In any case, here our work and predictions link.springer.com/article/10.1... - ThreadSky

irisvanrooij.bsky.social • 204 days ago

In any case, here our work and predictions

https://link.springer.com/article/10.1007/s42113-024-00217-5

Comments

This feels like a modified version of Gold's Theorem (see Language Identification in the Limit). I gave a talk last year where I covered a bit of Gold's Theorem in the context of LLMs: https://youtu.be/xiDPfLpFxK0?t=574

Do you have a link to the full paper?

bsky.chewxy.com•204 days ago

Nvm. Saw the link. Thought that was a link to the box. Thanks for the read.

efratfurst.bsky.social•204 days ago

📌

velezbeltran.bsky.social•204 days ago

I like the back of the envelope calculation and also disagree with the hype! However, I don't think I agree with this argument. It implicitly assumes that the data is not structured which is not true. Not all functions are possible and the inductive biases are much stronger for the real world.

velezbeltran.bsky.social•204 days ago

As an extreme example, if I could assume that the function which I am learning is constant, the problem would be trivially learnable. Although we are not in that case, I still think we are in a much more structured scenario than the one you assume.

A related paper: https://arxiv.org/abs/2304.05366

irisvanrooij.bsky.social•204 days ago

The argument does not assume the data is unstructured; in fact the proof argument grants that the data is generated by a structured and tractable algorithm.

velezbeltran.bsky.social•204 days ago

Yeah you are right. I think I wrote something not fully correct. Here is my argument a bit more laid out. I would be super interested in knowing what you think! 🧵 (1/n)

velezbeltran.bsky.social•204 days ago

What you demand in the AI-BY-LEARNING problem is very strong. You require an algorithm that simultaneously works for every single integer K and every possible distribution D. In other words, for every single possible K and D your algorithm must return a reasonable L_A.
(2/n)

velezbeltran.bsky.social•204 days ago

You use this fact strongly to reduce PERFECT-VS-CHANCE to AI-BY-LEARNING. However, what the proof doesn't rule out is a situation where some learning algorithm M works very well for a subset of all possible distributions D while still failing in the learning problem more broadly.

(3/n)

velezbeltran.bsky.social•204 days ago

What I meant was that I think that in real life we find ourselves in a situation where we want learning algorithms that perform well only for a subset of all distributions D. The properties that define such a set of distributions is what I imprecisely called structure.

(4/n)

moultano.bsky.social•204 days ago

Why would this not rule out human intelligence if it were true?

arhiippa.bsky.social•204 days ago

While the argument allows for much better technology than currently available, there might be (are) more efficient ways to produce human like intelligence. We however do not not have access to those computationally and would need something paradigm changing to reach those.

ahundt.bsky.social•191 days ago

@cpaxton.bsky.social you should read this if you haven’t yet!

irisvanrooij.bsky.social•204 days ago

kellyms.bsky.social•204 days ago

📌

maketsufi.bsky.social•204 days ago

I tend to think that AGI/ASI is our only and last chance to survive as a species. If it doesn't pan out, humanity is royally screwed.

geodesicvoyager.bsky.social•204 days ago

Citation needed. I mean where is that claim even coming from?

There is near universal consensus among environmental and climate scientists that we already have most of the technology needed for stopping climate change and climate cascades. What we lack is the political will. AGI won’t change that

maketsufi.bsky.social•204 days ago

Rapid technological advancement provided by AGI could force the hand of the politicians, or bypass them altogether. Current political systems are way too slow and rigid to address our problems as a species (climate change being the most urgent).

geodesicvoyager.bsky.social•204 days ago

How will that even work man? Corporate AGI will be beholden to corporate interests; maximize next quarterly earnings. And it’s unlikely there’s will only be one AGI when it is here, you’ll have a bunch of them moving fast and braking things.

geodesicvoyager.bsky.social•204 days ago

Whats the confidence that this technological advancement take us on a positive trajectory under the guiding hand of corporate world? Or is the hope that AGIs will forcibly take over the world govts.

maketsufi.bsky.social•204 days ago

I see AGI as our ultimate hail Mary technological solution to our problems (which we have created ourselves). I know it is a very slim chance, but that may be all we hot anymore at this point.

geodesicvoyager.bsky.social•204 days ago

There are a lot of ifs hiding in this long shot. If all the people (and there are a lot of brilliant people) who are trying to make this happen instead focus on lobbying and advocating to decarbonize now, we’d be solving this problem much more easily. But I guess there isn’t a lot of money in that.

mnemonicer.bsky.social•204 days ago

Considering the energy, hardware resources, semiconductors, personnel, data and IP resources needed to stand up a semi-functional AI let alone achieve AGI, this is an odd take.

therion5.bsky.social•202 days ago

we can’t exclude possibility such super intelligence already exists, but making our own with black jack and hookers doesn’t make sense for me. The humanity is a big experiment of nature, either we learn next step or we are off the table.

brightmarsh.bsky.social•204 days ago

Why do you think a thing humans will build can magically solve all our human problems? You're deluding yourself. Human problems need human solutions. Something for profit ai companies are wholly incapable of delivering on. Find your hope somewhere else

maketsufi.bsky.social•204 days ago

Like I said, AGI is a desperate hail Mary attempt. We can't solve the climate change with mere human intellect anymore.

psychicteeth.bsky.social•203 days ago

It's not about intellect. We already know how to solve climate change. We've known for over half a century.

irisvanrooij.bsky.social•203 days ago

💯 https://bsky.app/profile/irisvanrooij.bsky.social/post/3lcccrs3xcc22

irisvanrooij.bsky.social•204 days ago

It won’t pan out. Sorry

maketsufi.bsky.social•204 days ago

Then it is extinction for us as a species. We need vastly accelerated scientific and technological development in order to survive the enviromental problems we have created for ourselves.

leonderczynski.bsky.social•203 days ago

Political ones haven't saved us, quite the opposite, but tech isn't coming to save us either

essaywells.bsky.social•203 days ago

Rubbish. The technological solution to climate change, for example - silicon solar PV - was invented in 1939. It has been a political and economic choice not to use it.

maketsufi.bsky.social•204 days ago

For example, plankton will likely die out when the oceans warm up enough. We need a solution for that, otherwise we will need to learn not to breathe anymore (plankton produces most of our oxygen).

monkey69420.bsky.social•204 days ago

Which model is projecting that?

irisvanrooij.bsky.social•204 days ago

AI will not provide solutions. Only humans can have scientific breakthroughs.

irisvanrooij.bsky.social•204 days ago

We already knew scientifically what we needed to do to stop climate change. We as a species decided not to do it. All AGI can give us is vastly accelerated pollution.

ericandriolo.bsky.social•204 days ago

"as a species" makes this seem way more of a democratic decision than it ever was

ournewhomecoach.bsky.social•204 days ago

Bugger.

maketsufi.bsky.social•204 days ago

Then we have at least one solid solution to the Fermi Paradox: technological species wreck their planet so fast that it leads to their extinction as a technological species. The universe is likely filled with once-alive planets that are now trash heaps of plastic.

chrislarsen.bsky.social•204 days ago

Agreed. We already have the technology to "clean up our act," but our politicians are in the pockets of billionaires who are actively opposed to it.

vash2695.bsky.social•203 days ago

The intractability proof is elegant math, but it's like saying motion is impossible because you can't solve Zeno's Paradox. Both biology and physics show us that perfect theoretical solutions aren't required for practical emergence. General intelligence may work in the same way!

irisvanrooij.bsky.social•203 days ago

Our proof is not about human-level intelligence itself, which the proof argument grants is tractable. The proof argument is about the *making* of human-level intelligence. Here the full paper: https://link.springer.com/article/10.1007/s42113-024-00217-5

vash2695.bsky.social•202 days ago

Your proof demonstrates why directly engineering intelligence is intractable, but given that biology achieved general intelligence through emergence rather than explicit design... maybe soup is all you need?

irisvanrooij.bsky.social•202 days ago

AI-by-Soup? Interesting idea

vash2695.bsky.social•202 days ago

Right? Maybe evolution had the right idea - throw interesting ingredients into the soup and let complexity emerge on its own. Nature didn't need to solve intractability proofs to cook up consciousness, so why should we?

Comments

Posting Rules

Reply