Tuesday's Rundown was mostly about OpenAI putting its own coding agents to work on research — plus an early hint that an AI-designed lung drug might also nudge "biological age" down. I rewrote it the way I would tell a colleague over coffee. Short version: OpenAI published numbers on how hard its internal agents work (about 3.1 agent workdays for every human one), says it hit Sam Altman's "automated research intern" goal for September, and researchers are burning hundreds to thousands of dollars a day in agent tokens. Insilico Medicine reported that rentosertib, an AI-designed drug for idiopathic pulmonary fibrosis, left a small trial cohort looking younger on six blood-based aging clocks. A how-to builds a family planner app with Astra. An NBC poll finds most U.S. adults more worried than excited about AI, across both parties. Reader Sanjay built a poker luck detector with Claude.
The newsletter also carried ads for Unwrap and Pave. Skipping those.
Words worth knowing
| Word | In one line |
|---|---|
| Coding agent | An AI helper that can write and run code, open tools, and take multi-step jobs — not only chat answers. |
| Automated research intern | OpenAI's name for an AI that can do intern-level research work. Altman wanted this by September 2026; a fuller "automated researcher" target is March 2028. |
| Tokens | The little text chunks models read and write. More tokens = more compute cost. "Agent tokens" means the usage bill for those helpers. |
| Astra | OpenAI's new flagship model (also called GPT-6 Astra). The letter says internal use before public release was a big edge. |
| Frontier lab | A company building the most capable cutting-edge AI models (OpenAI, Anthropic, Google DeepMind, and peers). |
| Idiopathic pulmonary fibrosis (IPF) | A lung disease where scar tissue slowly stiffens the lungs and breathing gets harder. "Idiopathic" means the cause is unknown. |
| Rentosertib | A drug Insilico's AI helped design for IPF. The aging signal came from blood samples in a small prior trial. |
| Aging clock | An AI or statistical model that estimates "biological age" from markers (often in blood). Different teams build different clocks. |
| UK Biobank | A huge UK research database of health and genetic data used as a comparison pool here. |
| Codex /product brief / mockup | Codex is OpenAI's coding environment. A product brief is the written "what to build." A mockup is a design sketch before real code. |
| Data center | A warehouse of computers that run AI (and other) services. Locals often worry about power, water, noise, and land use. |
| World model | An AI that tries to simulate how the physical world works (physics, scenes, cause and effect), not only chat or make clips. |
| Open-source model | Weights and (often) code released so others can run or fine-tune them. MiniCPM5-2B is a small open model from OpenBMB. |
OpenAI opens the books on its AI research intern
OpenAI published a look at how much of its own research its coding agents handle. Agents logged about 3.1 workdays for every human workday. The company says it hit the "automated research intern" goal Sam Altman had set for September.
Details named in the letter:
- Altman posted the timeline last October: intern-level system by this month, fully automated researcher by March 2028.
- A typical OpenAI researcher now burns $600+ a day of agent tokens; the 90th percentile is above $7,000. Token output is up 124× since December.
- Around 80% of researchers use 4+ agents at once. Experiments per researcher are said to be at an all-time high inside the company.
- Agent tasks are getting harder and longer, with more success on complex jobs and a surge in tech troubleshooting help.
The newsletter's why-it-matters: OpenAI's Tibo Sottiaux called internal use of Astra before public release the "biggest competitive advantage." Labs that already have unreleased frontier models — and the budget to burn tokens on them — compound that lead in ways outsiders struggle to match.
A lab ops report, not a clinic blueprint
These are OpenAI's own usage numbers about its researchers. They do not mean your practice should run four unsupervised coding agents on patient data. The transferable idea is familiar: helpers that can act need spend caps, logs, and a human who still owns the outcome — same instinct as capping who can bulk-export from the EMR.
Insilico's AI-designed drug hints at slowing aging
Insilico Medicine published trial-linked data on rentosertib, a drug its AI designed for idiopathic pulmonary fibrosis. On all six "aging clocks" they checked, treated patients read as biologically younger — an early signal on a small sample that the drug might affect aging itself, not only the lung disease.
- Rentosertib treats IPF (scar tissue slowly choking the lungs). Insilico's AI picked the target protein and designed the molecule.
- Last year's trial included 42 patients. Blood samples from that work fed the aging analysis, compared against a broader UK Biobank pool.
- Six aging clocks (models from separate teams that read age markers from blood) rated treated patients younger; one group's estimates fell about 2.7–3.5 years.
- The dose with the clearest aging signal was not the same as the dose with the biggest lung-function gains — a hint the aging effect may not be "just feeling better because the lungs improved."
The newsletter's why-it-matters: Anthropic CEO Dario Amodei has argued that a real medical breakthrough can shift public AI sentiment more than marketing. Stories like this are early, small, and not a prescription — but they show why AI-designed drugs are racing toward markets.
Read this like a journal club, not a script
Small n, secondary aging endpoints, clocks that estimate biological age rather than prove "you get X more years." Interesting signal. Not a reason to start or stop any therapy tomorrow, and not TGA advice. If a patient brings the headline in, "early research on a lung-fibrosis drug, not an anti-aging pill you can get at the chemist" is the honest frame.
Practical: turn a family scheduler idea into an app with Astra
The how-to walks from a Workflow Hub idea to a working family planner in Codex with Astra — brief, mockups, then a build you can click-test.
- Find an idea in the Rundown University Workflow Hub and adapt it (they turned Shawn's eldercare log into a family planner).
- Open Codex, select Astra, and use
/planto ask for a product requirements document covering calendar and layouts, with multiple-choice questions to clarify details. - Ask for four design mockups (desktop and mobile side by side), then pick a direction before building.
- Tell Astra to build the chosen design, open it in its browser, add an event, reload, and switch users to check that saves and sharing work.
Pro tip from the letter: ask Astra to find five skills on GitHub that could improve the app's design, then choose which to install.
Steal for clinic
Same shape as prototyping a shared on-call or leave calendar: start from a clear brief, lock the design before code sprawl, then test "does it save and can the right people see it." Keep real patient identifiers out of consumer coding agents unless privacy has been reviewed.
U.S. voters' rare bipartisan issue is fearing AI
NBC News released a poll of 7,105 adults on how Americans feel about AI. About 70% feel more worried than excited, and the concern spans both major parties — even as more people say they use the tech.
- 52% use AI very often or sometimes (up six points from June 2025), but only 18% trust AI-generated information most or almost all of the time.
- Local data centers are unpopular: 69% against one nearby, 45% strongly against, versus 31% who would somewhat or strongly support.
- 70% believe AI is costing people jobs. Majorities expect harm to schools and elections. Science and healthcare were the only areas with a net positive view.
- 81% say Washington's AI rules fall short. Asked which party they trust on AI policy, 44% say neither (Democrats 20%, Republicans 16%).
The newsletter's why-it-matters: industry CEOs keep watching sentiment, and the poll is still moving the wrong way for them. Data centers have become the public face of the fight, and backlash is turning into a campaign topic. Voters also do not trust either party to handle it.
Why a GP should even care
U.S. midterm polling is not Australian regulation. The useful bit is the trust gap: people use AI more and trust AI-written answers less. That matches what you already hear in consults — patients paste ChatGPT output and still want a human to say whether it applies to them. Healthcare showing up as a rare "net positive" domain is reputation to protect, not a blank cheque for every vendor pitch.
Community: Sanjay's poker luck detector
Reader Sanjay Singhal built a luck-detection app after a rough poker month. He wondered whether bad cards or bad play was the real problem. No off-the-shelf tool measured luck from hand histories, so he started in a Claude chat (not as a coder). Claude built the UI in the chat and helped with formulas for luck on dealt cards, on the flop, and when going all-in. He says he fixed two major leaks and felt better knowing bad luck was a big part of the story. He later moved it to a full web app with Claude Code, and is working toward an iPhone app he may offer free.
Hobby analytics, not a clinic product. The transferable shape is "I had a gut feeling, I wanted a score I could audit, I iterated with AI until the maths matched the question" — familiar if you have ever wanted a simple dashboard for something your PAS does not track well.
Quick hits
- U.N. rights chief Volker Türk said he shares fears that advanced AI could pose an "existential risk," and that a few men hold "almost unlimited power over AI."
- ByteDance (TikTok's parent) is reportedly building a world model on top of its Seedance video AI, with founder Zhang Yiming involved and a possible October launch.
- OpenBMB open-sourced MiniCPM5-2B, a ~2-billion-parameter model that ranks first among open models under 4B on AA's Intelligence Index.
- OpenAI and Microsoft face a new lawsuit from the Seattle Times and Newsday over alleged illegal training; the outlets say AI is breaking journalism's economics "beyond repair."
- Trending tools named (non-sponsored): MiniCPM5-2B, Grok Bot Marketplace, Gemini 3.8 Flash, Muse Voice Transcribe.
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. OpenAI's agent-spend report is a reminder that the labs with the newest models and the fattest token budgets pull further ahead — relevant only when a vendor brags about "same model the labs use" without saying which snapshot, with what audit trail. The Insilico aging signal is the one patients may screenshot: small trial, secondary clocks, IPF drug first — interesting, not a lifestyle prescription. The NBC poll's trust gap (more use, less trust) is the consult-room reality. Astra family-planner builds and Sanjay's poker tool are the stealable ops pattern: narrow problem, human check, measurable output. U.N. warnings, ByteDance world models, and newspaper lawsuits are industry noise unless they show up as a patient worry or a partner email.
Sources
- The Rundown AI, 8 September 2026 — Inside OpenAI's agent-powered research boom.
- OpenAI — Research acceleration: a view inside OpenAI.
- Nature Biotechnology — Insilico rentosertib aging-clock analysis.
- The Rundown — Turn a family scheduler idea into an app with Astra.
- NBC News Decision Desk poll (DocumentCloud) — September AI topline.
- NBC News — Poll: polarized America unites on deep concerns about AI.
- The Rundown Community — Sanjay Singhal poker luck detector.
- Reuters — U.N. rights chief on AI existential risk.
- Bloomberg — ByteDance world model / Seedance.
- Hugging Face — OpenBMB MiniCPM5-2B.
- CourtListener — Seattle Times / Newsday lawsuit filing (OpenAI and Microsoft).