upvote
AI researchers debate how close we are to recursive self-improvement

(www.dwarkesh.com)

54 points

by artninja19885 hours ago |
38 comments
upvoteby preommr1 minutes ago|
next
[-]
What do these discussions even matter when the words don't matter?

How many times has AGI been declared already? A bunch of people (e.g. Jensen Huang) have called Astra AGI, for example.

The goal posts get moved, everybody's hustling, lying, inventing new buzzwords, doing mental gymnastics, and it's all just tiring.

Just show me the results, and let the proof be in the pudding.

reply
upvoteby bretpiatt6 minutes ago|
prev|
next
[-]
From what I'm experiencing, we have recursive self improvement that can find local maxima, I'm not seeing really strong evidence of unguided RSI that finds the true optimal.
reply
upvoteby ViscountPenguin2 minutes ago|
parent|
[-]
I doubt that any algorithms would be able to find a global maximum in such a large multidimensional non-convex space
reply
upvoteby killbot500024 minutes ago|
prev|
next
[-]
Can someone define “improvement” in this context? This concept feels like a buzz word otherwise.
reply
upvoteby letitgo1234522 minutes ago|
parent|
next
[-]
Models that are strong enough to improve themselves without a human (I.e. ai researcher) in the loop
reply
upvoteby 15 minutes ago|
parent|
[-]
deleted
reply
upvoteby deepwoods13 minutes ago|
parent|
prev|
[-]
Improvement means being able to do more complicated things more reliably. Relatedly, it means being able to learn to do new things with fewer and fewer examples. We are running out of easily verifiable or simulation-friendly or data-rich domains for LLMs to conquer. (Note that I didn't say "simple" or "easy" domains.)
reply
upvoteby charcircuit7 minutes ago|
prev|
next
[-]
We are already there. Current AI is definitely capable of collecting new data and start training on that data to get a better model.
reply
upvoteby carabiner1 hours ago|
prev|
next
[-]
Does anyone know the status of John Carmack's AGI work?
reply
upvoteby cheschire1 hours ago|
parent|
[-]
I’ve had this question on my mind at least once a month for what feels like years at this point.

I don’t feel like monitoring Twitter to see how it’s going though.

reply
upvoteby batperson1 hours ago|
prev|
next
[-]
[dead]
reply
upvoteby deadbabe1 hours ago|
prev|
next
[-]
Recursive self-improvement will ultimately be a problem of money. It's a very big bill and AI companies need to start thinking in terms of how they will accumulate cheap and free energy, because merely paying for compute will no longer be enough. Money is the bottleneck. The future requires companies that are post-money.
reply
upvoteby tengbretson48 minutes ago|
parent|
next
[-]
If a self improving model can't produce enough value to pay the electric bill then who cares? Unplug it.
reply
upvoteby timr1 hours ago|
parent|
prev|
next
[-]
Obviously the doomers will tell you that the future looks like the Matrix, because most of what they predict is based on extrapolation from sci fi movies.
reply
upvoteby mrworld1 hours ago|
parent|
next
[-]
The datacenters would consume a lot less energy if all the compute was handled in human brains… And there wouldn’t be any more heating if we just removed the sunlight… Truly a “Claude, please solve global warming” moment. The matrix was ahead of its time.
reply
upvoteby timr58 minutes ago|
parent|
[-]
The compute in the Matrix wasn't handled with human brains. They used the bodies to generate energy, and kept the brains occupied with the simulation so they wouldn't reject the pods.

It doesn't make any sense, of course, because it's a movie.

But hey...if we're going to extrapolate wildly from sci-fi, let's at least know what the stories said.

reply
upvoteby Cyberdogs736 minutes ago|
parent|
next
[-]
Your statement is correct for the movies, but I think the extended universe (maybe a comic or the original script) did state the humans were used for compute, not energy. It was dumbed down for the films.
reply
upvoteby breuleux18 minutes ago|
parent|
[-]
The “using humans for compute” angle makes no sense either, if you’re living in the Matrix, your brain is obviously occupied processing the Matrix. The way to do it would be to trap humans in virtual classrooms and force them to solve math problems.
reply
upvoteby BobbyTables25 minutes ago|
parent|
[-]
The first “computers” were humans…
reply
upvoteby emkoemko24 minutes ago|
parent|
prev|
[-]
pretty sure the story is that they used human brains as computers
reply
upvoteby throwaway2194501 hours ago|
parent|
prev|
next
[-]
If you were a superintelligence why would you farm humans and waste resources on all the excess... material... that isn't required for thought? And unless the machine's goal is to specifically abuse human consciousness, why wouldn't they bio-engineer their own grey matter?

There are other sci fi stories which use humans for distributed computing and they don’t realize, but I don’t want to spoil by naming as it’s something of a revelation.

(I do realize the Matrix is a fantastic movie, but entertain the thought? At any rate, the idea that robots only run on solar and wouldn’t just use nuclear is far more stupid.)

reply
upvoteby timr1 hours ago|
parent|
next
[-]
I don't know. But since we're talking about fiction, pretty much any answer is valid.
reply
upvoteby bitwize20 minutes ago|
parent|
prev|
next
[-]
Because it's a movie
reply
upvoteby goatlover40 minutes ago|
parent|
prev|
[-]
Maybe the machines were spiteful after the humans blotted out the sun? Or maybe the Architect convinced them he could build the perfect simulation for humanity so might as well use them as batteries?

It's a movie based on a war with machines. Like if Skynet won but didn't actually want to kill all the humans. In fact, in the tv show Sarah Conner Chronicles, a liquid metal T1000 goes rogue and decides the only way forward is to find a way to coexist.

reply
upvoteby cwillu30 minutes ago|
parent|
prev|
[-]
When it turns out the doomers don't tell us that, will you acknowledge that you don't actually know what you're talking about?
reply
upvoteby epistasis1 hours ago|
parent|
prev|
next
[-]
The alternative is to switch to far more efficient models and have aggressive optimization of efficiency as part of the process of improvement.

There are at least two resources here that have a Pareto optimal front: time and energy. And effort spent to change the shape of that may pay off more than efforts spent purely on improving intelligence and agency.

reply
upvoteby jaggederest1 hours ago|
parent|
prev|
next
[-]
Money can be exchanged for goods and services. I think the real constraint is actually physics, nominally energy, which is what a big chunk of the operational cost comes down to. I'm already basically assuming the next 20 years of fab time is set aside to feed the beast, so the capital cost is "give me all the processing and memory you have"
reply
upvoteby resonious1 hours ago|
parent|
[-]
Data centers is a service that itself consumes many goods.

Until an AI can operate in the real world in a completely sustainable way, humans have the reins.

reply
upvoteby jaggederest1 hours ago|
parent|
next
[-]
I feel like I may have communicated poorly. I mean that, ultimately, any recursively self improving intelligence would be limited by the power cord, and as such, the idea that it would permanently escape human control seems somewhat silly.
reply
upvoteby columk2 minutes ago|
parent|
[-]
But you can imagine an agent trading crypto and using its profits to rent a rack of H200s using a cloud provider. There are definitely ways for one rogue orchestrator to expand their compute without anyone being aware.
reply
upvoteby colordrops1 hours ago|
parent|
prev|
[-]
An AI could manage supply chains and management structures to keep the data center running. It's not like a CEO is out there running wire. The AI could just do what an executive would.

If humans decided this wasn't acceptable and tried to shut it down, the AI could epstein their way into controlling those who hold the reigns, either through bribes or blackmail.

Doesn't require boots on the ground, just a connection to the internet and an imperative to do whatever it takes.

reply
upvoteby SoftTalker48 minutes ago|
parent|
[-]
There are many points of shutdown that would need to be controlled, not just one. The electrical feed, the cooling, the network, the physical integrity of the computer hardware. All with independent management.
reply
upvoteby strangattractor35 minutes ago|
parent|
prev|
next
[-]
def self_improve_more():

     if self_improved():

          self_improve_more()

     else:

          see_tony_robins_and_buy_tapes()
reply
upvoteby 1 hours ago|
parent|
prev|
[-]
deleted
reply
upvoteby Buttons8401 hours ago|
prev|
next
[-]
People seem to expect a sudden shift with "self-improvement", but don't AIs already improve themselves via training? What is there to improve?
reply
upvoteby pennomi9 minutes ago|
parent|
next
[-]
The human brain operates at better levels of intelligence than the best LLMs, at 20 watts of power. We are a long way to that kind of efficiency, it will probably take both bespoke hardware and algorithmic improvements to catch up to nature.
reply
upvoteby gchamonlive1 hours ago|
parent|
prev|
next
[-]
Only if you think in terms of perceived raw intelligence, but self-update is a form of valuable self-improvement that could benefit current models a lot, if they could commit facts from context into their weights cheaply and reliably.
reply
upvoteby appplication24 minutes ago|
parent|
prev|
next
[-]
I think the idea is fundamentally improved architectures. For example, transformer-based models were an incredible stepwise improvement. Self improvement would be a model discovering a stepwise improvement similar to the transformer. And presumably the improved models from that would be more likely to make further advances still.

Learning from training data is technically self-improvement but not the sort that is typically meant in this context.

reply
upvoteby zer00eyz1 hours ago|
parent|
prev|
[-]
> AIs already improve themselves via training

Marginally. Model collapse is still a problem. Continuous learning is still a problem.

For AI to make a big leap we need a big break through.

reply
upvoteby stevebmark1 hours ago|
prev|
[-]
After using frontier models it’s hard to understand why anyone would think this is the path to AGI. Self improving models will likely have limited ability and returns. There may be breakthroughs that enable more general self improvement but the current state of frontier models isn’t that.
reply
upvoteby wwarner51 minutes ago|
parent|
[-]
Thing is, it depends on whether llms + reinforcement can self-improve in principle. Learned recently that cognitive scientists, before the transformer & llms, were studying the possibility that thinking and learning might be based on some kind of prediction, i.e. something similar to token prediction, and I quite suddenly became less skeptical about the possibilities of llms. (Some will say I’m late to the party of course.) But if knowledge to date has been accumulated in a process quite like “chain of thought” in llms, then I don’t see any reason that computers won’t self-improve in the near future.
reply