Hacker News Reader

Comment by agnosticmantis | Hacker News Reader

It's amusing how Sutskever kept musking Musk over the years (overpromising with crazy deadlines and underdelivering):

In 2017 he wrote

"Within the next three years, robotics should be completely solved, AI should solve a long-standing unproven theorem, programming competitions should be won consistently by AIs, and there should be convincing chatbots (though no one should pass the Turing test)."

"We will completely solve the problem of adversarial examples by the end of August."

Very clever to take a page from Musk's own playbook of confidently promising self-driving by next year for a decade.

View on HN

That’s embarrassing and should be noted when he’s treated as a guru (as in today when I guess he gave a talk at Neurips conference) Of course, he should be listened to and treated as a true expert. But, it’s becoming more clear in viewing public people that extreme success can warp people’s perspective.

I mean, he wasn't that far off. The Turing test is well and truly beaten, regardless of how you define it, and I sure wouldn't want to go up against o1-pro in a programming or math contest.

Robotics being "solved" was indeed a stupid thing to assert because that's a hornet's nest of wicked problems in material science, mechanical engineering, and half a dozen other fields. Given a suitable robotic platform, though, 2020-era AI would have done a credible job driving its central nervous system, and it certainly wouldn't be a stumbling block now.

It's been a while since I heard any revealing anecdotes about adversarial examples in leading-edge GPT models, but I don't know if we can say it's a solved problem or not.

> The Turing test is well and truly beaten, regardless of how you define it

Unless the question the human asks is 'How many l's in llama'

Yeah, snark really settles the question, right up until the model gets better. Go try to fool o1-pro with that schtick.

This month, a computer solved the first Advent of Code challenge in eight seconds.

Everyone on Hacker News was saying "well of course, you can't just feed it to a chatbot, that's cheating! the leaderboard is a human competition!" because we've normalized that. It's not surprising, it's just obvious, oh yeah you can't have an Advent of Code competition if the computers get to play as well.

Granted it took seven years. Not three.

I think the achievements in the past couple of years are astonishing, bordering on magic.

Yet, confidently promising AGI/self-driving/mars landing in the next couple of years over and over when the confidence is not justified makes you a conman by definition.

If the number 3 means nothing and can become 7 or 17 or 170 why keep pulling these timelines out of their overconfident asses?

Did we completely solve robotics or prove a longstanding theorem in 2020? No. So we should lose confidence in their baseless predictions.

Self-driving is not so much a technological problem as it is a political problem. We have built a network of roads that (self-evidently) can't be safely navigated by humans, so it's not fair to demand better performance of machines. At least, not as long as they have to share the road with us.

'AI' landings on Mars are the only kind of landings possible, due to latency. JPL indisputably pwned that problem long before anyone ever heard of OpenAI.

Theorem-proving seems to require a different toolset, so I don't know what made him promise that. Same with robotics, which is more an engineering problem than a comp-sci one.

The cars are still worse than humans.

On an uneven playing field, yes. If we'd designed our roads for the robots, the robots would do better.

In any case the robots are getting better. Are we?

Reply on news.ycombinator.com