Introduction
Welcome to today's Daily Pulse from Nicolas's AI Lab - the AI briefing for busy professionals, founders, and business owners. Around 6 minutes. Straight to what matters.
The theme right now is speed and cost. New models are cheaper, faster and starting to render interfaces the way video streams render frames. Healthcare and enterprise are already feeling the effects, while the money question behind AI's giant compute build-out grows louder.
Today at a Glance
Runway's Solaris generates full interfaces frame by frame, no code underneath.
An AI model spots heart disease from a routine ECG in under two seconds.
ChatGPT Ads hit a $1bn annual run rate in under 200 days.
Runway's Solaris renders software like video
Summary: Runway unveiled Solaris, an Interface World Model built on its Gen-4.5 video model paired with a language model. Instead of writing HTML, CSS or JavaScript, it predicts and renders each next screen based on what the user clicks and drags. In tests, 250 users preferred Solaris over code-based interfaces in more than 60% of instruction-following cases and close to 70% of natural-behaviour cases. It is early access by request only and still struggles with text clarity and drift in long sessions.
Why it matters: If this holds up, interfaces could adapt to each user on the fly and agents could work with any layout, changing how software is built.
No underlying code means interfaces render in real time like a video feed.
Consistency, state, accessibility and security remain open problems.
AI catches heart disease in two seconds
Summary: Researchers at Imperial College London built an AI model that detects heart failure and valve disease from routine ECGs in under two seconds. Diagnosing these conditions usually meant waiting months for an ultrasound. The model was trained on 10.6 million ECGs and tested on 65,000 patients, hitting 81% accuracy for heart failure and 90% for valve disease. A trial is running at six hospitals with the aim of reaching the NHS within two years.
Why it matters: Hospitals perform over a billion ECGs a year, so a fast screen on existing tests could catch disease far earlier without new equipment.
The tool works on tests hospitals already run every day.
Six-hospital trial underway, NHS rollout targeted within two years.
Cheaper, smarter models pile pressure on the field
Summary: A wave of releases focused on filling gaps rather than opening new frontiers. Google's TimesFM-3 forecasting model now reads multiple data streams at once and tops several benchmarks, though it carries a non-commercial licence. Anthropic's new Claude Fable 5.1 and Mythos 5.1 cut costs by around 25% for regular use and up to 45% for automated pipelines, with fewer false alarms and a privacy mode that keeps enterprise data off training. DeepSeek added vision to its model at no extra cost.
Why it matters: A genuine cost war is forming, and free upgrades like DeepSeek's vision push rivals to match on price and transparency.
Claude's cheaper cache reads slash costs for automated workflows.
DeepSeek adding vision for free resets expectations on pricing.
Source: https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/
An AI TV channel that never stops
Summary: A developer built Infinite Slop, an interactive livestream that generates endless content from viewer suggestions. It runs on a tuned version of MiniMax H3, nicknamed Max, which can render video faster than people can watch it. The stream even has a live news segment that scans X and Hacker News for breaking stories. Running it without sponsorship would cost an estimated $4,000 a day.
Why it is interesting: Real-time video models are cheap and fast enough to power always-on, audience-driven entertainment that did not exist before.
Viewers steer the scenes and the stream generates them live.
Running costs of roughly $4,000 a day show the compute bill.
Star in your own 15-minute movie
Summary: Koyal Experiences lets anyone drop themselves and their friends into a 15-minute interactive film that changes based on the choices they make. It is part of a broader surge in instant AI video, including TikTok-style apps that generate fresh clips as you scroll and prompt-as-you-go video games.
Why it is interesting: Interactive, personalised video suggests entertainment where the audience shapes the story in real time.
The film evolves with your decisions as you watch.
Part of a fast-growing wave of interactive AI video apps.
Source: https://xp.koyal.ai/
A $399 robot duck that roller skates
Summary: Hugging Face and Pollen Robotics opened preorders for the Microduck, a 25cm open-source biped that costs $399. It weighs 800g, runs on 15 motors, and carries a camera, a LiDAR sensor and two inertial sensors. It ships with seven trained moves, including walking, sitting, kicking, grabbing with its articulated beak, and roller skating once the skates are fitted. The SDK, the simulator and the full reinforcement learning training stack are on GitHub under a permissive licence.
Why it is interesting: Cheap, open hardware is where a lot of robotics learning actually happens, and this puts a trainable biped and its entire training stack at hobbyist money.
The whole reinforcement learning stack is open, not just the robot.
$399 puts a research-style platform within reach of individual developers.
AI Tools
Clipto: Search your local computer files by describing what you need, with most processing done privately on device. clipto.com
Snapdown: Turns any part of a Mac screen into clean markdown for writing and documentation. snapdown.com.au
Descript: Edit podcasts and videos simply by editing the transcript text. descript.com
Gemini 3.5 Transcribe: Converts speech into neatly formatted text across more than 85 languages. blog.google
FnScribe: Private, on-device Mac dictation triggered by the fn key. github.com/AlgorithmicResearchGroup/fnscribe
Expert Prompt of the Day
Context: Once Claude is connected to Gmail or Google Drive through Connectors, you can ask it to draft replies and reorganise files while keeping human oversight. It confirms before taking any action.
Prompt: Read the latest messages in this email thread and draft a reply that answers each open question, keeps my usual tone, and flags anything I need to decide before sending. Show me the draft first and do not send anything without my confirmation.
Example use case: A manager clears a backlogged inbox by having Claude draft context-aware replies to client threads, then reviews and sends each one.
ChatGPT Ads reaches a billion-dollar run rate
Summary: OpenAI's ChatGPT Ads has hit a $1 billion annualised revenue run rate in under 200 days, backed by tens of thousands of advertisers. Ad buying is expanding across Europe, India, the Middle East and North Africa. OpenAI says ads are picked from interests in the current chat and do not influence answers or expose private conversations.
Why it is important: A fast-growing ad business gives OpenAI a major new revenue line alongside subscriptions and enterprise deals.
The $11 trillion compute question
Summary: An estimated $11 trillion AI infrastructure build-out is projected through 2029, and analysts are split on whether it is justified. Optimists argue demand is underestimated, while sceptics warn subsidies and cross-company funding may mask weaker real demand. Some suggest OpenAI and Anthropic alone could consume half of this compute.
Why it is important: Whether this spending meets genuine demand or inflates a bubble will shape the whole industry's stability.
OpenAI cuts off Cursor after its sale to SpaceX
Summary: OpenAI told SpaceX it will wind down the contract that supplies its models to Cursor, the AI coding tool, with a proposed shutoff date of 12 November. SpaceX had just acquired Anysphere, the company behind Cursor, and OpenAI invoked the change-of-control clause in their agreement. Its stated reason is that it cannot be confident SpaceX will stay inside its terms of service, citing earlier contract breaches by Musk-owned companies. OpenAI says it is giving the maximum notice the contract allows, and that Cursor will not receive its future models. The two had worked together for nearly four years.
Why it is important: A business built on another company's models can lose them for reasons that have nothing to do with how it behaves. Here a change of ownership was enough, with about eleven weeks' notice.
That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas
