AI has been one of the biggest let down in this pandemic.
There was so much noise for over a decade and when it came to complex systems even Google's of the world didn't do shit (location tracking app is NOT AI).
Bill and Melinda Gates's philanthropic venture has been phenomenal while Microsoft's AI team has been and still taking a nap (chatbots don't count).
Finally, someone spoke what it needs to be said (and I do modeling n AI for a living).
Thank you, Cheryl Rofer - Read her entire twitter thread and I am posting it here too:
1. There are a hundred gazillion models out there. Few of their owners have bothered to compare their model to others to see what is working and what isn't.
2. By the standards of the models I've worked with, they are all simple - a few differential equations, a curve fit. I've worked with a hundred or more elementary reactions and then a mass- and heat-transfer model that incorporated those in. (Hint: we had to boil them down to six)
3. The only model I have seen that is at all transparent about its parameters is the Imperial College model. All the others I have seen are curve fits. They mumble about social distancing as a variable but never say which parameter it fits into.
4. All the curve fits are with different functions. At least, back in the day when chemical kinetics was curve fitting, we always used the same function.
5. It looks like amateur hour. Everyone's got their pet model, but they're not telling us what it is. And a further layer of amateurs on Twitter say solemn words like "assumptions" that they have nothing to back up.
6. I've spent a lot of time on the Imperial College and IHME models. I don't intend to do that for every single model out there. (The Imperial College Modeling Of The Coronavirus)
7. My assumption from here on in is that any model but the Imperial College model is crap until it's explained. With explicit connection of assumptions to parameters.
8. Tyler Cowen (and probably other economists who think they've got something to say) has never looked closely at the epidemiological models, judging from this. His questions are obnoxious. Ignore stuff like this - What does this economist think of epidemiologists?
Yes, this is a huge issue now - people cannot shut up. Most of them became overnight economists, virologists, epidemiologists, doctors, modelers and experts in everything under the sun (so much for epistemological modesty).
As J.D Salinger's Holden would say - I could puke next time someone gives an intellectual fart talk on "the great AI threat".
But to be fair:
It takes lots of data, domain knowledge, a range of diverse knowledge (foxes and not hedgehogs) and more importantly patience and time to train a generalized model. One cannot instantly "deep learn" the way out of a pandemic.
There are some great exceptions - Five-Thirty-Eight had a good post on Why It’s So Freaking Hard To Make A Good COVID-19 Model and of-course Nassim Taleb.
There are some hidden treasures who have been modeling this for a long time - kudos to them:
There was so much noise for over a decade and when it came to complex systems even Google's of the world didn't do shit (location tracking app is NOT AI).
Bill and Melinda Gates's philanthropic venture has been phenomenal while Microsoft's AI team has been and still taking a nap (chatbots don't count).
Finally, someone spoke what it needs to be said (and I do modeling n AI for a living).
Thank you, Cheryl Rofer - Read her entire twitter thread and I am posting it here too:
1. There are a hundred gazillion models out there. Few of their owners have bothered to compare their model to others to see what is working and what isn't.
2. By the standards of the models I've worked with, they are all simple - a few differential equations, a curve fit. I've worked with a hundred or more elementary reactions and then a mass- and heat-transfer model that incorporated those in. (Hint: we had to boil them down to six)
3. The only model I have seen that is at all transparent about its parameters is the Imperial College model. All the others I have seen are curve fits. They mumble about social distancing as a variable but never say which parameter it fits into.
4. All the curve fits are with different functions. At least, back in the day when chemical kinetics was curve fitting, we always used the same function.
5. It looks like amateur hour. Everyone's got their pet model, but they're not telling us what it is. And a further layer of amateurs on Twitter say solemn words like "assumptions" that they have nothing to back up.
6. I've spent a lot of time on the Imperial College and IHME models. I don't intend to do that for every single model out there. (The Imperial College Modeling Of The Coronavirus)
7. My assumption from here on in is that any model but the Imperial College model is crap until it's explained. With explicit connection of assumptions to parameters.
8. Tyler Cowen (and probably other economists who think they've got something to say) has never looked closely at the epidemiological models, judging from this. His questions are obnoxious. Ignore stuff like this - What does this economist think of epidemiologists?
Yes, this is a huge issue now - people cannot shut up. Most of them became overnight economists, virologists, epidemiologists, doctors, modelers and experts in everything under the sun (so much for epistemological modesty).
The best minds of my generation are thinking about how to make people click ads. That sucks.This is what we get from AI when we forget the sense of the reality of complex systems vs. earning million-dollar salaries by making people click on links.
- Jeff Hammerbacher
As J.D Salinger's Holden would say - I could puke next time someone gives an intellectual fart talk on "the great AI threat".
But to be fair:
It takes lots of data, domain knowledge, a range of diverse knowledge (foxes and not hedgehogs) and more importantly patience and time to train a generalized model. One cannot instantly "deep learn" the way out of a pandemic.
There are some great exceptions - Five-Thirty-Eight had a good post on Why It’s So Freaking Hard To Make A Good COVID-19 Model and of-course Nassim Taleb.
There are some hidden treasures who have been modeling this for a long time - kudos to them:
No comments:
Post a Comment