Hacker News Reader

> 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.

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.

> 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?

I don’t agree this is settled ethically unless you can come up with a better example. This leaves full agency in the hands of the driver throughout the application, unlike a drug which is a point-in-time decision which leads to a long chain of events outside the patient’s control. A better example would be something where a manual operator was displaced by an automated one, especially where the operator was a non-professional and the use of the thing is risky and already kills many people. Perhaps there is an analog in trucking or machinery.

In most cases, drugs need to be taken regularly for a period of time to have any observable effect, positive or negative. Sure, it's still much less immediate than a driving decision, but it's also not simply a point-in-time decision.

More importantly though, I believe that the statistical characteristics are more relevant here - I have no idea to how much risk I'm exposing myself by activating the Tesla "FSD", because (1) I don't know how much I can trust myself to switch from supervision to split-second control, and (2) I have no idea how badly this has been going for other drivers of the Tesla, or how likely the Tesla is to make really poor decisions (will it stop before the intersection, or plow directly into traffic? What percentage of the time will it do either?)

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

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.

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).

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

> How often do drivers intervene?

Extremely often, if the videos linked in the article are anything to go by.