Excellent review. I am likely to buy and read the book.
In the extreme and weird scenarios the basic pitch is that when we separate the mechanism from the objective, bad things can happen, like hypno-drones telling people to buy paperclips. It feels like we should employ the same basic trick when evaluating the current things people are worried about, like deepfakes.
Deepfaked videos aren’t a meaningful threat because video just isn’t that important. But what if we could deepfake medicine? According to a WHO article from 2010, counterfeit medicine was worth ~$75B/yr then. That seems plenty big enough to merit throwing similar ML techniques at designing a pill with no active ingredient but that passed various kinds of basic tests for whether a medication is genuine.
The problem with deepfakes isn’t that there are fake videos, it is that we are on track for a general and reliable method of fakery.
That seems plenty big enough to merit throwing similar ML techniques at designing a pill with no active ingredient but that passed various kinds of basic tests for whether a medication is genuine.
Ahm… the way we test pills that are FDA approved is by feeding them to humans and seeing if they have the desired effect upon disease-related markers based on assays that imperfectly capture those markers.
So, this is already happening I’m afraid, no drugs are designed to cure anything in particular, they are designed to optimize for the marker we can test which lead us to think they will pass the tests that say they are a cure for a diseases.
Drugs that are arguably not very useful or even harmful (e.g. statins) have been designed this way already.
Excellent review. I am likely to buy and read the book.
In the extreme and weird scenarios the basic pitch is that when we separate the mechanism from the objective, bad things can happen, like hypno-drones telling people to buy paperclips. It feels like we should employ the same basic trick when evaluating the current things people are worried about, like deepfakes.
Deepfaked videos aren’t a meaningful threat because video just isn’t that important. But what if we could deepfake medicine? According to a WHO article from 2010, counterfeit medicine was worth ~$75B/yr then. That seems plenty big enough to merit throwing similar ML techniques at designing a pill with no active ingredient but that passed various kinds of basic tests for whether a medication is genuine.
The problem with deepfakes isn’t that there are fake videos, it is that we are on track for a general and reliable method of fakery.
Ahm… the way we test pills that are FDA approved is by feeding them to humans and seeing if they have the desired effect upon disease-related markers based on assays that imperfectly capture those markers.
So, this is already happening I’m afraid, no drugs are designed to cure anything in particular, they are designed to optimize for the marker we can test which lead us to think they will pass the tests that say they are a cure for a diseases.
Drugs that are arguably not very useful or even harmful (e.g. statins) have been designed this way already.