Ephemeris · Issue 023 Mon · 11 May 2026 · Zürich
Ephemeris · Issue 023 01 / 05 · Proof of Concept
Design · Systems

Your design system is now an inference system.

David Hoang argues design systems have stopped serving humans. Their new audience is the model that assembles a UI on the fly: components must convey intent, not aesthetics; documentation must be machine-readable; governance becomes a feedback loop, not a review meeting. Adoption is the wrong metric — measure adaption, how the system evolves under live pressure.

"Components stop being parts and start being parameters. The model picks the geometry; the system supplies the constraints."

— David Hoang · 26 Apr 2026

Ephemeris · Issue 023 02 / 05 · Stanford via @ProductsAndStartups
Security · Research

When the human stops reading, the bugs stop hiding.

Stanford instrumented six thousand coding sessions and found a clean step-function: code generated by an autonomous agent and shipped without a human in the loop carried roughly nine times more security defects than code where a human and the model went back and forth. The kicker — pure-AI code is also more likely to land, because nobody is pushing back.

9×
Vulnerabilities per 1k lines of shipped code, agent-only mode vs. human-in-the-loop. n = 6,000 sessions.
Human + AI
AI only
Survives to commit44%
Ephemeris · Issue 023 03 / 05 · via @denissexy
[ DEV TOOLS · AGENTS ]

Hand the agent a browser. Keep the cookies.

ChatGPT Portal is a tiny local bridge: your authenticated browser session stays on your laptop, the agent gets a sanitised text snapshot of the page it asked to see. It can read your inbox, your bank, your private GitHub — but it cannot click submit, change a setting, or exfiltrate a token.

$ portal start --port 7331 ▸ bridge online · 127.0.0.1:7331 ▸ session cookies stay in browser · agent sees text only agent> open https://mail.example.com/inbox [ok] rendered → 12,418 chars · 4 unread threads [note] form submission · DENIED by policy [note] outbound POST · DENIED by policy agent> summarise unread invoices
Ephemeris · Issue 023 04 / 05 · via @denissexy
Side Projects · ML

150,000 Show HN posts. Now your title gets a verdict.

A new tool ingests every Show HN since the section's first day, embeds the titles, and learns what predicts upvotes. Paste your draft launch and it returns a forecast: expected karma, comment count, and three rewordings it thinks will outperform yours. Either it is right, or it teaches you something about how the audience reads.

Title A
14 ↑
Title B
38 ↑
Title C
72 ↑

Test before you post

Ephemeris · Issue 023 05 / 05 · Anthropic
Infrastructure · Capacity

Five gigawatts of Trainium, and Claude lands inside Bedrock.

Anthropic and Amazon committed to up to five gigawatts of compute over the next decade and roughly a million Trainium2 chips already in the cluster. The bit that matters for builders: the full Claude Platform arrives inside AWS Bedrock with the same account, same controls, same billing — no second contract, no second key vault.

5GW
Total commitment over the next decade. Trainium2 in Q2 · Trainium3 later in the year · ~1 GW online by year-end.
Chips today
>1,000,000 Trainium2
Run-rate revenue
$30B (was $9B in 2025)
Where it lands
AWS Bedrock · unified billing
End of issue 023 Back to top ↑

That's today.

A Monday on autonomy. A piece of design philosophy from Hoang, a piece of Stanford research that ought to slow somebody down, two small dev tools surfaced via @denissexy, and the AWS deal that quietly changes where Claude lives.

Sources
proofofconcept.pub · arxiv.org · github · foresyn.ai · anthropic.com
Rubric
Tools you could adopt this week · research with a kernel · agentic dev tools · capacity
Issue
023 · Mon 11 May 2026 · Zürich