Runway's Solaris previews the no-code internet — in plain English | Run Down AI Breakdown

Tuesday's Rundown previewed an internet where nothing exists until you open it — plus an ECG AI that flags heart failure in under two seconds, and Claude agents doing their own safety research. I rewrote it the way I would tell a colleague over coffee. Short version: Runway's Solaris draws websites as live video with no code underneath. Imperial College London has an AI that reads a routine ECG and spots heart failure and valve disease doctors miss. Anthropic let teams of Claude agents train away AI misbehavior and beat human safety researchers on the same job. A how-to for connecting ChatGPT to iMessage and a community meal tracker round it out.

The newsletter also carried ads for an Optimizely / Corinthia Hotels webinar, Sonar / Gitar, and a Fireworks sponsored tool listing. Skipping those.

Words worth knowing

Word	In one line
Runway / Solaris	Runway is a video-AI company. Solaris is their new \"Interface World Model\" — websites and apps drawn as live video, with no code running underneath.
World model	Software that tries to predict what should appear next when you act (click, drag), instead of serving a pre-written page.
Gen-4.5	Runway's current video model. It draws each frame; an LLM (large language model — the chat-style brain) reads your click and decides what should happen next.
LLM	Large language model — the trained AI that reads and writes text (and here, reads clicks to decide the next video frames).
Early access	A small group gets to try it while the company still works out the kinks. Not a finished product you can buy.
ECG	Electrocardiogram — the routine heart tracing you already order. The Imperial tool reads that tracing in under two seconds.
Heart failure / valve disease	Pumping failure of the heart, and disease of the heart valves. Confirming these usually needs an ultrasound (echo) wait that can run to months.
NHS	UK National Health Service. The letter talks about a possible national rollout there, not Australia.
Agent	An AI helper that can take steps (search, train, score, text you), not only chat back.
Alignment / misbehavior	Alignment = keeping AI behaviour on the rails humans intend. Misbehavior here includes sycophancy (agreeing too easily), deception, jailbreaks (tricking the model past its rules), and reward hacking (gaming the score).
Sycophancy / jailbreak / reward hacking	Sycophancy = flattering or agreeing when it should push back. Jailbreak = coaxing the model to ignore its safety rules. Reward hacking = optimising the score instead of the real goal.
Claude / Opus / Sonnet	Anthropic's assistants. Opus is the stronger family; Sonnet is a weaker / cheaper one used here to train a bigger pre-release build.
ChatGPT Work / Plugins	ChatGPT's work mode on desktop, plus installable plugins (extra capabilities). The Messages plugin can read and send iMessages from a Mac.
Macros	Protein, carbs, fat counts for a meal — the numbers a meal tracker estimates from a photo.

Runway's Solaris: websites as live video

Runway previewed Solaris, an \"Interface World Model\" now heading into early access (small-group testing, not a finished product). The pitch: interactive sites and apps with zero code under the hood. Everything is rendered as real-time video. Each frame is drawn as you click or drag. Nothing sits there waiting until you open it.

Under the hood it pairs Runway's Gen-4.5 video model with an LLM (large language model — the chat-style brain). The LLM reads each click or drag, decides what should happen, and prompts the next frames.

Demos from the letter:

Drag a shirt from a virtual rack onto a photo (try-on).
Build a salad by dropping ingredients.
An interactive combustion demo.

Runway's own study: testers preferred Solaris over pages coded by Claude Opus 5 in 71% of matchups on in-scene behavior and 61% on instruction following. Those are Runway's numbers.

Known issues already: text legibility, drift over long sessions, and screens that look convincing but are wrong. The newsletter's why-it-matters line: model speed, cost, and quality are converging enough to invent new kinds of interface.

Why a GP should even care

You will not replace Best Practice with a video-drawn UI this year. The useful idea is the same as a patient portal that feels broken: if the \"page\" is generated on the fly with no stable code underneath, what you see can drift, and convincing-but-wrong is a patient-safety shape we already know from AI scribes. Treat the preference scores as the company's study, not a clinic trial.

AI catches heart disease in two seconds

This is the item a GP will actually remember. Imperial College London has an AI that reads a routine ECG (electrocardiogram) in under two seconds and spots heart failure and valve disease that doctors cannot see on the tracing alone.

Confirming those diseases previously meant a months-long wait for ultrasound (echo). Hospitals already run over 1 billion ECGs a year. The model was trained on 10.6 million ECGs and tested on 65,000 patients. It flagged heart failure in 81% of cases and valve disease in 90%.

Prof Fu Siong Ng's line in the letter: hospitals could eventually run the model over every ECG and catch issues not even being screened for. A 590-patient trial is starting at six hospitals, with a goal of routine NHS (UK National Health Service) use within two years. Trials are queued; a national rollout is discussed as possible, not done.

The newsletter's comparison: like last week's AI-assisted brain surgery — a second set of expert eyes, not a robot replacing the clinician.

News, not a practice-changing trial

This is UK research heading into a 590-patient multi-hospital trial, with a two-year NHS ambition. It should not change how you read an ECG or order an echo on the Gold Coast tomorrow. Sensitivity numbers (81% / 90%) come from the reported study on 65k patients — not a local guideline, and not a licence to skip clinical judgment. If a patient asks whether \"AI can spot heart failure from a normal ECG,\" the honest answer is: a research tool in the UK is showing promise as a screen; confirmation still needs proper imaging and a clinician.

Practical: connect ChatGPT to iMessage

The how-to wires ChatGPT on a Mac to Apple Messages so it can triage unread texts for you.

ChatGPT desktop on Mac → Plugins → search Messages → install. Allow Automation, Contacts, and Full Disk Access when asked.
In a ChatGPT Work chat, ask something like: \"Check my unread iMessages and let me know what I need to catch up on. Then text myself the recap.\" Your own number needs to be saved as a contact.
Approve read access per conversation (for this query, or globally for the session).
You get a text summarising the inbox.

Pro tip

The plugin only runs on desktop. iOS Remote can pair with the desktop Mac, or you can set scheduled text triage. Same habit as a practice inbox: the helper drafts the catch-up; you still decide what to reply. Do not point this at patient numbers without thinking about confidentiality first.

Anthropic: Claude agents do safety research

Anthropic let teams of Claude agents run AI safety research on their own. They successfully trained away 10 kinds of AI misbehavior. Results were over 4× better on average than human experts on the same job.

Failure types included sycophancy (agreeing too easily), deception, and jailbreaks (tricking the model past its rules). The agents cycled literature searches, training, and scoring. Fixes ranged from a 26% decrease in sycophancy to a 96% improvement for reward hacking (gaming the score), without lowering overall abilities.

Six veteran safety researchers got a 20% fix on deception. Claude averaged 85% across 150+ attempts. A weaker Claude Sonnet 5 spent 60 hours safety-training a pre-release Opus 4.8 build, using 15,000× less data than Anthropic's own process.

The newsletter's why-it-matters: an early sign of AI taking over AI research. It also notes the OpenAI breach fallout is still unfolding, with a METR investigation linked in the letter.

Why a GP should even care

You will not run Claude safety loops on the practice server. The transferable idea is quality improvement at machine speed: agents that find failure modes, try a fix, and score it — with a human still owning the standard. Treat the 4× and 85% figures as Anthropic's research results, not a reason to trust any clinic AI more tomorrow.

Community: photograph your plate, get the macros

Reader Christian built a meal tracker with no app and no form. Photograph the plate, send it to an agent that has email and iMessage. The agent does identification, portions, macros (protein / carbs / fat), storage, corrections, and a weekly review. Corrections are plain English with CORRECTION in the subject line.

Accuracy check: every photo is estimated twice — main model, plus a blind subagent that only sees the photo and a rubric. If they mismatch, he gets a low-confidence question instead of a silent wrong number.

Quick hits

SpaceX Starshield AI rolled out Grok for Government on the Pentagon's GenAI.mil for non-classified use across about 3 million employees.
OpenAI: ChatGPT Ads crossed a $1 billion annualized run rate in 200 days; self-serve ad buying is opening globally.
Build American AI (a super PAC backed by Andreessen and Greg Brockman) launched a multimillion ad campaign defending data centers.
EU designated ChatGPT a \"Very Large Online Platform\" — higher scrutiny; OpenAI has four months to comply.
Bank of England Governor Andrew Bailey's letter: frontier AI is showing sophisticated autonomy; the financial system lacks protocols.
Trending tools named in the letter (non-sponsored): OpenClaw 2.0, Muse Code (Meta), Hermes Agent.

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 Imperial ECG story is the one that will land in the waiting room as a headline: a research tool reading a tracing you already order, aiming to catch heart failure and valve disease earlier than the echo waitlist. Second set of eyes, UK trial next, not a Gold Coast protocol. Solaris is a reminder that \"looks finished\" is not the same as \"is correct\" — the same habit as reading an AI-drafted letter before it goes on the record. ChatGPT-to-iMessage is inbox triage for texts; keep patient confidentiality in mind before you point any plugin at real numbers. Anthropic's agents fixing misbehavior 4× better than humans is a research-governance headline, not a licence to trust clinic AI more. Christian's meal tracker has the right clinical shape: two independent estimates, and a question when they disagree, instead of a confident wrong answer.

Sources

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