Two things landed this week. YC’s new Requests for Startups, and a podcast where an immunologist explains how GPT-5 cracked a problem his lab had been stuck on for three years.

Read together they say something neither says alone.

The bottleneck in AI is no longer intelligence. It is observability. Models already out-reason us anywhere the data exists. Where they fail, they fail because nobody is recording anything.

YC frames this batch as AI moving into the physical world. I think that is the wrong frame. The physical world is just where the sensors are missing.

I have been in this field since 2012. That shift is why I think now is the best moment to start something I have seen.

📌 AI Insider members get the full edition, including this week’s Maker Playbook. Start your free trial →

Multiplayer AI. Every tool that won the last twenty years won by going multiplayer. Docs beat Word. Figma beat Photoshop. AI is still single-player. You open a chat, you get an answer in a box only you can see. Agents now run tasks for days. That work was never meant to be done alone.

New operating systems for the physical world. 80% of the global workforce does not sit at a desk. Construction, maintenance, fleet ops. The software there has not changed in twenty years. Now you have three kinds of worker to coordinate: agents, robots, and humans wearing sensors. These industries spend 10 to 100x more on labour than on software.

A cloud for small software. Agents made it trivial to build a tool for one person or one team. Deploying and sharing it is still painful. AWS and Azure were designed for software that scales to millions, at the cost of complexity. Small software should be as easy to share as a Google Doc.

Proving you’re human. A finance worker joined a video call with his CFO and colleagues and wired out $25 million. Every other person on that call was a deepfake. Every trust signal we have was built for a world where faking a human was expensive. That world is gone.

Compute at sea. Data centres are running out of land and power. Permitting takes years and local government can kill it anyway. The ocean is 70% of the surface, has no permitting process, and is an enormous heat sink. The physics works. Maritime insurance is going to be the hard part, not the engineering.

Also on the list: an AI tutor good enough to teach a child to read, AI for the aging population, physical-world data collection, compliance infrastructure, crypto rails, and self-maintaining APIs.

For the first time the list includes a request from a sitting U.S. Secretary of the Army. Low-cost interceptors, drones, resilient logistics. That is a real budget attached to a door being held open.

Source: Y Combinator, Requests for Startups, Fall 2026

Rhonda Patrick released a long episode with Dr. Derya Unutmaz, immunologist and aging researcher. It is on my list this weekend.

The claim that matters is not a prediction. It already happened. GPT-5 helped Unutmaz’s lab solve an immunology problem they had been stuck on for three years. OpenAI has cited his work as evidence that frontier models can contribute to real discovery.

The rest follows from that. Months of biological analysis compressed into hours. Digital twins to personalise treatment and shorten trials. Disease predicted years before symptoms appear.

His strongest position: there is a point at which avoiding AI support becomes medically irresponsible. Not optional. Irresponsible.

You do not have to buy the timelines. Unutmaz thinks the next ten years could add decades to human lifespan. That is a big claim on a lot of unknowns. Take it as a direction, not a date.

The three-year mystery is not a forecast though. That is a result.

Source: FoundMyFitness with Dr. Derya Unutmaz

Here is the part that made me rewrite this issue.

Unutmaz’s lab did not need a new sensor. Three years of immunology data was already sitting on their drives. GPT-5 added inference, not observation.

YC’s own entry on physical-world data gives away the other half: once you can model a system, you can control it.

So there are two kinds of opportunity on that list, and confusing them will cost you two years.

Cognition-limited. The data already exists and exceeds human working memory. Immunology, code, contracts, compliance, claims. The models are ready now. Nothing is stopping you today.

Instrumentation-limited. The data does not exist yet. Construction sites, elder care, the battlefield, the atmosphere. You are a sensor company before you are an AI company. Triple whatever timeline you just estimated.

Most founders I meet pick an instrumentation-limited problem and budget like it is cognition-limited. That is the mistake.

The rare thing about right now is that both windows are open at the same time. That is not cheerleading. It is the reason the answer to when is now rather than soon.

Until next week,
Martin

Was this forwarded to you? Start your 7-day free trial →

I build AI agents and GenAI systems for companies. If you’re working on something like this — just reply to this email. I read every response.