3 ms·

What's an acceptable error rate for such a model? What's the human baseline? I ask these questions[1][2] for nearly every model I'm requested to develop. It us

What's an acceptable error rate for such a model? What's the human baseline? I ask these questions[1][2] for nearly every model I'm requested to develop. It usually throws off PMs who aren't used to any acceptable errors. It's always fun to get a request that's several orders of magnitude better than human performance. Sometimes it's actually possible to deliver, but often it shifts the conversation into feasibility and expectations. Anyone developing a model like this will have errors and definitionally makes systemic. Regardless of performance it's important to address that. 1. In this order 2. Sometimes I study human baselines but ask before revealing because it prevents anchoring.

tekacs 3y ago
Absolutely -- an academic environment I was in used to have a lot of government callers come by, asking for predictive power to solve any number of humanitarian crises, terrorism, crime and so on. We collectively used to ask them every time, 1. What data they had to work with. 2. What was the human baseline that they were currently executing with. 3. What performance would make a difference to this process. ... and as you say, the answers we would get would tend to sound something like. 1. Almost nothing. 2. Unclear, un-measured. 3. Something well beyond incredible, well-intentioned humans with incredible data. In that case the lesson was often to simply not try, but it does sometimes feel like other areas of government like E2E-breaking legislation, in that well-intentioned folks will try and try and try again and eventually get someone to fund and execute on the project, feasibility notwithstanding. In this instance with United... the data that we're getting via Ars and this case is so unclear and one-sided that it's hard to know how well things actually work... it'd be fun to consign them (for now) to "don't even try" if we had the info to do so, but it feels like it's a matter of time either way and so... I think the ideal with most of these cases is to eventually ship something that's actually high quality.
raverbashing 3y ago
> What's an acceptable error rate for such a model? What's the human baseline? Funny, I though it was socialized medicine that had "death panels" and such /s
- kelnos 3y ago
  If anything, the "death panels" on private insurers (whatever form they might actually take) are likely much worse than a government-run "death panel".
  - willcipriano 3y ago
    I haven't heard the death panel canard brought out since Canada started it's death panels. It's like the classic idea, but instead of letting you die from lack of medical care they drive you into depression prior to killing you. > Are Canadians being driven to assisted suicide by poverty or healthcare crisis? https://www.theguardian.com/world/2022/may/11/canada-cases-right-to-die-laws https://www.theguardian.com/world/2022/may/11/canada-cases-r... > Why is Canada euthanising the poor? https://www.spectator.co.uk/article/why-is-canada-euthanising-the-poor-/ https://www.spectator.co.uk/article/why-is-canada-euthanisin...
kelnos 3y ago
> What's an acceptable error rate for such a model? Ideally, 0%. We're talking about people's lives and health here. But, right, realistically: > What's the human baseline? Good question, but I would hope it's much better than a 90% error rate. And if it's not, any sane company in any sane industry would figure out why things are that bad, and fix it. Of course, insurance companies are the scum of the earth, and have a huge incentive to get things wrong in this way.
- kelseyfrog 3y ago
  Why is there a different standard between the model and a human? From the perspective of the business, whether a model or human causes the error is immaterial. > Any sane company in any sane industry would figure out why things are that bad, and fix it. The function of a thing is its consequences. So we have to conclude that the reason the model had a 90% error rate was to have a 90% error rate. It's aligned with business incentives to be so.

tekacs 3y ago

Absolutely -- an academic environment I was in used to have a lot of government callers come by, asking for predictive power to solve any number of humanitarian crises, terrorism, crime and so on. We collectively used to ask them every time, 1. What data they had to work with. 2. What was the human baseline that they were currently executing with. 3. What performance would make a difference to this process. ... and as you say, the answers we would get would tend to sound something like. 1. Almost nothing. 2. Unclear, un-measured. 3. Something well beyond incredible, well-intentioned humans with incredible data. In that case the lesson was often to simply not try, but it does sometimes feel like other areas of government like E2E-breaking legislation, in that well-intentioned folks will try and try and try again and eventually get someone to fund and execute on the project, feasibility notwithstanding. In this instance with United... the data that we're getting via Ars and this case is so unclear and one-sided that it's hard to know how well things actually work... it'd be fun to consign them (for now) to "don't even try" if we had the info to do so, but it feels like it's a matter of time either way and so... I think the ideal with most of these cases is to eventually ship something that's actually high quality.

raverbashing 3y ago

> What's an acceptable error rate for such a model? What's the human baseline? Funny, I though it was socialized medicine that had "death panels" and such /s

kelnos 3y ago
If anything, the "death panels" on private insurers (whatever form they might actually take) are likely much worse than a government-run "death panel".
- willcipriano 3y ago
  I haven't heard the death panel canard brought out since Canada started it's death panels. It's like the classic idea, but instead of letting you die from lack of medical care they drive you into depression prior to killing you. > Are Canadians being driven to assisted suicide by poverty or healthcare crisis? https://www.theguardian.com/world/2022/may/11/canada-cases-right-to-die-laws https://www.theguardian.com/world/2022/may/11/canada-cases-r... > Why is Canada euthanising the poor? https://www.spectator.co.uk/article/why-is-canada-euthanising-the-poor-/ https://www.spectator.co.uk/article/why-is-canada-euthanisin...

kelnos 3y ago

> What's an acceptable error rate for such a model? Ideally, 0%. We're talking about people's lives and health here. But, right, realistically: > What's the human baseline? Good question, but I would hope it's much better than a 90% error rate. And if it's not, any sane company in any sane industry would figure out why things are that bad, and fix it. Of course, insurance companies are the scum of the earth, and have a huge incentive to get things wrong in this way.

kelseyfrog 3y ago
Why is there a different standard between the model and a human? From the perspective of the business, whether a model or human causes the error is immaterial. > Any sane company in any sane industry would figure out why things are that bad, and fix it. The function of a thing is its consequences. So we have to conclude that the reason the model had a 90% error rate was to have a 90% error rate. It's aligned with business incentives to be so.