to the system's cognitive surface, and a passing test suite is not ownership. His rule is never ship code you do not own.

The scarcity of consumer AI companies comes down to token costs, because every request now carries real inference cost where a pre-AI consumer website did not. The author's personalised news-site eval cost about $1 a run on Sonnet-class models and about $0.10 on gpt-5.6-luna, which he clocks at roughly 100 tokens per second. He expects frontier demand to keep compounding while fast, cheap, good-enough models absorb routine business work.

Stripe's agreement to buy OpenRouter for more than $7 billion reads as a move on the spend side of AI, not routing. OpenRouter carried a $1.3 billion valuation three months earlier and says it processes more than 10 trillion tokens a day across 400-plus models. Stigg's CTO frames every AI company as running two economies, customer revenue and token cost, with Stripe now holding a seat inside the second.

Mehta's case is that there is no objectively good taste, only curation judged against your intended audience, and that slop accumulates when a creator lets AI make those choices. Better models raise the defaults, but a default is still a default, so the tells shift rather than disappear. His prescription is to use AI aggressively for exploration and cautiously for decision-making, replacing ask-generate-accept with frame, explore, critique, select, refine.

agent-manager is a Go and tmux session manager for coding agents, Apache-2.0 licensed, installed with brew on macOS, Linux and WSL. Space spawns an agent from one sentence or replies into a blocked one without attaching, tab cycles between claude, codex, grok, gemini and others, and ctrl+r opens a diff whose line comments go into the agent's pane. Reach for it when you run more agents than tab switching tracks.

agent-device is Callstack's CLI and MCP server that lets AI coding agents inspect, control and verify apps on iOS, Android, HarmonyOS, TV, web, macOS and Linux. Agents read token-efficient accessibility snapshots instead of reasoning over screenshots alone, act through refs and selectors, and save evidence for review; the MIT-licensed CLI needs Node.js 22.12 or newer. Reach for it when an agent edits mobile code it cannot run and check.

RealDiff is an MIT-licensed CLI that finds runtime behaviour changes source review misses, running the same tests on both sides of a pull request and comparing method arguments and return values. It covers .NET, Java, Node, Go, Rust and Python, reruns the base to subtract self-noise, and reports the first changed behaviour whose descendants behaved identically. Reach for it, at early preview, when a refactor looks purely local.