7 min read · 14 sources

Anthropic has spent the last year teaching Claude to write code, browse the web and call APIs. Today it taught the model to drive a laser. The Model Hardware Standard, shipping as a research preview aimed at scientific labs, is a set of standardized drivers plus a common data format that lets AI agents talk to microscopes, cameras and other lab instruments without bespoke glue code. Pair it with the Model Context Protocol and an agent can update instrument parameters mid-run, recover from hardware faults, or just take the wheel in plain English. Anthropic says the goal is to compress weeks or months of experimental setup into “hours or minutes,” and the bet is that lab robotics, like software before it, eventually ships with drivers that already know how to talk to a competent agent.

Cutting one byte per entry was worth 250GB; cutting the per-entry footprint in half across 250 billion entries freed roughly 100TB of fleet RAM.

Nvidia's circular money is load-bearing

Source: wsj.com ↗

Nvidia is funnelling capital back into the AI labs that buy its chips, and the WSJ frames this not as a circular-financing scheme but as deliberate infrastructure. Frontier labs are growing faster than their balance sheets and credit profiles can support, and the only thing keeping the flywheel spinning is ever-greater compute. Nvidia can borrow at better rates than any of its customers, so the company is effectively fronting the capex those labs cannot yet finance themselves. The bet is that the labs eventually fund their own growth, and the question is what happens if they do not.

The Microduck is real, $399, and open source

Source: techcrunch.com ↗

Hugging Face has a duck robot. The Microduck stands 25cm tall, costs $399 and ships before Christmas, with an SDK, a simulator and a full RL training stack on GitHub. Built by Pollen Robotics, which Hugging Face acquired in April 2025, it can waddle, pick up 800g with its beak, self-right, crouch and roller-skate, with camera, lidar and two IMUs for sensing. Behaviours trained in simulation deploy straight to the hardware. The timing is loud: Nvidia is reportedly set to buy Hugging Face at a $13B valuation, so the world’s largest chipmaker is about to be in the duck business.

An AI for the invisible half of every protocol

Source: nytimes.com ↗

Transfyr came out of stealth with a $25M seed and a thesis that the reason experiments fail has less to do with the published method and more to do with what scientists do with their hands. The startup’s system, covered by the NYT, ingests video, audio and instrument logs and tries to extract the tacit variation that no protocol captures. Its early finding is blunt: lab workers carry out the same protocol in wildly different ways, and those differences are doing real work. If the model can name them, the replication crisis in wet-lab biology gets a quieter, more useful adversary than statistics.

Your AGENTS.md has a half-life

Source: addyo.substack.com ↗

Addy Osmani argues that agent configuration files, the CLAUDE.md, AGENTS.md and skills directory entries that steer coding agents, decay. Models improve, harnesses gain features, codebases change, and instructions that were correct six months ago are now confidently wrong. His practical advice: run Claude’s /doctor every few weeks, review memory separately from skills, and force every line in the file to earn its slot again. Recent research is harsh on personalised skills, showing inconsistent value even when the abstraction is sound. The mental model worth holding is that of a config file as a living system, not a one-time setup.

The $5.9M Asana migration, audited

Source: blog.pragmaticengineer.com ↗

OpenAI published a case study claiming Asana saved $5.9M by using Codex to migrate from Enzyme to React Testing Library in two weeks for about $12K. Gergely Orosz read it carefully. The frameworks are not drop-in: Enzyme operates on component instances, React Testing Library operates on the rendered DOM, so the migration is a real rewrite, not a search-and-replace. His read is that the prior four-engineer-five-years estimate was inflated, but the underlying point still holds: AI is making migrations that were previously too expensive to schedule finally worth scheduling. The original post is gated, and the published excerpt leaves the dollar figure under-defended.

Meta is Anthropic's frenemy with a $10B line item

Source: nytimes.com ↗

Mark Zuckerberg published a 6,500-word essay warning that leading AI labs are consolidating power while warning the rest of us about doom. The NYT notes that Meta is internally projected to spend as much as $10B a year on Anthropic’s services, making it one of the lab’s largest customers. The essay did not name Anthropic, and Anthropic did not name Meta. The money, on the other hand, is very specific.

Small models crossed a line

Source: calv.info ↗

Calvin French-Owen writes that gpt-5.6-luna runs at roughly 100 tokens per second and turns a research thread that used to cost about $1 into something that costs tens of cents. He flags GLM 5.3 as another point on the same Pareto frontier. The interesting consequence is economic: traditional ad-supported consumer apps assume near-zero marginal inference cost, and at a dime per request rather than a dollar, personalised daily news, always-on assistants and the long tail of consumer AI start to add up to actual businesses. Frontier models (Fable 5, 5.6 Sol) are still preferred for serious coding work, but the floor of “good enough” just moved.

Stripe paid $7B for an AI bank

Source: thefinancialengineer.substack.com ↗

The acquisition closed on August 16 at more than $7B, roughly 5x OpenRouter’s $1.3B Series B from three months earlier. OpenRouter is misread as a thin proxy; it actually sits between 80+ model providers, independent inference vendors and 10M+ developers, routing over 10 trillion tokens a day across 400+ models. The author’s case is that this is marketplace infrastructure, not plumbing, and that Stripe’s pitch is to become the bank every AI company is currently forced to build for itself. Stripe’s DNA is marketplace settlement, so the fit is closer than “payments company buys a router” sounds.

Agent swarms are not databases

Source: trychroma.com ↗

Chroma’s Robert Escriva argues that multi-agent systems are distributed systems problems, and borrowing transaction semantics from databases is the wrong move. Agent “transactions” run for minutes and cost real money to retry, so abort-and-restart discards more useful work than it saves. Git-based conflict resolution is also wrong for agent-written prose, since natural-language edits get summarised, reorganised and rewritten in ways that do not line up with line-based diffs. The metric Chroma settled on is “goodput”: the share of paid reasoning that survives rather than being thrown away on abort. The implication is that the next round of agent infrastructure looks more like a distributed runtime than a database with extra steps.

Hypergrowth is still accelerating

Source: epochai.substack.com ↗

Epoch AI updates the revenue run-rate chart that has become the single most-watched number in tech. OpenAI tripled from $13B to $40B in the last year. Anthropic went from $1B to $9B in 2025, then more than tripled again in Q1 2026 to a reported $65B by end of July. Combined, the two labs grew from $30B to $105B in 2026 alone, a 3.5x year. The conventional model says growth at this scale has to slow to 2x, then 50%, then 20-30% annually. The data says it has not. Whether that is a sign of durable economics or a temporary pull-forward is the open question.

Harness engineering: the missing layer

Source: habitat-thinking.github.io ↗

A term from Birgitta Boeckeler at ThoughtWorks, harnessed here into a fuller definition. A coding harness is not a test harness; it has to verify invariants tests cannot see, like architectural decisions, naming conventions, security constraints and structural rules, because AI assistants produce plausible code that drifts quietly. The three components are context engineering (a HARNESS.md capturing stack and conventions), agent-based review, and periodic entropy checks. The framing is that the harness is not a constraint on what gets written but a continuous verification mechanism around the model.

Cloudflare cut 100TB of DNS bloat

Source: blog.cloudflare.com ↗

The Big Pineapple platform that powers 1.1.1.1, Gateway DNS, DNS Firewall and AS112 holds over 250 billion cache entries, so a single wasted byte per entry is more than 250GB across the fleet. Five successive changes to how CacheKey and CacheEntry are laid out in memory cut per-entry footprint by more than 50%, freeing roughly 100TB, which is the equivalent of 130 modern servers’ RAM. Insert throughput improved 43% and lookup latency dropped 19% on top. The largest gains landed in ECS-heavy locations, where the savings compound. Cloudflare published the struct layouts, the benchmark methodology and the workload mix (56% A, 25% AAAA, 19% TXT) so anyone running a similar cache can copy the wins.