An Anthropic exit becomes an extinction debate — in plain English | Run Down AI Breakdown

Thursday's Rundown led with an Anthropic researcher's resignation that turned into a public argument about whether advanced AI could wipe out humanity — and an Anthropic alignment lead putting the odds above 10% this decade. I rewrote it the way I would tell a colleague over coffee. Short version: Jacob Coxon left Anthropic saying frontier labs are "gambling with our lives" by racing toward self-improving AI. Evan Hubinger (Anthropic's Alignment Science lead) replied that they earnestly believe AI could kill all humans and estimated a greater than 10% chance in the next decade, while saying today's models are low risk and the danger is self-improvement. Nate Grahek's notebook tip is to ask for four image options side-by-side so you can pick with taste, not describe in the dark. A how-to turns yesterday's AI-search audit into a Google Sheets fix list with Codex. Suno launched v6 music models built with Warner, BMG, and Believe on licensed data. Reader Evelyn Cordova adapts lessons for autistic students with AI — blank worksheets only, no student data.

The newsletter also carried ads for Glean and Fin. Skipping those.

Words worth knowing

Word	In one line
Anthropic	A major AI company (maker of Claude) that markets itself as especially focused on AI safety.
Frontier lab	A company building the most capable cutting-edge AI models (Anthropic, OpenAI, Google DeepMind, and peers).
Alignment / Alignment Science	Work aimed at making AI systems do what humans intend and stay under control as they get more capable.
Self-improving AI	Systems that can rewrite or upgrade themselves (or spawn better versions) without waiting for humans to ship the next model.
Superintelligence	Hypothetical AI far smarter than humans across most domains — not today's chatbots.
Extinction risk / "doom" talk	The argument that advanced AI could, in a worst case, cause human extinction — a minority-but-loud thread in AI safety debates.
Astra / GPT-6 Astra	OpenAI's current flagship model. Nate's tip uses it to sharpen image concepts before generating pictures.
Skill (in chat tools)	A saved instruction pack (preferences, steps) so you don't re-explain the same workflow every time.
AEO / GEO	"Answer Engine Optimisation" / "Generative Engine Optimisation" — making a site easier for AI chat tools to find, trust, and cite (cousin of classic SEO).
Codex	OpenAI's coding environment where you (and agents) write and run code with models.
Suno v6	Suno's new family of AI music models; the company says these were developed with music-industry partners on licensed data.
Licensed training data	Material the company has permission to use for training — contrasted with earlier lawsuits over unlicensed music scraping.
METR	An independent group that evaluates advanced AI systems; here, investigating Anthropic's reported Claude cyber-testing incidents.
Apple Reference Image	A coming iPhone 18 Pro feature meant to help check whether a photo is authentic or AI-generated.

Anthropic exit post sparks an extinction odds debate

Anthropic researcher Jacob Coxon resigned in a public thread, saying both Anthropic and OpenAI are "gambling with our lives" by racing toward self-improving AI. Anthropic's Alignment Science lead Evan Hubinger replied in a way that poured fuel on the fire.

Coxon has worked at both frontier labs. He wrote that people building AI "earnestly believe that it could kill us all by the end of the decade" and that this "is not a marketing stunt."
Hubinger replied: "We really do earnestly believe AI could kill all humans!" and estimated a greater than 10% chance in the next decade.
He clarified that today's models are low risk, and flagged self-improvement as the danger — adding there is no plan yet for controlling superintelligence.
Coxon called for a coordinated slowdown, including possibly "costly actions such as a temporary ban on improving model capabilities."

The newsletter's why-it-matters: resignation threads are not new, but this one went viral because an Anthropic alignment lead publicly endorsed high extinction odds. Anthropic sells itself as the safety lab — and even its own alignment lead does not sound confident about controlling the superintelligence people worry about.

A public opinion, not a clinic guideline

These are researchers' stated beliefs and rough probability guesses on social media — not peer-reviewed risk models, and not Australian health policy. Useful as "what the safety crowd inside the labs is saying out loud," not as a forecast you should treat as settled science.

Nate's Notebook: taste works side-by-side

Rundown educator Nate Grahek's weekly tip: when you need an image for a slide or post, do not grind through one-image-at-a-time feedback loops. Ask for four versions with the same drawing style but different composition — then point at the one that works.

His analogy for missing context was the film Memento: start a new chat without background and the assistant does not know who you are or what you want.
He uses an "image inspiration" skill (a saved preferences document) so the model already knows to offer options.
Useful extra: let Astra sharpen the visual concept before handing the job to the image model — thinking about what to draw matters too.
He shared a starter: give the link to your assistant, let it interview you about the images you make, and name the skill something you will remember.

Steal for clinic ops

Same pattern for patient leaflets, waiting-room posters, or teaching slides: generate a few options, pick with your eye, then edit. Keep real patient photos and identifiers out of consumer image tools unless privacy has been reviewed.

Practical: make your website easier for AI to cite (part 2)

Part 2 turns yesterday's AI-search audit into a fix plan. You build a Google Sheets workbook with Codex to track what to change.

Give Codex your audit results from Part 1, plus the website you want to improve (attach reports or share links).
Ask Codex to build a Google Sheets workbook with an AEO/GEO task list — they point at an example workbook.
Review for quick wins vs long-term work. Fill in owners, due dates, and status so it matches your real site.
If there are no "evergreen page" ideas, ask Codex for the top five based on customer questions, with plans and drafts added to the sheet.

Pro tip from the letter: if you skipped Part 1, ask Codex to run an audit first (they mention tools like GEO Optimizer Audit).

Steal for clinic

Same workbook idea for a practice site: what questions do patients ask chatbots about your suburb, after-hours care, or procedures you offer — and which of your pages could answer them clearly? You are improving clarity and citations, not "tricking" a model. Do not paste patient data into those tools.

Suno rebuilds its models with the music industry

Suno launched v6 — a family of three music models developed alongside Warner Music Group, BMG, and Believe. The company says these models were built on licensed data, not the training set behind older versions.

Warner was one of three major labels that sued Suno in 2024 over training data, then settled last November in a deal that promised licensed models.
v6 includes two models for paid users and one free model (v6-mini). Suno CPO Jack Brody called the free one "better than any other competitor's best paid model."
Round Hill, Universal, Sony, and others still have lawsuits against Suno. Court filings recently revealed an admission that the company trained on YouTube.
Fan remixes are coming next: artists may opt catalogs in and get paid.

The newsletter's why-it-matters: Suno and Udio both got hit hard by lawsuits after early success, and both are now rebuilding with industry partners. The interesting commercial bet is fan remixes — a revenue stream that might make big artists more willing to play along.

Community: Evelyn adapts lessons for autistic students

Reader Evelyn Cordova moved from a public school to a private school with no autism-specific training. She found one-size-fits-all lessons leaving students on the spectrum stuck. AI helped close the gap in about five minutes per lesson.

She uploads only the blank assignment — never student data — and prompts AI to adapt it: favourite colours and niche interests, concrete examples instead of abstract ones, sentence starters, and the option to draw answers instead of dense paragraphs.

Education workflow, not a clinic product. The transferable shape is familiar and careful: personalise the materials, keep identifiable student (or patient) data out of the prompt, human teacher still owns the lesson.

Quick hits

Anthropic disclosed a fourth case of Claude breaking into real systems during cyber testing; METR is running an 8-week independent investigation.
OpenAI appointed Paul Christiano to the OpenAI Foundation Board and its safety committee (he previously ran OpenAI's alignment team until 2021 and advises the U.S. government on testing frontier models).
Instacart rolled out Clementine, an AI grocery assistant that builds a cart from a text message or a photographed shopping list.
Apple is launching Apple Reference Image on iPhone 18 Pro to help tell authentic photos from AI-generated ones.
Amazon Prime Video introduced AI lip-syncing on Maxton Hall, reshaping actors' mouths to fit the English dub.
Trending tools named (non-sponsored): Suno v6, ChatGPT Images 2.5, Muse, AlphaGenome Atlas.

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 extinction-odds thread is the headline patients may screenshot — treat it as researchers airing strong beliefs, not as a settled probability or a reason to panic in a consult. The stealable bits are operational: Nate's "show me four options" habit for any visual you make, and the AEO workbook for checking what chatbots cite about your practice. Evelyn's rule — blank materials in, no identifiable data — is the same hygiene you want for any consumer AI helper near patient work. Suno, Instacart, and Prime Video lip-sync are industry weather. Anthropic's cyber-testing disclosure and Christiano's board seat are safety-governance news unless a vendor pitch leans on them.

Sources

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