4 ms·

Without releasing data, Tesla is forcing analysts to focus on anecdotes, which suffer from all the usual problems like selection bias. What we need to know: -

by gfodor 6y ago

Without releasing data, Tesla is forcing analysts to focus on anecdotes, which suffer from all the usual problems like selection bias. What we need to know: - How often do drivers intervene? - How often do drivers fail to intervene and the system causes an accident? - How often does the system take an action which has a high likelihood of having prevented an accident? - How often does the system do so when it seems likely a human driver would have failed to do so? - Integrated together, what's the expected dynamics of these probabilities and their net impact insofar as the system being releases more widely creates training data to help improve them more quickly over time? It could very well turn out to be the case that this system is purely positive, strongly net positive, or neutral in harm reduction. The question then is, given that, what ethics should inform its release: is putting the stress on the driver sufficient if it say, saves 1000 lives in the next six months and will harm noone in exchange, other than some drivers having to intervene and have moments of stress? What if it will save 1000 and 10 people will be harmed by failing to intervene? What if it's an even swap, harming people who fail to intervene in exchange for saving people who would have inevitably been lost to accidents they couldn't have prevented? Saying "this is wrong to do" based upon anecdotes is dumb analysis. Saying "this is wrong to do" based upon an absolutist form of ethics is fair, but it also means you reject that we should be trying to solve self driving. If you think it's wrong to do and are not doing either, you ought to articulate what scenario in terms of data would justify the action vs not, even if that data is unknown right now.

kanox 6y ago
Tesla released data for autopilot, FSD beta has been out for ~1 week so there's not much real data.
- gfodor 6y ago
  Yeah, I'm not arguing that Tesla should have done something differently so much that I'm encouraging them to make sure they get data out ASAP.
tsimionescu 6y ago
> It could very well turn out to be the case that this system is purely positive, strongly net positive, or neutral in harm reduction It could also turn out to be negative, right? Certainly the anecdotes shown here would point much more to that - for example, its inability to turn sharply enough is indicative of some deep flaws. > If you think it's wrong to do and are not doing either, you ought to articulate what scenario in terms of data would justify the action vs not, even if that data is unknown right now. It's wrong to do as long as you rely on CUSTOMERS to ensure safety. There is no further argument to be had: as long as the system is not PROVEN to be safer than a human driver even in the absence of human attention, it can't be put into the hands of customers. That is not to say that they shouldn't be developing it, and paying and training people to operate it safely. That is what all other companies are in fact doing - taking a safe and responsible route, which is apparently too much to ask of the most highly valued car company in the world.
- gfodor 6y ago
  Certainly could be negative, I didn't list them not because I don't think they're possible, but because I was responding to the negative perspectives on this story which in many cases presumes the ones I listed are not possible. Saying "there is no further argument to be had" is just a way of shutting down debate. Why exactly is asking customers to opt-in to testing this somehow off the table? I'm not saying you're wrong, but you haven't given an argument and just are stating your belief as if it is some immutable fact.
  - tsimionescu 6y ago
    > Why exactly is asking customers to opt-in to testing this somehow off the table? Because they are likely to not just harm themselves, but also others. And also because they are not given access to enough information to be able to make an informed decision on whether the technology is safe to use, under what parameters. Not to mention that the lack of accountability this creates for the company introduces perverse incentives - the company is getting free labour, and the more it hides about real stats, the more free labour it is likely to get. This is all pretty well settled ethics - it is for the same reason that we don't allow companies to run beta tests for potentially life-saving drugs. This is long settled both legally and ethically, and the self-driving case is essentially identical, so why reopen that debate?
FartyMcFarter 6y ago
> How often do drivers intervene? Extremely often, if the videos linked in the article are anything to go by.
6gvONxR4sf7o 6y ago
It's not enough to know rates. If you know incidents per day, but don't know hours driven per day, you can't compare safety of two modes of driving. One mode might just be parked more often, and that's why it has fewer incidents per day. Similarly, if you know interventions per mile, but don't know miles of challenging circumstances per mile, you can't compare two modes. Maybe people only turn self driving on in particular circumstances. Or maybe tesla opened the beta to people in easy locations. Apples to apples safety comparisons require a lot of detailed data when they aren't part of a randomized trial (which is still hard to do correctly).
mlindner 6y ago
The system has been out and available for a tiny selection of drivers for 1 week. How can there be statistics?

by gfodor 6y ago

kanox 6y ago

Tesla released data for autopilot, FSD beta has been out for ~1 week so there's not much real data.

gfodor 6y ago
Yeah, I'm not arguing that Tesla should have done something differently so much that I'm encouraging them to make sure they get data out ASAP.

tsimionescu 6y ago

> It could very well turn out to be the case that this system is purely positive, strongly net positive, or neutral in harm reduction It could also turn out to be negative, right? Certainly the anecdotes shown here would point much more to that - for example, its inability to turn sharply enough is indicative of some deep flaws. > If you think it's wrong to do and are not doing either, you ought to articulate what scenario in terms of data would justify the action vs not, even if that data is unknown right now. It's wrong to do as long as you rely on CUSTOMERS to ensure safety. There is no further argument to be had: as long as the system is not PROVEN to be safer than a human driver even in the absence of human attention, it can't be put into the hands of customers. That is not to say that they shouldn't be developing it, and paying and training people to operate it safely. That is what all other companies are in fact doing - taking a safe and responsible route, which is apparently too much to ask of the most highly valued car company in the world.

gfodor 6y ago
Certainly could be negative, I didn't list them not because I don't think they're possible, but because I was responding to the negative perspectives on this story which in many cases presumes the ones I listed are not possible. Saying "there is no further argument to be had" is just a way of shutting down debate. Why exactly is asking customers to opt-in to testing this somehow off the table? I'm not saying you're wrong, but you haven't given an argument and just are stating your belief as if it is some immutable fact.
- tsimionescu 6y ago
  > Why exactly is asking customers to opt-in to testing this somehow off the table? Because they are likely to not just harm themselves, but also others. And also because they are not given access to enough information to be able to make an informed decision on whether the technology is safe to use, under what parameters. Not to mention that the lack of accountability this creates for the company introduces perverse incentives - the company is getting free labour, and the more it hides about real stats, the more free labour it is likely to get. This is all pretty well settled ethics - it is for the same reason that we don't allow companies to run beta tests for potentially life-saving drugs. This is long settled both legally and ethically, and the self-driving case is essentially identical, so why reopen that debate?

FartyMcFarter 6y ago

> How often do drivers intervene? Extremely often, if the videos linked in the article are anything to go by.

6gvONxR4sf7o 6y ago

It's not enough to know rates. If you know incidents per day, but don't know hours driven per day, you can't compare safety of two modes of driving. One mode might just be parked more often, and that's why it has fewer incidents per day. Similarly, if you know interventions per mile, but don't know miles of challenging circumstances per mile, you can't compare two modes. Maybe people only turn self driving on in particular circumstances. Or maybe tesla opened the beta to people in easy locations. Apples to apples safety comparisons require a lot of detailed data when they aren't part of a randomized trial (which is still hard to do correctly).

mlindner 6y ago

The system has been out and available for a tiny selection of drivers for 1 week. How can there be statistics?