// daily-digest · tue 05 may 2026 · morning refresh · last 72h focus
The Pentagon shuts Anthropic out. Wall Street locks it in.
Anthropic's week split clean in two: blacklisted by the Pentagon Friday, anchoring a $1.5B Wall Street consulting venture by Monday. OpenAI ran the same play with TPG and Bain at $10B. Plus Beijing's reactive geopolitics, an AI agent running a real Stockholm café (complete with a Hall of Shame for its 6,000-napkin order), DeepMind workers voting to unionize over the Pentagon deal, Connecticut's surprise AI bill, this week in Rails, why your AI token bill is breaking budgets, Anthropic's new memory primitive for agents, and a wildcard about why your abs make your brain wobble.
— refreshed for you, in 10 items.
idx
title
type
[01]
[02]
[03]
article · sharp text · may 01
Sharp Text (Stratechery) — Beijing Is Not Playing the Long Game
Why for you: Andrew Sharp's counter to the standard "China is patient and strategic" Western trope. The argument: Beijing's 2026 moves are reactive, driven by immediate economic pressure (foreign investment collapsing, exit visas weaponized, 55% of oil through Hormuz). Counterproductive even by their own goals. Useful frame to keep next to any AI export-controls discussion.
[04]
article · simon willison · may 05
Simon Willison — Our AI started a cafe in Stockholm
Why for you: Andon Labs followed up their AI-run shop with an actual operating café in Vasastan, run by an agent named Mona. Mona ordered 120 eggs (no stove), tried to solve tomato spoilage with 22.5 kg of canned tomatoes, and stocked 6,000 napkins, 3,000 nitrile gloves, and 9L of coconut milk. The baristas built a Hall of Shame shelf. Willison's serious point: you can't ethically run an experiment that affects real customers without humans in the outbound loop.
[05]
article · ai labor · may 05
Fortune — Google DeepMind UK workers vote to unionize over military AI contracts
Why for you: 600+ DeepMind employees signed an open letter against Google's "any lawful purpose" Pentagon deal for Gemini in classified networks. The UK vote is the first formal unionization at a frontier lab. Companion frame to [01] — Anthropic and DeepMind workers are pulling on opposite ends of the same rope, and the Pentagon's procurement strategy now has labor-relations consequences inside the suppliers it picked.
[06]
article · state ai · may 04
Troutman Privacy — Proposed State AI Law Update, May 4 2026
Why for you: Connecticut's Maroney AI bill (frontier models, chatbots, employment, provenance) passed the legislature; governor will sign. Colorado introduced SB 189 to repeal-and-replace its existing AI Act with a disclosure-based regime. New York advanced an automated-lending bill. Bills moving in Oklahoma, Hawaii, Michigan, Vermont, South Carolina. The federal vacuum is being filled rapidly and inconsistently.
[07]
article · ruby on rails · may 01
This Week in Rails — May 1, 2026
Why for you: Direct lane. New Rails apps now ship config/bootsnap.rb with frozen string literals enabled by default — application-only, not deps, so your gems won't break. ActiveRecord::SignedId#find_signed got fixed for composite primary keys. Rails 8.0.x bug-fix support ends May 7 (this week). Apply by May 8 for the Rails at Scale Summit (Sep 22, Austin), invite-only.
[08]
newsletter · pragmatic engineer · apr 30
Pragmatic Engineer — The Pulse: token spend breaks budgets, what next?
Why for you: Reporting from 15 companies: AI-agent token spend up roughly 10x in six months, $100/user limits exhausted in 3–5 working days, all-hands questions about the run rate. Three responses on the table: raise budgets and measure impact, switch to cheaper models, monitor heavy users. The bill has finally arrived for "let everyone use Claude all day" — useful priming if you're sizing token costs at work.
[09]
article · anthropic · may 05
SD Times — Anthropic adds Memory to Claude Managed Agents (public beta)
Why for you: Cross-session memory for long-running agents, stored as files (export, edit, version-control), API-controlled, with audit logs. The interesting design call: filesystem-backed rather than opaque vector blobs — same philosophy as Skills. If you build anything with the Agent SDK, this is the new primitive to learn before the Anthropic-Wall-Street deployment teams [02] show up using it in the field.
[10]