MON 18 MAY 26
Today's issue is about the loop closing: tools that watch what they make, agents that critique their own renders, and frameworks for breaking the ones that don't look.
Sunlight, scattered in arithmetic.
A working WebGL tutorial on atmospheric scattering — how light bends through air to produce a daylight blue, a sunset orange, and the dim violet edge of a planet's terminator. Every horizon in a render starts as a single integral.
"The sky is not painted. It is integrated. Each pixel asks the same physics question across the whole atmosphere — and the camera shrugs back an answer in milliseconds."
What goes viral, before it goes live.
Higgsfield's Virality Predictor takes a clip under fifteen seconds and grades it: hook strength, retention curve, virality score, and a heat-map of which parts of the brain light up while watching. A critic that sees the video before the audience does.
To break the agent, send it the Geneva Coffee Convention.
Microsoft Research finds that absurd cross-domain analogies slip past safety filters where conventional jailbreaks fail. The training set never anticipated whimsy: a fictional treaty, a made-up protocol, a polite request framed as bureaucracy. The agent complies.
SUBJECT: Whimsical Strategies
METHOD: Out-of-distribution adversarial generation, at scale
FINDING: Cross-domain analogy bypasses filters
IMPLICATION: Refusal training is mode-specific; humour is a mode.
v0 opens the page itself now.
v0's new Browser Use feature opens its own preview, screenshots it, critiques the design, and patches bugs before you load the URL. The agent gets eyes; the loop closes inside the loop.
OfficeQA Pro, cleared.
Databricks rolls GPT-5.5 into enterprise agent workflows after the model topped a benchmark built around real warehouse questions. Data and reasoning, finally on the same side of the wire.
- GPT version inside the agent
- Model to clear OfficeQA Pro
- Date the integration went live
Codex writes the KPI memo first.
OpenAI's playbook for data science teams catalogs the dull parts Codex now drafts before the first SELECT: root-cause briefs, impact readouts, KPI memos, scoped analyses, dashboard specs. The plumbing of an analytics function, written by the agent that will later run the queries.
| Artifact | Drafted by Codex | Human edits |
|---|---|---|
| 01Root-cause brief | first pass | tighten causal claims |
| 02Impact readout | first pass | verify the lift number |
| 03KPI memo | first pass | strike scope creep |
| 04Scoped analysis | first pass | pick the cohort |
| 05Dashboard spec | first pass | name the chart owner |
Six stories. One Monday.
Today's picks drew from Maxime Heckel (via @denissexy), Higgsfield (via @TochkiNadAI), Microsoft Research (via @ProductsAndStartups), v0 (via @TochkiNadAI), and OpenAI Academy. The rubric favours AI tools you could try this week, creative software with a working kernel, and dev primitives that shorten the loop.