The people building AI want to slow down — in plain English | Run Down AI Breakdown

Monday's Rundown flipped the mood again. Last week, people inside the industry were warning that AI could kill us all. This week, several of the biggest lab bosses lined up behind the idea of slowing capability gains so safety work can catch up. I rewrote it the way I would tell a colleague over coffee. Short version: Anthropic CEO Dario Amodei published an essay arguing labs should deliberately pace the frontier; OpenAI's Sam Altman, Elon Musk, Google DeepMind's Demis Hassabis, and Microsoft's Satya Nadella publicly backed the direction, while Donald Trump dismissed slowdown talk as exaggerated. Staff at The Rundown shared their own AI workflows (story scouting and a running coach). A how-to shows how to draft a one-page proposal minutes after a sales call. And a randomized trial in five Chinese hospitals found an AI assistant helping sonographers catch more fetal brain malformations without raising false alarms. Skipping the Oracle and AWS ads.

Words worth knowing

Word	In one line
Anthropic / Dario Amodei	A major AI company (maker of Claude); Amodei is its CEO and often writes long public essays on AI risk and policy.
Frontier / pacing the frontier	The leading edge of AI capability. "Pacing" means deliberately slowing how fast those capabilities improve so safety checks can keep up.
Recursive self-improvement	AI systems helping build better AI systems, so progress can accelerate without proportional human effort — Amodei's core worry here.
Agents / agent swarm	AI programs that can take multi-step actions (not just chat). Amodei warned that in 6–12 months a swarm of agents might be able to take over much of the internet.
Third-party evaluators	Outside safety testers embedded inside labs to probe models independently, rather than labs only marking their own homework.
OpenAI / Sam Altman	Maker of ChatGPT; Altman said he would commit to outside evaluators.
Google DeepMind / Demis Hassabis	Google's leading AI lab; Hassabis backed the general direction of Amodei's pacing argument.
PAICS	A deep-learning tool trained to flag 10 specific fetal brain malformations in real time during prenatal ultrasound.
Sensitivity / false positives	Sensitivity = how often real problems are caught. False positives = healthy scans wrongly flagged. The trial raised sensitivity without more false alarms.
Sonographer	The clinician who performs and interprets ultrasound scans.
Zapier / Copilot (automation)	Zapier connects apps; its Copilot helps you describe a workflow in plain English so it builds the automation for you.
IPO	Initial public offering — floating a company on the stock market. Altman said OpenAI will not IPO in 2026.
Fields Medal / METR Evals	The Fields Medal is maths' top prize; 25 medalists warned AI maths races risk skipping understanding. METR Evals is an independent AI evaluation group that two safety researchers joined after leaving Anthropic and DeepMind.

Dario Amodei makes the case for slowing AI

Amodei published an essay arguing that labs should slow the rate of capability gains so safety can catch up. His core worry is recursive self-improvement: models already helping build the next models, so progress can run ahead of oversight.

He said AI is now speeding its own development, and in 6–12 months a swarm of agents could be capable of taking over the entire internet.
For pacing, he proposed embedding third-party evaluators inside labs, agreeing safety standards among democratic labs, and coordinating even with authoritarian regimes.
Fellow leaders agreed in public: Sam Altman committed to evaluators, Elon Musk said "Dario is right," and Demis Hassabis and Satya Nadella backed the direction.
Donald Trump dismissed slowdown advocates as "negative forces" pushing "things that won't happen," arguing the U.S. should keep its lead because "whoever wins AI wins."

The Rundown's take: this rare show of unity, alongside recent maths leaps and researcher warnings, hints that unreleased lab models may already be showing unusually strong capabilities. Slowing down still will not be simple — some want open development, some call pacing "regulatory capture," and others do not want to slow at all.

The Rundown Roundtable: staff AI use cases

Two staff examples from The Rundown community:

Darren (media): built a morning story-scouting workflow. It maps past short-form coverage, scans newspapers, tech blogs, Rundown letters, papers, and company updates, then keeps only stories with a clear visual hook, a 45–90 second explainable idea, broad interest, and real novelty. Ranked briefs arrive as one Slack DM.
Kenny (sales): built a running coach in a Claude Project linked to Google Calendar. For six months it has planned his week, logged watch data, flagged mileage ramps that risk injury, tracked shoe wear, and rewrote the calendar when life disrupted training.

How-to: one-page proposal after a sales call

The letter's training piece walks through auto-drafting a proposal after a call (useful pattern for any templated follow-up, not only sales). Full guide: Send a one-page proposal minutes after every sales call.

Keep meeting notes in one AI-notes folder. Build a one-page Google Docs template with placeholders (prospect name, needs, and so on).
In Zapier, ask Copilot to: after a sales call, pull notes, have AI extract the fields, fill the template, and draft a prospect email.
Set the last step to draft in Gmail, not send automatically.
Always read the draft first — in their tests, one run left the AI's own editorial comment about budget sitting in the proposal instead of a clean number.

Pro tip from the letter: log each run to a Google Sheet (time, prospect, company, key needs) for a simple audit trail.

AI helps catch more fetal brain defects in hospitals

A randomized trial across five Chinese hospitals, published in The Lancet Digital Health, found that giving sonographers an AI assistant during live prenatal ultrasounds raised detection of specific fetal brain malformations without more false alarms.

The tool, PAICS, is a deep-learning system trained to flag 10 specific fetal brain malformations in real time — one of the first tests of AI actually in the scanning room.
With PAICS, sonographers' detection sensitivity rose from 78.6% to 87.3%, with gains holding across most subgroups and false positives staying steady.
Human judgment mattered: PAICS alone often scored worse than the assisted sonographer, who correctly overrode about 60% of the AI's mistakes.
Limits: high-risk pregnancies only, brain malformations only, all sites in China, and AI-assisted scans took about 40 seconds longer.

The Rundown's framing: while public AI talk fixates on existential risk, healthcare detection keeps advancing — here as a second set of eyes beside the human, not a replacement.

Quick hits

Community: reader Thy Toeung used Claude Code and Fable to build a phone gardening app for his wife (local data, climate-zone settings after friends elsewhere wanted it). His write-up.
Tools named: Smaug Flash (Abacus open-weight fine-tune), SWE-2 (Cognition coding model inside Devin), ElevenLabs Music v2.5, Meta's Muse personal agent.
Sam Altman told Fortune OpenAI will not IPO in 2026 — calling now an "ill-advised moment" given safety concerns.
25 Fields Medalists signed a declaration warning that AI races to solve maths problems risk sacrificing understanding, attribution, and the human process that advances the field.
The Pentagon is considering a $5B loan to AI cloud startup Fluidstack for U.S. data-center supply chains (WSJ).
At the BRICS Summit in New Delhi, China proposed an open-source AI community for shared development across member nations.
Anthropic and Google DeepMind safety researchers Joe Benton and Josh Engels resigned to join METR Evals, both warning that capabilities are outpacing safety.

What this means for a GP

None of this is clinical advice, and none of it should change how you treat a patient tomorrow. It is why a busy doctor might still skim the letter. The Amodei slowdown essay is the kind of screenshot patients may share after a news cycle about AI risk — treat it as public lobbying among lab bosses, not as a clinical protocol, and remember Trump's public pushback sits in the same news cycle. The fetal-ultrasound trial is the most clinic-adjacent item: AI as a real-time second reader that raised sensitivity without more false alarms, but only for specific brain malformations in high-risk pregnancies in China, with humans still correcting most AI mistakes and scans taking a little longer. Darren and Kenny's workflows are the stealable ops pattern — map what you already cover, filter for novelty, and put guardrails on anything that schedules your week. The sales-proposal Zap is the same idea for admin: draft automatically, send only after a human read.

Sources

Plain-English notes on The Rundown AI · Dr Kotha · Gold Coast GP. Not medical advice. Source: The Rundown AI.