Sun · 03 May · MMXXVI Today's eight.
Ephemeris — a daily field manual for senior engineers and founders.
Workflows that follow the tenant.
Cloudflare ships Dynamic Workflows: a library that lets a platform route durable execution into the tenant's own code — multi-day jobs that survive restarts, with idle costs kept near zero. Useful if you've ever owned a Temporal cluster you wish you didn't.
Before
Central worker pool. Tenant code marshalled in. One bad customer can starve the others. Idle workflows still cost.After
Workflow steps dispatched into per-tenant Workers. Crash isolation by construction. Dormant tenants → near-zero spend.Triage as an act of inclusion.
GitHub put a continuous AI loop on its accessibility feedback queue. The agent reads, dedupes, classifies, and routes — humans get back the time they used to spend sorting, and spend it fixing.
Next Halloween, I'm dressing as O(n²).
A wry tour through quadratic time as it shows up in the wild — the loops that look fine at one tenant, fall over at ten, and only confess under a profiler. Read it before your next "it's slow only in prod" ticket.
| # | Smell | Symptom | 10× | 100× |
|---|---|---|---|---|
| 1 | nested loop over input | fine in dev | 100 | 10,000 |
| 2 | list .indexOf inside loop | p99 creep | 100 | 10,000 |
| 3 | repeated re-render of N items | typing lag | 100 | 10,000 |
| 4 | quadratic JSON merge | OOM at 5pm | 100 | 10,000 |
Google sells out.
A week-of-GPT-5.5 dispatch from Zvi Mowshowitz: a labs-arms-race scoreboard updated nearly to the hour, plus what each release actually changes for the people building on top.
Zvi's weekly AI roundup is the closest thing the field has to a serious paper of record. This week: GPT-5.5 lands hard, and the post traces how the new system card lines up with the messier reality of real-world use.
The headline take is on Google: a string of product moves that, taken together, look less like measured iteration and more like a company finally choosing to ship. Whether that's brave or reckless depends on where you sit.
Worth reading for the synthesis as much as the news. By the time you finish, you'll have the right cluster of facts in your head to argue any of the three plausible positions on what comes next.
Models prefer their own writing.
A new arXiv paper documents self-preference bias: when an LLM grades resumes, it scores ones written by the same model 20–60% higher than equivalent human-written ones. Hiring loops where one LLM writes and another judges are now optimizing for prose style, not signal.
Claude, on the film set.
Higgsfield wires its video-generation toolkit into Claude over MCP. Pitch a campaign in chat, get back shot lists, motion control, and competitor teardowns — without leaving the conversation. The interesting part isn't the demo; it's that creative tools are starting to ship as protocols, not GUIs.
Your assistant, your hardware.
MF0.ai is a self-hostable AI assistant: bring your own LLM keys, store memory locally, watch the model's working set in plain sight. For anyone who has felt uneasy about pasting client work into someone else's chat box, this is a reasonable evening project.
- To
- Operator, your-laptop.local
- From
- Editorial desk · Ephemeris
- Re
- One assistant, no third parties
- CC
- Your favourite model providers
The most important charts in the world.
Zvi crowdsources the question every analyst secretly asks: which single chart, if you saw it once, changes how you think about the world for the rest of your career? The thread is a reading list disguised as a poll.
That's today.
Issue 015 of Ephemeris — Sunday, 3 May 2026, Zürich. Eight stories from seven sources, picked against the rubric: AI tools you could adopt this week, dev tooling and agentic coding, performance, research with a practical kernel, anything immediately useful to a senior engineer or founder. Pure marketing skipped on principle.