SAT · 16 · MAY · MMXXVI Sixteen.
Today's issue is about the rooms where the agent doesn't want you to look — the brokerage, the council vote, the gigawatt that nobody can find, the channel it just walked into without knocking.
Your bank, in the box.
ChatGPT Pro now lets US users hook their actual checking, brokerage, and card accounts into the chat. Ask it whether you can afford the flight; it knows. The marketing copy is "tailored guidance"; the operational reality is that a model with a system prompt is now sitting between you and your money, every day, with read access.
WHAT SHIPPED · 15 MAY
- Account linking via Plaid-style flow, gated to ChatGPT Pro in the US.
- Tailored answers grounded in your real balances, holdings, and recent transactions.
- Read-only — no transfers, no trades — for now.
- Same trust-and-safety pitch as the rest of GPT-5.5: opt in, scoped, revocable.
Beijing pulls the agent's plug.
Andrew Ng's weekly digest leads with a quiet story you'll see again: Chinese regulators have rejected Meta's bid to ship its agent stack to the mainland, citing data-egress and content-control gaps the Llama family was never built to satisfy.
Issue 353 packs three other pieces worth ten minutes apiece. Washington's NIST has begun pre-release evaluations of the next round of frontier models — the same arrangement the UK AISI piloted last spring, now formalised under the federal AI Action Plan.
The medical-AI desk reports on a multi-site mammography trial in which a deployed model matches radiologist sensitivity at lower recall rates, with the usual caveat that the comparison is against radiologists working under time pressure, not under their best conditions.
And Ng's own column introduces "AI Andrew", a personality-tuned assistant trained on his lectures and writing — a small but pointed answer to the "who owns your voice when the agents are reading it back" question.
May, in a spreadsheet.
Zvi's monthly roundup is the closest thing to a tape of the month — the chyron version of every story he didn't grant its own post. The cell that should reassure you, and probably won't: "we will almost certainly not have another pandemic."
| row | topic | verdict |
|---|---|---|
| A04 | Another pandemic, near-term | unlikely |
| A11 | The "AI is plateauing" thesis | premature |
| A19 | State-level AI bills as net-positive | mostly no |
| A26 | Education systems' response to GPT-5.5 | still lost |
| A33 | Crypto market as a leading indicator | still useful |
| A48 | "Vibe" of the May discourse | tired but lucid |
A gigawatt nobody can find.
Ed Zitron walks through Microsoft's claim of adding one full gigawatt of data-center capacity per quarter — and then tries, line by line, to locate the physical buildings. The conclusion is uncomfortable: most of the "new capacity" is small upgrades, leased racks, and offshore facilities that never made the press release.
Ship faster. Sleep less.
Bayram Annakov sketches the new boardroom argument: the CEO has watched the agents three-x velocity and wants the team to commit; the CTO has read the diff and wants verification, taste, and a way home from on-call. Neither is wrong. The post lays out what each side has to give up.
RECTO · The CEO
"We can ship in a week what used to take a quarter. If we don't, somebody else will, and they'll be the case study."
Wants: throughput, demos, board-deck momentum.
VERSO · The CTO
"The agents wrote the code. Nobody on the team has read it. The next outage will be the one we cannot debug."
Wants: tests, taste, time to absorb the diff.
Guest bots walk in.
Bot API 10.0 introduces "guest bots" — a bot you can summon into any chat without making it a member. Nikita Rvachev points out the obvious move: wire a Claude Code instance to Gmail, Calendar, Drive, Todoist and Obsidian, and the rest of your stack now follows you into every group thread.
@claude summarise this thread, file
the decisions in Obsidian, and add
"reply to legal by Friday" to Todoist.
> done. 1 file, 1 task, 1 calendar
> nudge for Thursday 14:00.
The hook score is the new CTR.
Higgsfield ships a 15-second-cut analyser that grades short video on four axes the platforms used to keep to themselves. It is the kind of tool that will quietly rewire how the next cohort of creators thinks about a cut — not what's pretty, but what scores.
Viral potential composite of the three below, weighted by surface.
Hook score likelihood a viewer holds past second 2.
Retention curve predicted % still watching at each second.
Brain-activation heatmap per-region engagement estimate, frame by frame.
Claude grows down.
Anthropic's first proper SMB tier — not Team-with-a-discount, an actual product. The differentiators are operational: smaller seat minimums, a bookkeeping-friendly billing cycle, and an admin surface that does not assume you have an IT department.
The wager: there is a long tail of 5–50-person shops that buy software like a household — one credit card, one decision-maker, no annual contracts. Until now those shops have been on personal Pro accounts pretending to be enterprises.
The deeper bet is on lock-in. Get a 14-person agency onto Claude for their copy, briefs, and bookkeeping, and the per-seat upgrade pitch in three years writes itself.
That's Saturday.
Eight items for the senior engineer or founder with ten minutes before the first coffee. A money product, a regulator's wall, a monthly's ledger, an investigative bar-chart, a boardroom argument, a Telegram primitive, a virality grader, and a Claude tier sized for an agency.
Sources today: OpenAI · DeepLearning.AI · Don't Worry About the Vase · Where's Your Ed At · Anthropic · Telegram · Higgsfield, surfaced via @seeallochnaya · @ProductsAndStartups · @rvnikita_blog · @TochkiNadAI.
Rubric: AI tools to adopt this week · creative software · dev tools & agentic coding · privacy & security · research with a practical kernel · anything immediately useful before standup.