Hacker News Reader

Comment by atemerev | Hacker News Reader

This is what we used to call an "AGI" back in 2013-2014. Generalized artificial intelligence. An AI which is general enough to be used for diverse tasks in multiple fields, unlike specialized AIs of that time (e.g. for image recognition or self-driving cars). Nothing more than that.

Now the goalposts are moved way forward, and now "AGI" is something that have conscious, with superhuman intelligence, and probably evil intent. But back in a day, AGI was just what we have today -- an universal chat bot.

We live in the magic land.

View on HN

This doesn't match what I've seen of usage of the term AGI since ~2010. I think we've "solved" chatbots more than was imagined at the time, and that involves being able to produce readable and arguably useful answers about a range of topics, but we haven't got game playing sorted yet, we don't have proactive models, and there's nothing in these models for movement or robotics (which until recently has been considered an important necessary step).

We have taken a noticeable step forwards towards AGI in the last few years, but AGI is still a long way off.

It's not just about language.

https://general-pattern-machines.github.io/

>but we haven't got game playing sorted yet

4 can play chess, Minecraft fine. So I imagine a lot of games fall down to data.

>and there's nothing in these models for movement or robotics

The point is that there doesn't need to be.

There are many projects using LLMs to pilot robots off text prediction ability.

https://tidybot.cs.princeton.edu/ https://innermonologue.github.io/ https://www.deepmind.com/blog/rt-2-new-model-translates-visi...

I think most from the 2000s would be calling what we have today AGI.

Games are solved since AlphaZero. Just set up an adversarial model, and voila. Any game, any conditions.

Proactive models are simple — just generate the “internal monologue” periodically from the system prompt, mix it with the compressed state, then feed it to another model (the outputs are used as state updates). I think this is how human agency works as well.

Movement is a different story (LLMs do not work here, except as planners), but it pretty much solved too. Robot hardware is too expensive, though, and it is not clear how to power it autonomously.

AlphaZero is amazing, but it only plays 2D games with well defined win conditions.

Yes, like military strategy.

Indeed, this is something that neither AlphaZero nor anything else has managed to do at anything approaching human level.

You really shouldn't say that so confidently. I know for a fact that at least in the US military (well, DoD civilian analysts), they're using models like AlphaZero for theatre-win condition modelling. It's probably being used because it allows analysts to find better solutions and they probably started using it after winning wargames with it - both guesses on my part, all I know for sure is that they are using the tech to model westpac.

There is still no computer program (with or without “AI”) able to play bridge decently.

"there's nothing in these models for movement or robotics"

You should read the article. It talks about exactly this use case, using an LLM as part of robotics control systems.

But at that point neural networks started to dominate AI research already to the point that "AI" started to become synonymous with them... what about 1950(?)-2010 ?

> But back in a day, AGI was just what we have today -- an universal chat bot.

I don't know anybody that used AGI to mean that. I've always thought of AGI as an AI that can learn how to perform tasks with very simple instructions, rather than spending weeks training a neural network on a data center full of GPGPUs.

Basically, my idea of AGI has always been what you're calling the "new" definition of AGI.

Remember that people have decades of generalized training before becoming able to perform tasks with very simple instructions. I would consider years of pre-training to be acceptable if the model is usable and trainable quickly in practice as well

Mostly “training” to do abstract things, not really for survival or fun.

I could catch fish since a fairly early age, mostly because I enjoyed it.

At about age 12 I used to drive stolen cars around without any training at all except observation.

It too me decades to become a corporate robot though :)

I think training is a weird word, life is an experience, not just training. I think your comment demonstrated the limitations of language to describe things.

We are pre-trained by billions of years of evolution.

To use a fishing rod and drive a car ?

To have the social and physical capabilities to create and pass on the concepts of "fishing rod" and "car" as well as how to use them. We did start from scratch, after all. But I think the learning of simple, complex, abstract, and nuanced concepts throughout a person's development as their brain grows is more than enough to compare to an artificial general intelligence without taking into account development of a species, considering we're the ones building the chips the AI is running on.

Nothing starts from scratch...

This is true.

People think newborn babies know nothing. But even from birth, there are certain instincts and skills they already know, such as the ability to recognize faces.

Heck, our ability to think, analyze, etc. has come from hundreds of thousands of years of evolution. Even dogs, which we consider to be intelligent, often fail at simple tasks like trying to bring a stick through a doorway when the stick is wider than the doorway. Sure, sometimes they figure it out, but not all can.

It only takes like 5 minutes to teach someone to play Fluxx, not decades.

Aren't you effectively describing few-shot learning in LLMs? Teach the model a new task by describing it and maybe give a few examples?

I would call a chat bot a specialized application. And I dispute that a chat bot was ever considered the end goal of AGI. The idea is that if we had AGI, we could build a chat bot with it to test/demonstrate its general reasoning capabilities. Simply being effective at chatting is not sufficient to be called an AGI. And today's chatbots are quite stupid.

Sure we're starting to give chat bots the ability to query information and even interpret user-provided images, but AFAIK the image interpretation is done using a different model rather than being an extension of the LLM.

LLMs aren't chatbots. That's just a nice consequence of what they can do.

Their prediction abilities stretch beyond just language even if they're only trained on it.

https://general-pattern-machines.github.io/

There's nothing specialized about being able to predict meaningful completions from arbitrary token sequences.

There's a whole lot that can be done with that even if they were constrained to language.

https://tidybot.cs.princeton.edu/ https://innermonologue.github.io/

There's nothing stupid about GPT-4. There's no testable definition of general intelligence GPT-4 would fail that a chunk of humans also wouldn't.

Vision understanding does not have to be a separate model. https://arxiv.org/abs/2306.14824

Reply on news.ycombinator.com