Goodhart's law causes problems proportional to the availability and comprehensiveness of accurate data. If you have an abundance of available metrics that all contribute to the desired outcome, targeting an aggregate of those metrics is far less vulnerable to gaming than if useful metrics are few and difficult to collect.
Thus, a huge obstacle to understanding biology is how efficient it is to collect reliable data. Need to measure the contents of someone's bloodstream? Blood tests require a needle, labs, and reagents. Need to check on someone's brain in detail? You have to put them in a multi-million-dollar room-sized MRI machine.
Don't get me wrong, MRI machines are marvels of engineering and labs are doing critical work! Nevertheless, it is immensely frustrating how inefficient even our very best data gathering methods are for biological processes.
There are very few useful health metrics that *can* be collected in frictionless and automatic ways, things like heart rate and blood oxygen that can be tracked via non-invasive wearable devices.
If I want to understand a computer, I can search through the code and, with study, reliably identify each instance of cause and effect. I can look at each piece of data in memory and discover where and how it is being used. I can study the system architecture and hardware specifications.
If this scenario were medical science, it'd be like if we could no longer view the monitor and were forced to make inferences about the computer's internal state by measuring its power usage, temperature readings, and the weak electromagnetic fields surrounding the internal components.
In this analogy, our ability to talk to the human about their physical/mental state would be like if the monitor can be viewed...but only by a small child who cannot read and can only give vague descriptions of the colors and shapes on the screen. (The child also has an innate sense of whether particular pieces of hardware are 'hurting' or not.)
There is no documentation nor specifications except those created by those studying under these same limitations.
I applaud doctors, researchers, biologists, and everyone else working under these nightmarish limitations.
To this computer scientist, it seems like legibility hell.
No doubt, medical science is plagued by dimensionality reduction. We reduce the 3D dynamic organism to its 2D static shadows and try to draw inferences from those silhouettes.
Loved this and you are so correct. Thats one of the reasons i wrote about the causes of the “loss of Intellectual curiosity” in younger generations. The formal ‘education’ system has lost its education mission, to becoming a train to the test environment. Similarly I see (typically younger) people in gyms abandon general fitness and strength/functional conditioning in the extensive pursuit of micro improvements. Again ‘training to the test’ except that this time the test is typically vanity. Lovely to find your writing.
Thanks, Stuart! I found your newsletter from an infographic plotting various writers across the axes of educational/entertaining and short-form/long-form. Enjoyed your article on how shifting incentive structures kill curiosity in favor of conformity.
This loss of general intellectual curiosity seems to parallel the increasing trend towards specialization that began with the Industrial Revolution. As Freakonomics explains, if you understand people's incentives, you can predict their behavior.
In other words, never confuse association with causation.
Many measurements (e.g., grip strength) can be associated with a desired outcome (e.g., long term health) without causing the outcome. This problem shows up over and over again in all sorts of studies of human health and elsewhere. And it's often challenging to solve.
Goodhart's law causes problems proportional to the availability and comprehensiveness of accurate data. If you have an abundance of available metrics that all contribute to the desired outcome, targeting an aggregate of those metrics is far less vulnerable to gaming than if useful metrics are few and difficult to collect.
Thus, a huge obstacle to understanding biology is how efficient it is to collect reliable data. Need to measure the contents of someone's bloodstream? Blood tests require a needle, labs, and reagents. Need to check on someone's brain in detail? You have to put them in a multi-million-dollar room-sized MRI machine.
Don't get me wrong, MRI machines are marvels of engineering and labs are doing critical work! Nevertheless, it is immensely frustrating how inefficient even our very best data gathering methods are for biological processes.
There are very few useful health metrics that *can* be collected in frictionless and automatic ways, things like heart rate and blood oxygen that can be tracked via non-invasive wearable devices.
If I want to understand a computer, I can search through the code and, with study, reliably identify each instance of cause and effect. I can look at each piece of data in memory and discover where and how it is being used. I can study the system architecture and hardware specifications.
If this scenario were medical science, it'd be like if we could no longer view the monitor and were forced to make inferences about the computer's internal state by measuring its power usage, temperature readings, and the weak electromagnetic fields surrounding the internal components.
In this analogy, our ability to talk to the human about their physical/mental state would be like if the monitor can be viewed...but only by a small child who cannot read and can only give vague descriptions of the colors and shapes on the screen. (The child also has an innate sense of whether particular pieces of hardware are 'hurting' or not.)
There is no documentation nor specifications except those created by those studying under these same limitations.
I applaud doctors, researchers, biologists, and everyone else working under these nightmarish limitations.
To this computer scientist, it seems like legibility hell.
No doubt, medical science is plagued by dimensionality reduction. We reduce the 3D dynamic organism to its 2D static shadows and try to draw inferences from those silhouettes.
Loved this and you are so correct. Thats one of the reasons i wrote about the causes of the “loss of Intellectual curiosity” in younger generations. The formal ‘education’ system has lost its education mission, to becoming a train to the test environment. Similarly I see (typically younger) people in gyms abandon general fitness and strength/functional conditioning in the extensive pursuit of micro improvements. Again ‘training to the test’ except that this time the test is typically vanity. Lovely to find your writing.
Thanks, Stuart! I found your newsletter from an infographic plotting various writers across the axes of educational/entertaining and short-form/long-form. Enjoyed your article on how shifting incentive structures kill curiosity in favor of conformity.
This loss of general intellectual curiosity seems to parallel the increasing trend towards specialization that began with the Industrial Revolution. As Freakonomics explains, if you understand people's incentives, you can predict their behavior.
Ah! Was that Andrew Tsang’s article with the infographic?
Yes, that was the one!
This is a wonderfully succinct and important reminder. On you go, Nita!
Thanks, Thomas! I know you could teach us a thing or two as an endurance athlete.
In other words, never confuse association with causation.
Many measurements (e.g., grip strength) can be associated with a desired outcome (e.g., long term health) without causing the outcome. This problem shows up over and over again in all sorts of studies of human health and elsewhere. And it's often challenging to solve.