AI Builders Digest

AI diffusion, not capability, is the bottleneck for the next decade

The gap between how capable models are and how little they show up in GDP comes down to the laws of corporate physics: preparing data, reengineering processes, and agreeing on what the new workflow even is. Past that you hit the speed of the real world, waiting on a customer's response to a proposal, a permit, a drug pipeline that takes years. Levie also expects coding agents to go far wider than anyone assumed, into cyber defense, life sciences research, legacy system upgrades, and software for companies that could never afford it. His counterintuitive conclusion: cheaper code raises engineering leverage so much that we end up needing more engineers, not fewer.

• #agents
• #products

Claude Marketplace opens with CrowdStrike, Cursor, Factory, Gamma, and Vercel

Enterprises can now apply their existing Anthropic spend commitment toward buying Claude-powered products and agents from third parties. Companies building on the Claude Platform can get listed so customers have a path to purchase without a separate procurement cycle.

• #products
• #agents

The task-decomposition model of automation is missing a term

Two assumptions quietly do all the work in a well-designed report on AI and jobs: that jobs decompose into tasks, and that automation always reduces human labor or at best leaves it flat. Shipper's experience says the second is simply false, that automation often creates multiples more work for humans, and any honest economic analysis has to model work creation. On the first, task lists are a useful lens but not the full picture, because the model can't say where new tasks come from. Borrowing from Wittgenstein and Heidegger, he frames a job as a way of seeing and caring about the world: an editor notices what a nurse doesn't, and as tools change, the same cares generate entirely new tasks.

• #policy
• #evals

OpenClaw cloud sessions now run fast, with remote terminal and WebVNC

Cloud sessions in OpenClaw finally hit acceptable speed, shipping alongside a remote terminal, WebVNC, and CUA for computer use. The dashboards and mini-apps push from two months ago has since replaced most of the custom tooling the team built around OpenClaw, with everything collapsing into a sidebar entry, a dashboard, or a plugin on their team server.

• #agents
• #open-source

Vercel counts 8 price cuts and 16 model discounts in six months

Six months of pricing changes add up to eight price cuts, fee removals, and structure improvements, plus sixteen model discounts, with more promised. Rauch is also putting on an AI benchmarks event with Benchmark, pitching the companies that measure models as steering the industry toward skill, truth, and efficiency.

• #products
• #evals

Banked resets failed to apply in ChatGPT Work and Codex this morning

A bug during a window this morning meant banked resets did not fully apply for people using them in ChatGPT Work and Codex. Everyone affected is getting another reset plus an apology email, no support ticket required.

• #products

Testing a new model by asking it to name your blind spots

Yang's go-to check on a model's reasoning is to paste in a prompt asking for brutally honest feedback about the blind spots holding you back, then see how sharp the response is. Astra's answer came back especially blunt, which is the point of the test.

• #evals