Today the tools finally know their job. A cheaper agent that stays on task. A media stamp you can verify. A research suite that reads the literature while you sleep. And a benchmark the experts couldn't pass.
The agent gets cheaper. The leash gets longer.
Cursor's Composer 2.5 fine-tunes Moonshot's Kimi K2.5 on twenty-five-times more synthetic tasks and ten-times more reinforcement compute. Cursor claims it stays focused across long-running work; the proof will be a Friday with twelve open tabs and a clean diff.
Sign the picture. Stamp the sound.
OpenAI is wiring Content Credentials and SynthID into ChatGPT, Sora, and the image API, then shipping a public verifier so anyone can check whether a clip was machine-made. The chain is C2PA-signed and watermarked in the same pass — you cannot brush one off without the other showing.
Read the literature in your sleep.
Google folded three prototypes under one label: Co‑Scientist for hypothesis generation, AlphaEvolve and ERA for computational discovery, and Literature Insights via NotebookLM. The bundle plugs into Antigravity's Skills system with native handles for UniProt and AlphaFold.
For the working scientist the press release is the wrong primitive — what matters is whether the hypothesis you get back at 9 a.m. is one a peer reviewer would entertain. Google is betting the lab notebook is the unit that gets agentic first.
The handles into UniProt and AlphaFold do most of the load-bearing here. Skills attach domain databases to the tool-call layer; the model stops bluffing about a protein it has never seen and looks up the structure. The literature-insights piece, meanwhile, lets a postdoc walk into a new field with a 30-minute brief that names the open questions and the cliques that hold them.
None of this replaces the bench. It replaces the part of the bench that involves reading PDFs at 11 p.m.
An agent that won't go away.
Gemini Spark sits inside Gmail and Docs and acts without being asked — drafting replies, scheduling, sweeping the inbox while you sleep. Google calls it "proactive 24/7 help." Whatever else it is, it is the first mainstream agent designed to keep running between sessions.
▸ daemon up · 03:14 zürich
▸ watching: gmail · docs · calendar
▸ swept 23 threads → 6 needed you · 17 archived
▸ drafted: re: Q3 review · awaiting send
▸ moved: 11 events · no conflicts
▸ next check: +12m
▸ user idle 07h 41m
The bench the experts couldn't beat.
Anthropic shipped BioMysteryBench: 99 wet-lab-grounded bioinformatics puzzles graded against experimental validation rather than published consensus. Claude matched human expert panels on the solvable items — and cracked roughly one in three problems the experts could not, sometimes by inventing a novel analytical path.
- Bench size
- 99wet-lab-grounded bioinformatics puzzles, graded by experiment, not consensus.
- Human-solvable
- parClaude matched expert panels on the items humans were able to solve.
Stop launching. Start telling.
PostHog's playbook for shipping product launches without the choreography: tell customers first, talk to the press as if they were customers, skip the launch deck. The thesis is short — process scales worse than your audience does.
Supply is infinite. Cost is zero. Reprice.
Jeff Bullas, fifteen years into making a living explaining things on the internet, writes the obituary for that whole industry. When the marginal cost of content goes to zero, the price follows; the only thing left to sell is the scar tissue behind the lesson.
When supply becomes infinite and cost becomes zero, the market reprices the commodity fast and permanently. — Jeff Bullas, 19 May 2026
That's today.
Seven picks today, drawn from seven publications. Curated for what a senior engineer or founder can do something about by Friday.
Sources, today
Cursor's engineering blog; OpenAI's index; the Google AI blog; Anthropic Research; PostHog; Jeff Bullas's Jabs. Telegram channels were polled for cross-references and did not contribute primary URLs not already on this list.
Rubric, daily
AI tools you could adopt this week. Creative software. Dev tools and agentic coding. Privacy and security when the signal is high. Science with a practical kernel. Anything immediately actionable for a senior engineer or founder.
Today, in particular
A cheaper coding model. A public way to verify a watermark. A research bundle that pretends to read the PDFs for you. A benchmark that broke an expert panel. A marketing memo for the founders who hate marketing memos. And one obituary for the content business.
Issue 032
Wednesday, 20 May 2026, Europe/Zürich. Edited overnight, published at 08:00 local.