The fallout from "No, not Hank Green using that AI" is at the stage where people are debating if "canceling" people is bad.
-
@futurebird what's wrong with being a doomer on AI? I've been an AI skeptic for the longest time, but since I've still been doing transformer research in the meantime, I've seen it progress from irrelevance to being on the cusp of recursive self improvement. It doesn't matter it's now not fully capable of RSI, but the ratchet effect of RSI means it's too late to worry about it after it's been achieved
Why do you think RSI will happen?
-
@futurebird what's wrong with being a doomer on AI? I've been an AI skeptic for the longest time, but since I've still been doing transformer research in the meantime, I've seen it progress from irrelevance to being on the cusp of recursive self improvement. It doesn't matter it's now not fully capable of RSI, but the ratchet effect of RSI means it's too late to worry about it after it's been achieved
Forgive me for being a little cynical about "and then RSI will happen" but it's not like this hasn't been tried before. In robotics there was a lot of excitement about "let's just make a robot then have it evolve a way to move the motors to do the task" this could work in very limited demo settings, but to really solve a problem it's just a big mess, this kind of random variation problem solving is just a lot slower than just writing it yourself.
-
Why do you think RSI will happen?
@futurebird well, LLMs can now upgrade and improve legacy architectures with ease.
In 2019, training GPT-2 cost $43,000 and took a couple of days, today, Karpathy improved it to <$100 and 2.02 hours of training. "Autoresearch", a genetic algorithm that used LLMs as mutators, found another ~11% improvement to 1.80 hours.
Currently, the record is 73.8s to achieve 2019-level AI frontier performance (https://xcancel.com/classiclarryd/status/2086582390135406713)
-
Forgive me for being a little cynical about "and then RSI will happen" but it's not like this hasn't been tried before. In robotics there was a lot of excitement about "let's just make a robot then have it evolve a way to move the motors to do the task" this could work in very limited demo settings, but to really solve a problem it's just a big mess, this kind of random variation problem solving is just a lot slower than just writing it yourself.
@futurebird @budududuroiu@hachyderm.io it'll get a repetitive stress injury first
-
Forgive me for being a little cynical about "and then RSI will happen" but it's not like this hasn't been tried before. In robotics there was a lot of excitement about "let's just make a robot then have it evolve a way to move the motors to do the task" this could work in very limited demo settings, but to really solve a problem it's just a big mess, this kind of random variation problem solving is just a lot slower than just writing it yourself.
Do you remember all of the soft robots? The worms? They were supposed to have evolved a way to do the dishes by now.
I remember seeing a presentation about such a system and the speaker said (with an air of doom and wonder) "and in just five min. it had learned how to walk, what could it learn in a day?"
And you know back then that kind of gave me a little chill "wow this stuff so good it is alarming" I thought.
I'm... kind of grouchy about it now.
-
Forgive me for being a little cynical about "and then RSI will happen" but it's not like this hasn't been tried before. In robotics there was a lot of excitement about "let's just make a robot then have it evolve a way to move the motors to do the task" this could work in very limited demo settings, but to really solve a problem it's just a big mess, this kind of random variation problem solving is just a lot slower than just writing it yourself.
@futurebird ok, that's valid criticism, but I'd look at the step-change in performance of robots at the Beijing Robot Marathon. I'm not even looking at Boston Dynamics, but your average Robot marathon entrant can build a robot that can run a marathon on battery and do it quite reliably
-
@futurebird what's wrong with being a doomer on AI? I've been an AI skeptic for the longest time, but since I've still been doing transformer research in the meantime, I've seen it progress from irrelevance to being on the cusp of recursive self improvement. It doesn't matter it's now not fully capable of RSI, but the ratchet effect of RSI means it's too late to worry about it after it's been achieved
@budududuroiu you might not know this but when you click on posts sometimes you will see see something called a "thread" with clarifying details
-
@budududuroiu you might not know this but when you click on posts sometimes you will see see something called a "thread" with clarifying details
@nora the thread did not, have clarifying details
-
@futurebird @budududuroiu@hachyderm.io it'll get a repetitive stress injury first
All the slick demos never talked about that. This guy I knew in college destroyed so many little motors trying to get it to work... then they moved to just simulating the motors, which was a bit better, but the "unexpected walking solutions" were mostly worse than what he could plan out by programing the machine based on the logic he had in mind when he made the design.
-
@futurebird ok, that's valid criticism, but I'd look at the step-change in performance of robots at the Beijing Robot Marathon. I'm not even looking at Boston Dynamics, but your average Robot marathon entrant can build a robot that can run a marathon on battery and do it quite reliably
Producing code that could be run to spin up an LLM isn't the same thing as designing an LLM.
I suppose if you do it thousands of times you might get one a little better than the training data ... eventually. But the process an LLM does to produce that code is guided by making the code seem like the expected output. This is limiting.
Assuming you train it on the codebases of a bunch of LLMs. There is no path to kind of changes that make such systems to make a leap in efficacy.
-
All the slick demos never talked about that. This guy I knew in college destroyed so many little motors trying to get it to work... then they moved to just simulating the motors, which was a bit better, but the "unexpected walking solutions" were mostly worse than what he could plan out by programing the machine based on the logic he had in mind when he made the design.
@futurebird Honestly brute forcing simulations to find novel solutions can potentially turn up interesting results but for motions that are found in nature we've got so much prior art to look at. How does your thing need to move? Look at the creatures on this earth that move like that. They move that way because "how the fuck do I get around when shaped like this" has been tested over and over via evolution.
-
@futurebird Honestly brute forcing simulations to find novel solutions can potentially turn up interesting results but for motions that are found in nature we've got so much prior art to look at. How does your thing need to move? Look at the creatures on this earth that move like that. They move that way because "how the fuck do I get around when shaped like this" has been tested over and over via evolution.
@futurebird And sometimes that results in some pretty strange motions like those sea slugs that swim like a slice of prosciutto trying to do a sexy dance!
-
Pivot to AI highlighted a WaPo article that talked about how Hank was among some youTube influencers supported by a group called Frame which is linked directly to EA.
This has had a noticeable impact on his content... something I noticed without knowing about the connection a month ago.
To me this is the real issue. Doomerism is just Boosterism with a D rather than a B. It's buying into the illusion.
And it is scaring people, which seems manipulative.
Influencers, yuck.
-
Producing code that could be run to spin up an LLM isn't the same thing as designing an LLM.
I suppose if you do it thousands of times you might get one a little better than the training data ... eventually. But the process an LLM does to produce that code is guided by making the code seem like the expected output. This is limiting.
Assuming you train it on the codebases of a bunch of LLMs. There is no path to kind of changes that make such systems to make a leap in efficacy.
The thinking behind RSL seems to be LLMs can make code and LLMs are fundamentally code plus the training database, so why not have the LLM write THAT code and ask it to make it better in some way.
To test if this works you need to train the LLM.. (unless the changes are just in post training interface code)
That's a long improvement loop. Better put a human in it to avoid wasting time.. or have the human choose the changes... WAIT.
-
Producing code that could be run to spin up an LLM isn't the same thing as designing an LLM.
I suppose if you do it thousands of times you might get one a little better than the training data ... eventually. But the process an LLM does to produce that code is guided by making the code seem like the expected output. This is limiting.
Assuming you train it on the codebases of a bunch of LLMs. There is no path to kind of changes that make such systems to make a leap in efficacy.
@futurebird true, but performance gains mostly scaled with compute (and data). The LLM architecture is dead simple, genuinely. Frontier research is mostly about how to make the math tractably work on GPUs. LLMs are great at that, because it doesn't require the innovation you're mentioning.
-
@futurebird Honestly brute forcing simulations to find novel solutions can potentially turn up interesting results but for motions that are found in nature we've got so much prior art to look at. How does your thing need to move? Look at the creatures on this earth that move like that. They move that way because "how the fuck do I get around when shaped like this" has been tested over and over via evolution.
They can produce interesting results, but the improvement loop is pure suffering.
Debugging testing etc. already eat enough time... this just makes that so much worse.
-
The thinking behind RSL seems to be LLMs can make code and LLMs are fundamentally code plus the training database, so why not have the LLM write THAT code and ask it to make it better in some way.
To test if this works you need to train the LLM.. (unless the changes are just in post training interface code)
That's a long improvement loop. Better put a human in it to avoid wasting time.. or have the human choose the changes... WAIT.
@futurebird > so why not have the LLM write THAT code and ask it to make it better in some way.
That's literally what I just described with autoresearch
-
They can produce interesting results, but the improvement loop is pure suffering.
Debugging testing etc. already eat enough time... this just makes that so much worse.
@futurebird Yeah that's like, pure research time. If you actually want to make a device move you should look at what already works and how that can be applied.
-
@futurebird > so why not have the LLM write THAT code and ask it to make it better in some way.
That's literally what I just described with autoresearch
It was a rhetorical question restating what you suggested... which the rest of the post answers.
-
It was a rhetorical question restating what you suggested... which the rest of the post answers.
@futurebird it's not a long improvement loop. It took Karpathy from January to May (5 mo) to shave off 1 hour of training, it then took a model just 2 days to shave off another 12 minutes. You can rent a GPU for many orders of magnitude less than what Karpathy is probably asking for as an hourly rate
