Roon says optimizations could bypass data and compute constraints
@MatthewJBar I think you’re wrong and there’s 1,000x efficiency gains leftover in deep learning research that could lead to much smarter faster more agentic models given the same inputs
@tszzl I don't think it's a good development. I continue to think that RSI is an overrated risk vector due to data and compute bottlenecks, and that slowing down AI would accomplish little at enormous cost.
Positive users agree the 1000x efficiency gains claim for deep learning is accurate and already happening, while negative users call it vague speculation that ignores physical limits and won't boost performance.
what ML researchers do is improve compute efficiency and data efficiency. if you believe we’ve already covered most of the ground, then you’ve got little to fear. if you think we aren’t close to the “Landauer limit” (using this entirely metaphorically) of model training, then RSI can change things dramatically
@tszzl @MatthewJBar that's such a vague claim
@MatthewJBar I think you’re wrong and there’s 1,000x efficiency gains leftover in deep learning research that could lead to much smarter faster more agentic models given the same inputs
@tszzl @MatthewJBar that's such a vague claim
@MatthewJBar I think you’re wrong and there’s 1,000x efficiency gains leftover in deep learning research that could lead to much smarter faster more agentic models given the same inputs
@tszzl @MatthewJBar i agree RSI *can* change things dramatically. that's again an extremely weak claim that:
1. presupposes "RSI" exists 2. qualifies the "change things dramatically" prediction with "can" (which makes it impossible to disagree with, especially conditional on 1)
what ML researchers do is improve compute efficiency and data efficiency. if you believe we’ve already covered most of the ground, then you’ve got little to fear. if you think we aren’t close to the “Landauer limit” (using this entirely metaphorically) of model training, then RSI can change things dramatically
@tszzl @MatthewJBar matthew is just saying RSI is overrated as a *risk vector*, as in, he thinks it's not going to happen because there are other inputs that go into improving AI systems that will become bottlenecks with abundant researcher effort
your claim doesn't respond to that at all
@tszzl @MatthewJBar in a world where r&d progress in AI is itself compute-bottlenecked, it can simultaneously be true that:
1. we're far from the "landauer limit" of efficiently using compute and data during training, inference, etc
2. we can't close the gap to it by scaling cognitive effort alone
@tszzl @MatthewJBar matthew is just saying RSI is overrated as a *risk vector*, as in, he thinks it's not going to happen because there are other inputs that go into improving AI systems that will become bottlenecks with abundant researcher effort
your claim doesn't respond to that at all
@EgeErdil2 @MatthewJBar you’re right im just giving you squishy intuitions but these are my intuitions
@tszzl @MatthewJBar in a world where r&d progress in AI is itself compute-bottlenecked, it can simultaneously be true that:
1. we're far from the "landauer limit" of efficiently using compute and data during training, inference, etc
2. we can't close the gap to it by scaling cognitive effort alone
@tszzl @DicksonPau @MatthewJBar The question on my mind is if scale and capability are truly tightly coupled, or if this is a partial consequence of current inefficiency.
@MatthewJBar I think you’re wrong and there’s 1,000x efficiency gains leftover in deep learning research that could lead to much smarter faster more agentic models given the same inputs
@tszzl @MatthewJBar checking myself by skimming Machines of Loving Grace...but I'm failing to see the societal difference between RSI and what's currently available. we still have Amdahl's law
@tszzl @MatthewJBar wouldn’t that just risk letting other countries lead?
@kagayakikiki @tszzl @MatthewJBar Those workplaces will go away too. The economy will be a red queen's race between the first few companies that managed to have the ASI run everything end to end without a human outer loop slowing them down. It's a numbers game and RSI makes the numbers go vertical.
@absenteewarlord @tszzl @MatthewJBar I kind of envision most workplaces (non-software or other science categories) where AI makes recommendations and the humans have to talk about it to make sure they agree, but that’s kind of what’s emerging today without RSI
@kagayakikiki @tszzl @MatthewJBar I suspect Machines of Loving Grace very intentionally didn't make RSI load-bearing. RSI just makes the discovery/capability/productivity exponential *much* steeper. Normal, human-driven progress would lead to the same place on a slightly longer timescale.
@absenteewarlord @tszzl @MatthewJBar the question then would be which parts of the economy will humans not allow AI to take over. The Pope emphasizes the importance of people being able to work, and work being central to human identity. The scope of RSI & human pushback seems core over the next couple decades
Some thoughts on agent time horizons, on-task data being crucial: https://www.oliversourbut.net/p/a-slightly-mechanistic-theory-for
" As with humans, AI can extrapolate a bit, but need ‘experience’ and examples to succeed. Importantly, this means that vast in silico training ranges for software, cyber, and mathematics very likely won’t transfer much to other domains of interest, like interpersonal intelligence, medical discovery, bioweapons development, intelligence analysis, and robotic manipulation. Of course, like with every domain of human experience and activity, we have some relevantly-similar data already collected, and schemes can be devised to more rapidly expand that digitised experience bank for AI to learn from. Increasing adoption of AI in task-integrated contexts, industrial deployment, and even explicit approaches to gathering example data such as ‘hand movement farming’ are the leading indicators to watch for progress in particular domains — not just the headline benchmark metrics in software-like tasks.
" The best case I can make for a much more general explosion is if the speed and cost-effectiveness explosions rapidly accelerate the gathering and digestion of diverse task data — but I think that remains mostly rate-limited in the familiar ways: some domains easy and some more difficult. Don’t mistake me for ruling out across-the-board AI capability! Companies are charging ahead with data collection and set on automating much of their AI production pipeline. It just won’t happen overnight. "
Ironically vague response! Presumably Ege would say there's headroom, but the activities we're on track to automate early aren't bottlenecking progress (much); rather compute and data are binding. I'd probably agree. Of course data/curriculum collection/curation (soon?) and compute manufacture (later?) may also be in reach, which changes things.
@tszzl @EgeErdil2 @MatthewJBar Crude biological wet brains are obscenely powerful. Evolution built them with dice. Engineered synthetic compute has no obvious near term efficiency floor.
@tszzl @EgeErdil2 @MatthewJBar This is accurate posting
@tszzl @yacineMTB @MatthewJBar I dunno about 1,000X but I certainly believe there’s 10-50X. But this is just efficiency and optimization… which is both a roadblock and a moat for the major labs.
@tszzl @MatthewJBar Simply impossible, we've already hit the hard physical limits of intelligence possible in this universe. We have no other existence proof of something more efficient than current systems. 😫
Roon says optimizations could bypass data and compute constraints
@MatthewJBar I think you’re wrong and there’s 1,000x efficiency gains leftover in deep learning research that could lead to much smarter faster more agentic models given the same inputs
@tszzl I don't think it's a good development. I continue to think that RSI is an overrated risk vector due to data and compute bottlenecks, and that slowing down AI would accomplish little at enormous cost.
