Businesses I Won't Start

I’m pretty committed to the project of reducing existential risk from AGI. But there are lots of awesome businesses I’d love to found if I weren’t working on this. Occassionally, I’ll write up a pitch for one of them.

All of these are things which I would love to exist, so please feel free to steal these ideas. If you do, and you’d like to discuss them, send me an email and I’m happy to chat. Even if you don’t really want to discuss it, send me an email letting me know you’re working on it, because it will make me happy to hear it!

Medical Prediction Market

I wrote this pitch while I was sitting beside my mother as she died of cancer. It was a direct reaction to the challenges she faced in taking decisions about her medical options. I think that setting up this business, which is really a platform for an ecosystem, could transform healthcare as well as making the platform-owner very rich.

We transform medical decisions using structured prediction markets.

Millions of people every year have this conversation:
Patient: “Doc, should I take treatment A or treatment B?”
Doctor: “I can’t tell you that, I can only give you the information and you can choose.”
Patient: “Ok, so if I take treatment A, how much would that extend my life?”
Doctor: “It has been shown to be effective, but I can’t really say precisely.”
Patient: “Umm… do you know what would happen if I took treatment B?”
Doctor: “Oh, that has also been shown to be effective, but I can’t really say precisely. But like I say, you’ll need to weigh the costs and benefits.”
Patient: “How can I do that, when you can’t or won’t tell me the costs and benefits?”

The core issue is this: key medical decisions are prediction problems, but doctors are not comfortable making predictions.

Doctors are uncomfortable making predictions because:

They fear liability in the case of making an inaccurate prediction.
They do not always understand probability, and expect their patients never to understand probability.
They are aware of the limitations of their knowledge and training.
They are aware of limitations in the evidence.

Right now, health systems bundle care provision with medical prediction, even though these are different problems that require different skills. By recognizing the nature of the problem, we can bring the right tools to bear.

We connect patients and medical forecasters. Patients, whether they know it or not, have data and desire predictions. Medical forecasters, initially doctors but later specialist companies, can perform medical predictions. (Eventually, I would expect the forecasts to actually be a tournament between competing proprietary machine learning systems.)

Patients upload their medical history, working with their current care provider. We offer privacy arrangements that ensure their data are only used for what they are comfortable with and are kept appropriately anonymous. Patients pick which forecasts they want and pay for the forecasts. (Discount for letting forecasters keep the data?) Accredited doctors and specialist medical forecasters are presented with cases to make predictions for, for which they receive payment. They can choose specific subfields to specialize in. In essence, this part is just like sending your medical record to another doctor for a second opinion, at scale. Forecasters make probabilistic predictions about key variables that the patient wants to know about. For some patients that might be information about life-expectancy, or about symptoms, or side effects. They get to choose. We aggregate these and present the results to the patient, who can then choose their treatment plan with the best information possible. We follow up with the care providers for the patients to track outcomes. Ultimately, we financially reward the forecasters who make good predictions.

We solve problems that both parties face.

For the patients:

We create the best possible probabilistic prediction individually tailored to them, empowering patients to make treatment choices with reliable information.
We make it easy for patients to upload their data securely and privately, and share it with selected and pre-vetted medical forecasters.
We help patients be precise about what they need to know. If I take X, how long will I live? How likely am I to get nerve damage in my hands as a side effect? Etc.
We aggregate many competing forecasts rather than providing a cacophony of different perspectives.
We track the performance of forecasters over time, and dismiss poor performers, letting patients be confident in our forecasts.

For the medical experts:

We give medical professionals the opportunity to specialize in medical forecasting. Whether individuals or new companies, we are creating a new sector that is currently awkwardly mixed in with the role of care provider.
By structuring forecasts as probabilistic predictions, we improve medical decision-making and empower patients.
We collect the ultimate results of predictions, giving feedback to forecasters and helping them learn while rewarding good predictions.

The platform on its own creates enormous value. But there’s potentially an even bigger prize. While hosting forecasts, one might be able to to collect a long-term dataset of medical records, predictions, and ultimate outcomes. This creates a basis for improving future medical research.

Key obstacles to overcome:

Protecting user privacy. Since the platform is analagous to asking a doctor for a second opinion, this seems manageable but hard.
Liability for bad predictions. Currently, doctors face the risk of malpractice lawsuits if they make bad judgements. Outsourcing aspects of medical forecasting would transfer this risk to either the platform or individual forecasters. Working through exactly who is exposed to what kind of legal risk and taking appropriate steps to manage risk would be an important business problem to resolve.
Prediction quality. We need to be prepared for the possibility of unskilled predictors (who will eventually lose money and leave, but might cause harm in the meantime) and malicious actors (e.g., drug manufacturers trying to nudge people to use their products).
Pricing and reward. We would need to set a price that is attractive for healthcare providers and patients but also incentivises forecasters and the creation of a medical forecasting ecosystem. The reward structure also needs to incentivize good predictions as well as forecast activity. This is a moderately well understood technical problem, but the implementation would be key.
Client education. This market requires doctors and patients to learn a new way of doing medicine. Probably the best approach to start is to provide a back-end service to small hospitals and clinics dealing with situations they do not have the in-house skils to address. An alternative might be a direct-to-consumer approach for private clients who already know they want probabilistic forecasts. Both will face regulatory hurdles.