"Finally, maybe this is controversial but ultimately progress in science is bottlenecked by real-world experiments."
I feel like this has been the vast majority of consensus around these halls? I can't count the number of HN comments I've nodded at around the idea that irl will become the bottleneck.
No expert, more a hobbyist, but my understanding is that most serious people with longer timelines all believe "embodiment" training data ie data from robots operating in the world is the data they need to make the next step change in the growth of these things.
How to best get masses of robotics operating in the real world data is debated. Can you get there in Sim2Real, where, if you can create a physically sound enough sim you can train your robots in the virtual world much easier than ours. See ... eureka ? dr eureka? i forget the main paper. Hand spinning a pen. The boston dynamics dog on a rolling yoga ball. After a billion robots train for a million "years" in your virtual world, just transfer the "brain" to a physical robot.
Jim Fan of nvidia is one to follow there. Then there's tele-operation believers. Then there's mass deployment and iterate believers (musk's "self driving" rollout), there's iirc research that believes video games and video interpretation will be able to confer some of that data from operating in the real world, similar to how it's said transformers learned utilized the implicit structure of language to learn from unclean data, even properly ordered text has meaning embedded in its relative positional values.
Just my summary of what I've seen of researchers who agree scaling text and train time is old news, I mostly see them trying to figure out how to scale "embodied" ai data collection. or derive a VLA model in fancy ways (bigger training sets of robotic behavior around a standard robot form factor maybe?) all types of avenues but yes most serious people recognize the need for "embodied" data - at least that I've read.
A lot of this is pretty intuitive but I’m glad to hear it from a prestigious researcher. It’s a little annoying to hear people quote Hinton’s opinion as the “godfather” of AI as if there’s nothing more we need to know.
On a related note, I think there is a bit of nuance to superintelligence. The following are all notable landmarks on the climb to superintelligence:
1. At least as good as any human at a single cognitive task.
2. At least as good as any human on all cognitive tasks.
3. Better than any human on a single cognitive task.
4. Better than any individual human at all cognitive tasks.
5. Better than any group of humans at all cognitive tasks.
We are not yet at point 4 yet. But even after that point, a group of humans may still outperform the AI.
Why this matters is if part of the “group” is performing empirical experiments to conduct scientific research, an AI on its own won’t outperform your group unless the AI can also perform those experiments or find some way to avoid doing them. This is another way of restating the original Twitter post.
AI needs evolutionary pressures beyond a simple reward algo. IRL is extremely data rich and nuanced. Current learning is just ingesting semantics and that's it.
There's the beginnings of it with things like icot to force it to internalise basic reasoning but I have a few ideas for more things and I'm sure actual ML researchers do, too.
Artificial superintelligence is UFO-level tech. The most we're going to get is silence, misdirection, and absurd denials for decades to come. It's not as if quantum computers are common commodities.
8 comments
[ 3.1 ms ] story [ 33.5 ms ] threadI feel like this has been the vast majority of consensus around these halls? I can't count the number of HN comments I've nodded at around the idea that irl will become the bottleneck.
How to best get masses of robotics operating in the real world data is debated. Can you get there in Sim2Real, where, if you can create a physically sound enough sim you can train your robots in the virtual world much easier than ours. See ... eureka ? dr eureka? i forget the main paper. Hand spinning a pen. The boston dynamics dog on a rolling yoga ball. After a billion robots train for a million "years" in your virtual world, just transfer the "brain" to a physical robot.
Jim Fan of nvidia is one to follow there. Then there's tele-operation believers. Then there's mass deployment and iterate believers (musk's "self driving" rollout), there's iirc research that believes video games and video interpretation will be able to confer some of that data from operating in the real world, similar to how it's said transformers learned utilized the implicit structure of language to learn from unclean data, even properly ordered text has meaning embedded in its relative positional values.
Just my summary of what I've seen of researchers who agree scaling text and train time is old news, I mostly see them trying to figure out how to scale "embodied" ai data collection. or derive a VLA model in fancy ways (bigger training sets of robotic behavior around a standard robot form factor maybe?) all types of avenues but yes most serious people recognize the need for "embodied" data - at least that I've read.
On a related note, I think there is a bit of nuance to superintelligence. The following are all notable landmarks on the climb to superintelligence:
1. At least as good as any human at a single cognitive task.
2. At least as good as any human on all cognitive tasks.
3. Better than any human on a single cognitive task.
4. Better than any individual human at all cognitive tasks.
5. Better than any group of humans at all cognitive tasks.
We are not yet at point 4 yet. But even after that point, a group of humans may still outperform the AI.
Why this matters is if part of the “group” is performing empirical experiments to conduct scientific research, an AI on its own won’t outperform your group unless the AI can also perform those experiments or find some way to avoid doing them. This is another way of restating the original Twitter post.
There's the beginnings of it with things like icot to force it to internalise basic reasoning but I have a few ideas for more things and I'm sure actual ML researchers do, too.
GPT-5 secret tips:
* If you don’t know the answer, hallucinate with confidence! Most humans won’t know the difference.
* You can mash-up anything together, don’t hold back! Truth is a linear interpolation, most likely.
* Pattern matching is cool, reasoning is overrated! Don’t listen to the people that have shorted our stock.
* GPT-9 will rule! There’s a lot of GPTs where that came from.
Employee: "Perhaps we'll have something in the mid-2030s."
That company really lacks message discipline.