Artificial General Intelligence (AGI) News and Discussions

[Yuli Ban]

starspawn0

There are several steps separating this from what you might call "AGI" -- e.g. long-term memory (not just x number of tokens, where x is small); a sense of time (instead of just a sequence of tokens where each could be separated by an indefinite duration; this could be solved by training on voice and/or video, both of which contain a sense of time); and need for multi-modal, which might help make it learn representations quicker than just using text.

One capability that would be interesting to see in these models is self-awareness. They may already have some degree of it. What would that look like? If you ask it to answer a question, and then ask it to explain what its weakness was in doing so, and it comes back with, "I wasn't very confident with that answer. I guessed correctly." -- and then you check the neural net's internal dynamics, and find that, indeed, the probability distribution for answers was spread out in such a way as to suggest "uncertainty". If all that happens, then it would suggest that the model has some degree of self-awareness (and you'd want to check lots more examples of how its self-analyses compare with its own internal dynamics).

If models attain some level of self-awareness, then they would know what they are weak at, and suggest how to improve. They could tell you what you should be doing to make them smarter.

[Yuli Ban]

What does he know that we don't....?

[Yuli Ban]

[Yuli Ban]

Starspawn0

Incidentally, for those who don't know, Szegedy comes from a very distinguished mathematical family, and he happens to also be a math ph.d. with significant theoretical accomplishments -- though, most people in the machine learning world know him from his other work, for example on batch-normalization and also adversarial examples. His mathematical intuition is probably deeper than most of the other people at Google (certainly including Chollet, but also Hinton, Hassabis, Jeff Dean, and others...).
Addendum: Boaz Barak (theoretical CS guy at Harvard of considerable acclaim) weighs in:
https://mobile.twitter.com/boazbaraktcs/status/1512488217177575430

I would actually be surprised if it takes 500bn parameters. I would guess that:
(1) A moderate size neural net could easily learn to distinguish between the "left" distribution (identical shapes) and "right" distribution (different shapes) and also generate new examples.

Szegedy and Barak are in a whole other league from most of the people responding in that thread (far beyond poor Melanie Mitchell).
Addendum 2: Michael Nielsen also weighed-in:
https://mobile.twitter.com/michael_nielsen/status/1513283862905835521

My guess is that there will be a system - indeed, very possibly one based on zero-shot learning - much better than humans at solving Bongard problems sometime very soon, if people prioritize it.

....
This is the magic decade when all the major problems of AI get solved. It's a great time to be alive to witness it.

[Yuli Ban]

[Yuli Ban]

My prediction is that by 2025, we'll see personal assistants similar to Samantha from Her. And if the API is publicly available, anyone could talk to such assistants even on their phones.

Turing Test? Probably will be passed either this year or next. 2024 at the latest.

[wjfox]

The results of our poll (199 votes).

[andmar74]

My prediction is that by 2025, we'll see personal assistants similar to Samantha from Her. And if the API is publicly available, anyone could talk to such assistants even on their phones.

Turing Test? Probably will be passed either this year or next. 2024 at the latest.

Isn't Samantha AGI or very close? It's been a long time since I saw the movie, I will see it again if I can find it.

I think that's too early for the Turing test. The AI can't fool me yet. I would just ask it some tricky science questions and the answers will be gibberish or half-gibberish.

[Yuli Ban]

6 Year Decrease of Metaculus AGI Prediction -- cross-post from /r/slatestarcodex.
All-DayErrDay

It's now 9 years earlier. I would say if more big (or at least big feeling) things come out in a year that it will regress to 2029 - 2031. If things stay on track it isn't a bad guess for Weak AGI. Weak AGI in my opinion is not AGI but instead superhuman/on par at enough tasks to increase human productivity significantly, still lackluster enough to not be able to fill quite a few important human roles and is on the cusp of being an AGI (with maybe 5 - 10 more years if sufficiently difficult to truly puff out any of the harder weaknesses).

But I guess what I just said doesn't actually coincide with my personal distributional timeline. Somehow for me, 10 years simultaneously feels like a good estimate for when 'Weak' AGI will be available but that true AGI won't probably be here until roughly 20 years from now. However, that doesn't actually make sense in many ways. How could you get to a weak AGI without getting to a true AGI within 3-5 years of that?

I haven't really worked it out in my mind yet. I guess it feels like AGI is just too hard to attain in 10-15 years. Yet, 20 years feels good, but I also think weak AGI in 10 years. I don't know...

After thinking, my overall bet is weak AGI by 2032, then AGI by 2039-47, and then ASI within 1 month - 3 years of AGI. My probabilities are .8, .8 and .9. So by my current estimates I think there is an 80% chance of weak AGI by 2032, 64% chance of AGI by 2047 and a 58% chance of ASI by 2050 (some biases honestly push the odds a bit lower in my mind, like instead a .60, .60, .90 distribution, and another optimistic side that says .85, .90, .95, but I'll stick to my first estimate for the time being as slightly optimistic priors).

[raklian]

I think that's too early for the Turing test. The AI can't fool me yet. I would just ask it some tricky science questions and the answers will be gibberish or half-gibberish.

But wouldn't you get the same result if you ask a layperson the same questions?

[Yuli Ban]

I think that's too early for the Turing test. The AI can't fool me yet. I would just ask it some tricky science questions and the answers will be gibberish or half-gibberish.

A few thoughts on this:

First of all, that's part of a somewhat unfortunate fact: whatever AI you've used is almost certainly NOT the state of the art. Even GPT-3 is no longer leader of the pack in any regard except visibility— it had been surpassed last year and especially moreso this year.
https://super.gluebenchmark.com/leaderboard
Something like PaLM or Chinchilla would be a more effective gauge of where AI stands

Of course, second of all, the Turing Test is likely already passable. This because it's already been passed multiple times by earlier chatbots
https://en.wikipedia.org/wiki/Eugene_Goostman#Reaction
This because it only needs to convince 30% of judges that it's human.

If simple scripts, Markov Chains, and LSTM networks of the 2000s and early 2010s could do it, there's no reason why a modern transformer fine-tuned on dialogue couldn't crush the Turing Test as we speak. For whatever reason, it seems no one's tried.

Of course, there are some edits done to the Turing Test to make it more meaningful. For example, upping the score it needs to pass from 30% to 90% would go a long way. Also helping would be extending the amount of time the test is held, from a few minutes to half an hour or an hour or longer.



Personally, I look upon things like Eugene Goostman passing the Turing Test much the same way I look at that AI that beat a human at Go back in 2008.

Where it only won against a low-level champion, on a small 9x9 board. Goostman was given handicaps, such as being an ESL Ukrainian 14-year-old to "excuse" grammatical errors and non-sequiturs.
In comparison, AlphaGo won with no handicaps, fair and square. A modern transformer-powered dialogue agent should also be able to win fair and square, without need for any excuses.

[Ozzie guy]

Kurzweil has a much stricter version of the Turing test I suppose you could call it the Kurzweil test. https://longbets.org/1/

This could not be passed now and seems a better measure of us being close to AGI although I don't think anyone will care about his version unless he lives old enough to make good on his Turing test bet.


As a personal note I think I am going to read into the Metaculus AGI predictions and change my AI predictions to just following what they say.

I get the impression that the average view of those scientists will be the most realistic view possible (and it is also flexible to change over time) and I need to stray away from optimism in favor of realism to not let my self down.

[Tadasuke]

Proto-AGI may be a few years away, but I'm not seeing any substantial improvements to our daily lives thanks to AI.

Some examples from my life: Google search doesn't work better today than in 2016 (sometimes works worse), search functionality in Skyscrapercity doesn't work better than in 2002 (really), DeviantArt doesn't adapt to your preferences at all (your history, works, favourites, comments or things clicked as "see less like that in the future"), voice assistant function on smartwatches doesn't work better than a few years ago (still works terribly and is useless). I just don't see, hear or feel any exponential improvement in AI in my daily life at all, with the exception of language translators (which still can't translate from camera or photos) and YouTube recommendations.

I was so f***ing excited about AI when the 2010s were starting and here we are 12 years later and there has not been much of a difference. I'm typing this on a keyboard instead of dictating to a virtual assistant like Winston Churchill to his human assistant. IBM Watson won Jeopardy in February 2011. So little has changed since then in practically and widely used AI. I'm observing it, following it, but I can't feel the supposed big difference that exponential AI makes. All the stuff 'Two Minute Papers' shows, doesn't influence me a bit (for now). And I can't even run the advertised AI software, because I use Radeon cards, not GeForce RTX or Volta V100. And no, I'm not going to buy Nvidia, because they don't open-source and they started GPU price inflation in 2012-2013.

[peekpok]

AI in its current form isn't particularly well suited to consumer products in my opinion, although there are examples here and there of it slowly creeping into daily life. Being able to image search through your own google photos is pretty neat. I do feel that Google has gotten a bit better at question answering especially when it comes to basic questions, and voice recognition has seemed to get better as well.

But where AI is really going to be useful in the next 5-10 years is industry mostly. Things like weather prediction, agriculture, drug discovery, chip design and other stuff like that. In some cases AI can also help manage large industrial systems such as when Google used Deepmind's AI to control the cooling system for one of its data centers and reduced the energy bill by 40%. AI is going to quietly work behind the scenes making many businesses more efficient. It's hard to imagine it now, especially with the current bout of inflation that the world is suffering through, but probably over the next 10 years a variety of products and services will become inexplicably cheaper.

I briefly mentioned it before but it's worth repeating that AI is being used in computer chip design now, which means we're already starting down that path of "recursive self improvement" even if its not 100% automated. I suspect that AI will take a larger and larger role in the design and operation of new computer chips as its easy to imagine a lot of ways to improve them such as having hardware neural-network schedulers in complex CPUs and other similar types of optimizations. Also the concept of dedicated AI hardware is relatively new as most companies aside from Google have just been doing all their work with GPUs, but Tenstorrent is an up-and-coming company that is working to design small and energy efficient inference chips. This will make it much more attractive to use AI in a lot of every day machinery, mainly for the purposes of increasing efficiency. Imagine an electric car that uses a neural network to predict the driver's habits and tune the performance of the motors to save energy and extend range, for example.

[Yuli Ban]

[andmar74]

I think that's too early for the Turing test. The AI can't fool me yet. I would just ask it some tricky science questions and the answers will be gibberish or half-gibberish.

But wouldn't you get the same result if you ask a layperson the same questions?

True, but there are many human mumbling nutsacks. I would prefer a human astrophysicist. He/She would crush any AI of today.

[Tadasuke]

Kurzweil wrote in his book, that in 2009, a standard PC would have a 4 teraflops AI processor and 300 000 MIPS CPU (which is about equivalent to AMD Ryzen 1800X or current generation consoles CPU). And he estimated that 4 teraflops AI processor would do 25% of work. In 2019 he envisioned standard PC having 600 TIPS CPU and 8 petaflops AI processor that would do 90% of work. So we would have specialized processors for constantly evolving deep learning AIs that could help us with lots of stuff in most computers. And the share of work done by AI would grow with every year, exponentially. That's how I thought it would go, but reality is different. AI is on servers, connected with us, but it doesn't help much.

[Ozzie guy]

With Proto AGI being almost certain within the next 2 years and 8 months I have two questions on my mind.

Like AI will Proto AGI improve exponentially?

and more importantly

As Proto AGI improves exponentially and has generality will it improve until it is human/super human level AGI or do we need to make a new paradigm shift?
e.g. Narrow AI could not improve exponentially into proto AGI but maybe proto AGI can improve into human level AGI.


I ask this because if we don't need to create a new paradigm shift and proto AGI will improve exponentially into human level AGI we will have already "won" before 2025. We will get to sit back and watch proto AGI exponentially improve into human AGI over the course of some years.

[peekpok]

The funny thing about Kurzweil is that he overestimated the performance increases due to hardware (failing to anticipate Moore's law slowdown), but seems to have underestimated how far we could get with good software. It would be funny if we invent AGI sooner than he expected on worse hardware than we were supposed to have by then.

[Yuli Ban]

The funny thing about Kurzweil is that he overestimated the performance increases due to hardware (failing to anticipate Moore's law slowdown)

Actually, this is a common misconception. Kurzweil has always prominently declared that Moore's Law was going to slow down and crash to a halt, even calling that it would take place in the 2020s.
It's more his critics and fanatics who claim that he said that Moore's Law would continue unabated and that it alone would be responsible for the Singularity.
Actually, Kurzweil's point was that 3D computing architectures would begin taking over, and he may be right. But he did indeed fail to predict software acceleration.