Hacker News Reader

Comment by klabb3 | Hacker News Reader

> How exactly does this violate your privacy, assuming the data is analysed inside the car itself?

There is almost no remaining device-local data in consumer products, since everything that’s collected is usually exfiltrated or “shared with our partners to improve your experience”.

It’s not strange that people assume device-local info doesn’t exist anymore. It’s a reasonable simplification that almost always turns out to be accurate, sadly.

I worry we’re never gonna turn this around. Say gen Z, who have grown up after the compulsive data hoarding started. They won’t care because they don’t even remember another time..

View on HN

Given that this feature would have to be real-time and robust, and given what we know about how AI inference works, I think we can assume that the mechanism will be running locally on the car.

Otherwise you'd be able to drive drunk when out of cellphone range, drive drunk by messing with your car's antenna, drive drunk during system downtime, etc.

It seems plausible that car manufacturers will keep "kill-switch triggered" events local, if only because it's a huge liability, or because of the sentiment behind articles like these.

I think the more interesting question here is a discussion about the public policy tradeoffs of a local-only drunk driving kill-switch.

Cons: lack of user control over their own software and hardware, a slippery slope, cars will be self-driving on a roughly similar timeline anyway, and false-positive edge cases, like someone having a medical emergency getting blocked for driving weird on the way to the hospital.

Pros: 100s of lives saved in the first years, ~100,000s lives saved globally after full global saturation, a reduction in the need for drunk driving enforcement (which itself is extremely totalitarian, with traffic stops, checkpoints, court fees, jail time, suspended licenses, manslaughter convictions, lost jobs, etc.)

I lean against on software freedom principles alone, but it's an interesting decision space, because it's easy to limit specific harms and there is a lot to be gained, even in pure liberty terms. A world where someone who blacked out wakes up with a manslaughter or murder charge does not feel entirely free or fair either, since the conscious action they took was "drink until lose control of drinking." Many of us would rather get pulled over by their car than by a cop, I suspect.

If this means I stop getting woken up by idiots doing donuts in their motorcycles and Dodge Challengers and I’m less likely to be killed similar to the roughly half of all traffic deaths caused by drunk drivers, I’m not sure the cons outweigh the pros.

Driving is a licensed privilege anyway. You aren’t even allowed to drive at all unless you pass a written and physical test to prove your road-worthiness. There is no second amendment for driving. Your freedom to drive aggressively or drunk doesn’t outweigh everyone else’s physical safety.

It’s more unfortunate that the United States is built around driving where most people can’t get around and cover their basic needs without the ability to drive a personal vehicle.

I think the article is interpreting the law as a bunch of cops in the surveillance center monitoring us when what it’s asking automakers to do is make a basic feature that disables the car if it can’t detect an alert driver, just like how today’s driver assist systems disable themselves if they see you dozing off or not touching the wheel.

In my opinion, the vagueness of the language in the law is there for the benefit of automakers. As written, a carmaker that disables a vehicle after 20 minutes without detection of a hand on the steering wheel would be in compliance.

> I think we can assume that the mechanism will be running locally on the car.

It’s not an either/or situation. In 2023 it’s most likely to be cloud-first with local as a backup.

Exactly (although in this case I would assume that local is first because of flaky connectivity). But the point is, if any part of the system has determined levels of “impairment” using an AI system, that will most certainly be collected unless serious liability risks (such as with HIPAA or GDPR). However, the data must flow, so even with these the collection is still maximum within the law, and most certainly also outside too. So the deterrence argument is at best shaky, imo.

given what we know about how AI inference works, I think we can assume that the mechanism will be running locally

What do you know about AI inference that I don't? I thought most LLM models ran in the cloud because they are too power-hungry to run purely locally. If that's already true for text analysis and synthesis, doesn't that apply doubly so to 3D vision and motion analysis?

Training happens on the servers, existing trained models can be small enough to run locally and on embedded systems

This comment is a very weird one.

It already looks passed the idea that the kill switch and government determined ability to drive on a car level is already acceptable.

If the issue is people driving drunk, there's a lot you can do to mitigate it, such as making it unnecessary to involve operating a car in order to drink.

Reply on news.ycombinator.com