Thursday's Rundown was Meta and Google joining Anthropic's launch week — plus a safety scare about how OpenAI's next model "thinks." I rewrote it the way I would tell a colleague over coffee. Short version: Meta shipped Muse Spark 1.3, a near-top model that Zuckerberg says is almost too cheap to meter. Google shipped Gemini 3.8 Flash, a bounce-back after a rough year, still a notch below the frontier. A report says OpenAI's Astra uses repeated "thinking loops" that can hide readable reasoning from safety monitors. Nate's tip is blunt: your AI ceiling is your tech literacy. A how-to turns one AI workflow into a job-interview "proof project."
The newsletter also carried ads for Weights & Biases / CoreWeave, Render Workflows, Incogni, and a Claude skills guide. Skipping those.
Words worth knowing
| Word | In one line |
|---|---|
| Muse Spark 1.3 | Meta's new model. "Max" is the stronger setting. Zuckerberg pitches frontier-level performance at a very low cost. |
| Watermelon | Meta's codename for a larger follow-up model. Zuckerberg said it is "next up," plus open weights for Spark. |
| Weights / open weights | The saved numbers that make a model work. Publishing them means others can download and run it themselves. |
| Gemini 3.8 Flash | Google's newest fast / cheaper Gemini. Same price as 3.7; stronger on coding, reasoning, and agent jobs. |
| Intelligence Index / AA | Artificial Analysis's scoreboard for how "smart" a model looks on their tests. Spark Max 62; Gemini 3.8 Flash 59; Fable 5.1 still ahead. |
| Frontier | The current top-tier systems. DeepMind's Koray Kavukcuoglu admitted Gemini sits "a little below" that line. |
| Agent / agentic | An AI helper that can take steps (click, run tools), not only chat back. "Agentic tasks" = jobs that need those steps. |
| Astra | OpenAI's next big model, expected this week. The letter covers a reported technique behind its coding and computer-use skills. |
| Recurrent depth / loops | The model re-reads the same text in repeated internal loops before answering — extra "thinking" without making the model larger. |
| Monitorability | Whether humans (or safety tools) can still read what the model is thinking. Pure-math loop outputs are harder to watch than plain-English reasoning. |
| Tech literacy | Nate's phrase: basic fluency with your own machine — servers, cloud vs laptop files, organising things so you can find them. Not writing code. |
| Proof project | A short recorded walkthrough of one AI workflow you actually use, plus a five-slide deck, tailored to a job you want. |
| Human-in-the-loop | You stay in the process: the AI drafts or runs a step; a person still checks and owns the result. |
| Fair use | US copyright doctrine about when you can use someone else's work without permission. The US government brief argues limiting it for AI training would hurt US tech lead. |
| Quasar 438B | Spanish startup Multiverse Computing's first model. Score 43 on the Intelligence Index — tops European rivals per the letter, not the global frontier. |
Meta and Google join the launch party
Anthropic kicked off September with Fable 5.1. Meta and Google both shipped new models the next day — on very different vibes. Zuckerberg is pitching cost-plus-intelligence. Google is trying to show it still has a pulse after a year the letter calls one to forget.
Meta Muse Spark 1.3 (Max) lands at 62 on Artificial Analysis's Intelligence Index. That trails only Claude Fable 5.1 and Opus 5, per the letter, while being significantly cheaper to use. Zuckerberg's line: "frontier performance almost too cheap to meter." He also said "next up" is a larger model codenamed Watermelon, and that Meta will release the weights for Muse Spark (so others can download and run it).
Google Gemini 3.8 Flash keeps 3.7's pricing ($0.75 / $3.75 — input / output rates as the letter lists them) with gains in coding, reasoning, and agentic tasks. It slots in at 59 on the same index. DeepMind's Koray Kavukcuoglu admitted Gemini sits "a little below the frontier," and said being at the frontier is what matters for them.
The newsletter's why-it-matters: Meta is climbing on the blend of smarts and price that Google used to own. Watermelon could shake the industry if it lands. For Google, 3.8 is a step up, but with its Pro model seemingly out of the picture until Gemini 4, the pressure for a real frontier comeback is still on.
Why a GP should even care
Cheaper near-frontier models are how a practice tool's bill stops feeling silly. Spark at 62 and Flash at 59 are scoreboard numbers from Artificial Analysis, not TGA clearances. If a vendor swaps the model under your scribe or inbox helper, ask which one and what it costs per month of typical use — not which press release sounded louder.
Nate's Notebook: tech literacy is the ceiling
Rundown educator Nate Grahek's weekly note: the people getting the most from AI were already technical, or at least really tech-savvy. AI raises the floor. The ceiling stays low if you do not build basic tech literacy alongside AI fluency.
His habit: when someone says "webhook" or "environment variable," he no longer files it under "not my department." On the walk home he turns on voice mode and asks: explain that like I run the company but not the codebase. Re-explain at whatever altitude you need.
The bar he means is CEO-to-CTO level, not writing code: what a server is, which files live in the cloud versus on the laptop, and enough organisation to find one. If you closed that door years ago ("I'm not a technical person"), he says there has never been a better time to reopen it — with an explainer that will walk you through your own machine, including via screen recording or a browser agent.
Steal this for clinic
Same as learning what an MBS item actually pays, or how the secure-messaging path works, without becoming the IT person. Pick one term you have been nodding along to (API, cloud backup, "the agent failed") and ask an assistant to explain it at GP altitude this week.
Practical: the "Proof Project" for interviews
The how-to turns one AI workflow you already know into a short kit for a job application or interview.
- Pick one workflow you know well — something simple but useful (a reporting routine, a research process, a content pipeline).
- Record a five-minute Loom (screen + face). Walk through the solution, and emphasise the human-in-the-loop bit: where you still check and own the result.
- Paste the transcript, link, and job description into ChatGPT or Claude. Ask for a five-slide deck with the video on one slide, and the other slides explaining your thought process and how the workflow runs.
Pro tip
Record in one take. Light edits are fine; do not burn an evening polishing. Same habit as a teaching case: show the real workflow, including where the human stays in charge.
Report: Astra's loops and safety monitoring
The Information reports that OpenAI's upcoming Astra (expected this week) uses a technique that boosts coding and computer-use skills — and also keeps parts of its reasoning harder to watch.
Recurrent depth: the model analyses the same text in repeated internal loops before answering. Extra intelligence without making the model larger. The catch in the report: those loop outputs can be pure maths, not the readable chain-of-thought older models showed. That makes monitorability (can a human or safety tool still read what it is thinking?) harder.
OpenAI reportedly dialed the loops back so Astra still writes its reasoning out. Its Path to Astra blog promises extra monitoring of that reasoning at launch. Chief scientist Jakub Pachocki posted that he wants to "prevent a race into unmonitorability" from "confused reporting," while calling monitoring "fragile."
The newsletter's read: the alarm is less about Astra itself (OpenAI says it still shows its work) and more about a performance race that tempts labs to loop harder and monitor less. Pachocki also says current monitoring is already fraying for reasons unrelated to Astra.
A report, not a regulator's finding
This is The Information's reporting plus OpenAI's public replies. It should not change how you use any clinic tool tomorrow. The transferable idea is familiar: if a system takes actions on your behalf, you want a window into why — the same instinct as an audit log on an EMR change. "It thought in maths we cannot read" is a governance headache for labs; it is not a licence to trust or distrust any particular medical AI.
Community: a custom GPT for unified physics
Reader John Gaus built a custom GPT with the personalities of Einstein, Lorentz, Planck, and Compton. Starting from Einstein's 1920 Leiden lecture and equations like E=mc² and E=hf, he spent months pushing the model to challenge his assumptions, with tool calls and reference sites to write and test the maths. He says it now unifies equations from atomic to black-hole scales, and he summarised the work in a paper the GPT wrote for others to expand and test.
Hobby research with a named peer-review instinct (challenge the assumptions, test the maths). Not a clinic workflow — and not peer-reviewed physics just because a GPT co-authored the PDF.
Quick hits
- US Commerce Secretary Howard Lutnick told Axios "We trust Anthropic," saying the company is "back on the right side" of the relationship with government (after the earlier Pentagon blacklist fight).
- US government filed a brief supporting OpenAI in the New York Times copyright lawsuit, arguing that limiting fair use for AI training would cost the US its global tech lead.
- New York City public schools are banning AI through eighth grade. Mayor Zohran Mamdani rejected the idea that AI-powered education is "not only inevitable, but necessary."
- Spanish startup Multiverse Computing launched Quasar 438B, its first model, at 43 on the Intelligence Index — topping European rivals per the letter, not the global leaders.
- Trending tools named in the letter (non-sponsored): Gemini 3.8 Flash, Muse Spark 1.3, and Fal's H3 Max Turbo (video at twice the speed, half the cost, per Fal's post).
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. Meta and Google shipping near-frontier models at Flash prices is a cost story for whoever sells you a scribe or inbox helper — ask which model and what the monthly burn looks like. Nate's ceiling line is the most transferable: you do not need to code, but you do need enough tech literacy to brief and audit an assistant, the way you brief a registrar. The Proof Project is interview kit for people changing jobs; the shape (show one real workflow, keep the human in the loop) is also how you demo a clinic automation to partners without overselling it. Astra's "loops we can barely read" is a lab-safety headline about monitorability, not a reason to switch tools. John's physics GPT is a reminder that "the model wrote a paper" is the start of scrutiny, not the end. NYC banning AI in early school years is a policy choice patients may mention; it is not an Australian classroom rule.
Sources
- The Rundown AI, 3 September 2026 — Meta, Google join the AI launch party.
- Meta — Introducing Muse Spark 1.3 (near-frontier; Watermelon next; weights to be released).
- Google — Gemini 3.8 Flash and 3.8 Flash Cyber (same pricing as 3.7; coding/reasoning/agent gains).
- Artificial Analysis — Muse Spark 1.3 Max Intelligence Index 62.
- The Rundown — Nail job interviews with the Proof Project method.
- The Information — Report on OpenAI Astra recurrent-depth loops and safety monitoring concerns.
- OpenAI — Path to Astra (promises extra monitoring of reasoning at launch).
- The Rundown Community — John Gaus custom GPT for unified physics equations.
- Axios — Commerce Secretary Lutnick on Anthropic (“We trust Anthropic”).
- US government brief — support for OpenAI in NYT copyright lawsuit (fair use / AI training).
- NYC Mayor’s Office — AI ban in public schools through eighth grade.
- Multiverse Computing — Quasar 438B (Intelligence Index 43; tops European rivals per letter).
- Fal — H3 Max Turbo (video at twice the speed, half the cost).