Hacker News Reader

Comment by ampdepolymerase | Hacker News Reader

> To slim results. Elon Musk explained his new nihilism about the possibility of stopping AI advancement on the Joe Rogan podcast, when he said:

“I tried to convince people to slow down. Slow down AI. To regulate AI. This was futile. I tried for years. Nobody listened. Nobody listened.”

Rather rich coming from one of the self driving car market leaders. It certainly makes business sense to mislead academia and the policy sector into wasting resources on figuring out the best philosophical and ethical regime while large corporations benefit from regulatory capture. If he has his way, ML would become like the pharmaceutical industry, with multiple barriers of entry if you are not well-funded, well-connected, or established.

View on HN

I do believe it's reasonable to draw a distinction between what we're doing now (which is essentially just "statistics") and what Musk warns about.

It's a mistake to believe strong AI will just be a more powerful iteration of today's weak AI. He is arguing to slow developments toward general intelligence, not developments in any narrow field.

With high probability, 'just statistics' is an essential component of strong AI. Another essential component is embodiment, of which self-driving cars, and also military drones, are canonical examples. Researchers are taking the correct essential steps towards strong AI, it's a matter of (short) time until they succeed.

What are the requisite “correct essential steps” and how did you determine that they will necessarily lead to an AGI?

Elon wasn't calling for a hard halt like this article. He's talking about the dangers of general AI (AGI). I don't think most people would consider a self-driving car a likely path to accidental superintelligence; it's a highly-targeted application like a chess engine.

OpenAI's GPT-X engines OTOH, IMO, have a lot more potential danger because it's very unclear what they'll be used for.

If you do actual ML research, the technologies are two sides of the same coin. Three days ago there was a paper posted here in HN on using the GPT-like transformer architecture for reinforcement learning problems (of which self driving cars is a partial subset of).

https://news.ycombinator.com/item?id=27721037

* AI research already has high barriers to entry because the required computing resources are expensive.

* Self-driving AI will be heavily regulated for reasons not related to other AI regulation.

I don't see how more AI regulation would financially help Tesla or Elon Musk. In fact "more regulation of self-driving" is something that many Tesla competitors and critics support.

> Rather rich coming from one of the self driving car market leaders.

You don't need anything close to strong AI to do a reasonable job of driving a car. It seems like 90% of the problem is object recognition, and even in terms of brain-equivalent logic that's a really low bar.

It’s the 10% that gets you. To safely operate a car in all reasonable situations might very well require Strong AI. The car needs to be able to problem solve and make inferences about road conditions up ahead.

What kind of inferences need strong AI? How often does a driver need to figure out something with logic that couldn't be handled by current technology?

Level 4 self driving is fine and I really don't think it needs strong AI.

The remaining 10% that's 90% of the effort, is interrupting what is being done and changing context.

Without this capacity, "AI" is just a tiny shard of a complete mind.

I don't think anyone's really started grappling with this yet.

That's really not the 10% I was talking about. We don't need that part to follow some lanes.

Or the other way to put it is that all the other code is the first 90%, and then the "remaining 10% that's 90% of the effort" is the object recognition that was supposedly easy.

>We don't need that part to follow some lanes.

We do! I live near a highway and every time I use a particular on-ramp, the tar stripes make my car think it's going off the road. It never learns, and it has no plan B.

It's easy to do the lane-keeping, it's completely terra incognita to decide when to abandon the standard method that works 90% of the time. And what to do next.

For object recognition code, including lane-finding, to be sufficient, it needs to be looking far enough ahead that the car always has a plan B. If we reach that level of quality, and it still fails to find the lane, then the car can simply stop.

Deciding when to stop is a really easy threshold based on how confused the recognizer is and how far away the source of confusion is.

But also those sources of confusion will get rarer and rarer as the object recognition improves, to the point that even very mediocre handling of those situations will still give you an overall good experience.

Right now if a vehicle on auto pilot gets into an accident, the driver is scrutinized.

That is reasonable. The driver is expected to be alert incase intervention is needed.

If we take the manual override away, then we'll be squarely in the world that Mr. Musk is concerned about.

This is so obvious. Incumbents would like nothing better than a new law that makes it harder to operate (within “reason”, of course) in their industry. It is amazing that more people here don’t see right through the manipulation.

High entry barriers for self-driving are extremely good, for safety reasons.

Regulating research into general AI would not make those barriers considerably higher.

Tesla critics and competitors are also strongly in favor of self-driving regulations.

Obvious self-interest isn't proof that something is false.

For instance, it is obviously self-interested of me to not want the world to be blown up with nuclear weapons.

Reply on news.ycombinator.com