Ephemeris · Issue 022 Sun · 10 May 2026 · Zürich
Ephemeris · Issue 022 01 / 09 · OpenAI
Strategy · Product

Ads, with a label.

OpenAI is testing limited advertising inside ChatGPT, promising clear labelling, answer independence, and "strong privacy protections" to support free-tier access. Whether that survives contact with revenue targets is the test the next two quarters will run.

"Answers stay independent. Ads are clearly labelled. Privacy is preserved." — three claims a paid model would never have to make.

Ephemeris · Issue 022 02 / 09 · OpenAI
Privacy · Policy

What ChatGPT learns about you — and what you can keep.

OpenAI walks through what it stores, what it strips before training, and how to opt your conversations out. The doc is specific enough to act on tonight, not just bookmark.

Stored. Conversation text, model responses, account metadata, and per-message feedback. Retention windows differ by tier; enterprise data is excluded from training by default.

Stripped. The training pipeline runs deduplication, classifier-based PII removal, and topic-based filtering before a single token reaches the next checkpoint.

Yours. Plus and Free users can disable "improve the model for everyone" in privacy settings; temporary chats skip retention entirely. Worth one minute, once.

Ephemeris · Issue 022 03 / 09 · PostHog
Founders · Marketing

Stop subtweeting your rivals.

Public competitor takedowns reach the people who already love you and inadvertently introduce the rival to everyone else. PostHog's playbook is the un-glamorous one: honest comparisons, bid on the keywords, and put the energy back in the product.

Postmark NOT SENT Re: that smug LinkedIn carousel about $COMPETITOR
  1. Write the honest comparison page yourself; link to their docs.
  2. Bid on their brand keywords — quietly, not vindictively.
  3. Ship the missing feature; outcomes outlive screenshots.
  4. Praise them in public when they're right; readers notice.
Ephemeris · Issue 022 04 / 09 · PostHog
Tools · Analytics

Five real-time stacks, one sheet.

PostHog lines itself up against GA4, Plausible, Fathom, and Matomo on live dashboards, identified-user tracking, and replay. Read it once before you re-pick your default analytics — the picks have shifted since 2024.

ToolLive dashIdentified usersReplaySelf-host
PostHog
GA4
Plausible
Fathom
Matomo
Ephemeris · Issue 022 05 / 09 · Sentry
Observability · Agents

Watching the agents.A developer's handbook.

Sentry's guide to AI-agent monitoring is built for the people writing the agent loops, not for an auditor reading them after the fact. What to trace, what to alert on, and which compliance frameworks get it wrong.

? Trace the question, not the answer
Ephemeris · Issue 022 06 / 09 · via @seeallochnaya
Workflow · Agents

Specs first, agents second.

A growing camp of engineers writes a markdown spec, hands it to Claude or Codex, and reviews the patch — not the prose. OpenSpec documents the workflow as a small, opinionated convention you can adopt this afternoon.

01spec/ — one markdown file per change. State what, not how.
02agent run — agent reads the spec, drafts the patch, opens a PR.
03human review — you read code, not natural-language summaries.
04archive/ — merged specs live next to the diffs they produced.
Ephemeris · Issue 022 07 / 09 · via @denissexy
Creative · Tooling

Genternet rewrites the web in real time.

A Chrome extension that intercepts whatever you load and regenerates it through a fast image model and Gemini Flash-Lite. Two cents a page, fully open source, and the gag is also the demo: the medium is no longer the message.

$ genternet → grok-imagine-image · $0.02 / page
$ stack: chrome ext + gemini 3.1 flash-lite + diffusion image gen
$ source: github.com/DenisSergeevitch/genternet-chrome
$ verdict: an art piece that happens to be installable
Ephemeris · Issue 022 08 / 09 · via @denissexy DECLASSIFIED
Datasets · Curiosities

████ declassified UFO files, in markdown.

Denis Sergeevitch converted the U.S. government's released UFO archive — 4,000+ documents — into a clean markdown corpus and asked GPT-5.5 Pro to weigh in. Use it as a long-context reasoning eval; ignore the percentages.

Files4,063 markdown documents, structured by report-id.
SourceU.S. ██████ declassification project, 1947–2024.
Eval useLong-context retrieval; entity dedupe; date-extraction stress.
LicensePublic domain. Repo: github.com/DenisSergeevitch/UFO-USA.
Ephemeris · Issue 022 09 / 09 · Bullas
Business · Strategy

Truth, as a growth strategy.

Bullas argues authenticity has graduated from brand attribute to competitive moat — and that performative messaging is now a measurable drag on growth. A useful pressure check on whatever copy you most recently published.

Object · 2026.05 · Editorial
Truth is no longer the thing you put in the footnotes — it is the proposition itself. The companies that win the next decade tell the public what they are doing while they do it, and trust accumulates as a balance sheet item.
End of issue 022 Back to top ↑

That's today.

Nine picks, six sources, one throughline: what you choose to show. Today's feed leaned on the official OpenAI and PostHog blogs, Sentry's engineering site, two Telegram channels, and a long-running marketing essayist.

  • · OpenAI · openai.com/news
  • · PostHog · posthog.com/blog
  • · Sentry · blog.sentry.io
  • · @seeallochnaya · t.me/seeallochnaya
  • · @denissexy · t.me/denissexy
  • · Jeff Bullas · jeffbullas.com
Rubric: AI tools you could adopt this week · creative software · dev tools & agents · privacy & security · research with a practical kernel · anything immediately actionable.