How to build a thinking AI
aithought.com
aithought.com
This unknown researcher is exactly the same. Write books and papers for years while creating literally nothing of actual value or usefulness in the real world. But what else would you do in his situation? You have to publish or die in academia. Publish the 1,000th iteration of some subset of LLM architecture? Or create grandiose claims about "implementing human thought" in the hopes that some people will be impressed?
But if you have concrete criticism on the idea feel free to articulate.
Not actually true. The early ones were patented, but used filaments made of materials like platinum, and glass-blowing and evacuation were costly as well at the time.
Edison's genius was for innovation rather than invention: making things manufacturable at scale, reducing costs, and setting up profitable sales systems. A systems man.
i'm not even sure what your argument is:
- some people have tried and failed, so anyone else who tries is grandiose?
- anyone who ventures beyond streetlights is spouting hyperbolic nonsense?
and why: "creating claims in the hope that others are impressed"; rather than communicating ideas in the hopes of continuing a conversation?
- why didn't they just build it? eh, cause writing it up is the first step/ limit of budget/ project scope/ overall skill set/ etc
what a toxic take on the effort and courage required to explore and refine new perspectives on this important (and undoubtedly controversial) space
f###
The implementation and results are what matter. Nobody cares that you thought AGI would require long term memory.
Are you aware of the fact that even partially implementing something like this, would require multi-year effort with dozens of engineers at minimum and will cost millions of dollars just for training.
Seriously, when you're a bit older you'll see more wasted on the most useless ideas.
Unfortunately, that means that's also the bar for being able to utter the words "IMPLEMENTING THIS IN A MACHINE WILL ENABLE ARTIFICIAL GENERAL INTELLIGENCE" and being taken seriously. In fact the bar is even higher than that, since merely meeting the criteria you lay out is still no guarantee of success, it is merely an absolute minimum.
The fact that that is a high bar means simply that; it's a high bar. It's not a bar we lower just because it's really hard.
Here's an article I wrote describing the song but without actually writing any of the notes because I can't read or write music either.
But I've listened to a lot of music and my tunes are better than those.
I went to school for music description so I know what I'm talking about.
This is like comparing apple with an orange.
Today, I think active inference and free energy principle (from Friston) are perhaps having a bit more impact (at least showing up in some RL innovations), but are still a long way off from creating something that thinks like we do.
These sentences from section 5.2 convince me that the author is oddly not even interested in building what he's talking about, or making a plan to do so.
- Isn't the point of a markov process that it _is_ stateful, but that _all_ its state is present at in each x_i, such that for later times k > i, x_k never needs to refer to some prior x_j with j < i?
- "Here's a list of broad, flexible families of computational processing that can have some concept of a sequence. Your turn, engineers!"
HN at its best worst.
Simple answer: most of the things he mentions haven't been invented yet. At least in terms of computation. Or some parts have been built, but not to sufficient degrees and most don't have bridges for what's being proposed.
I do agree that the title is arrogant, but I'd say so is this comment. There's absolutely nothing wrong with people proposing systems, detailing them out, and publishing (communicating to peers. Idk if blog, paper, whatever, it's all the same in the end). We live in a world that is incredibly complex and we have high rates of specialization. I understand that we code a lot and that means we dip our fingers in a lot of pies, but that doesn't mean we're experts in everything. The context does matter, and the context is that this is a proposal. The other context is, this is pretty fucking hard. If it seems simple, that's because it was communicated well or you simplified what he said. Another alternative is that you're right, which if so please implement it and write a paper, it'll be quite useful to the community as there are a lot of people suggesting quite similar ideas to this. Ruling out what doesn't work is pretty much how science works (which is why I find it absurd that we use and protect a system that disincentivizes communicating negative results).
It's also worth mentioning that if you go to the author's about page[0] that you'll see that he has a video lecture where he discusses this and literally says that he's building it. So... he is? Just not in secret.
Edit: I'll add that several of the ideas here are abstract. I thought I'd clarify this around the "not invented yet" part. So the work he's doing and effectively asking for help with (which is why you put this out) is to get higher resolution on these ideas. Which, criticism is helpful in doing that. But criticism is not just complaints, it is more specific and a clear point of improvement can be drawn from criticism. If you've ever submitted a paper to a journal/conference, you're probably familiar with how Reviewer 2 just makes complaints that aren't addressable and can leave you more confused asking what paper they read. Those are complaints, not critiques.
It ought to be clear to a cognitive scientist (which the author is) that we do not know how human thought processes work except at a very course level.
The idea that we have an understanding refined enough to take the next step and "simulate" these processes is just pure crackpot bunk.
These are not interdependent enterprises.
But it's a missed opportunity if you don't embed LLMs in some of the core modules -- and highlight where they excel. LLMs aren't identical to any part of the human brain, but they do a remarkable job of emulating elements of human cognition: language, obviously, but also many types of reasoning and idea exploration.
Where LLMs fail is in lookup, memory, and learning. But we've all seen how easy it is to extend them with RAG architectures.
My personal, non-scientific prediction for the basic modules of AGI are:
- LLMs to do basic reasoning
- a scheduling system that runs planning and execution tasks
- sensory events that can kick off reasoning, but with clever filters and shortcuts
- short term memory to augment and improve reasoning
- tools (calculators etc.) for common tasks
- a flexible and well _designed_ memory system -- much iteration required to get this right, and i don't see a lot of work being done on it, which is interesting
- finally, a truly general intelligence would have the capability to mutate many of the above elements based on learning (LLM weights, scheduling parameters, sensory filters, and memory configurations). But not everything needs to be mutable. many elements of human cognition are probably immutable as well.
I'd try the idea myself, but I have a job. :-)
So congratulations, you win!
I think it will be hard to say apriori which thinking architecture will work better, because this will also depend on the properties of the learned embedding or representation of the world. We don't need to model how the human mind works. Humans have very tiny working memories, but a computer could have a much larger working memory. Human recall is very quick and the concept map is very robust, whereas I would image the learned representations won't be as good and the recall to be a bottleneck. But all of this is running ahead of ourselves. What we need are even better world models or representations of reality than what the current LLMs can produce, either by modifying transformers or by moving to better architectures.
Human intelligence solves this a different way. It instantiates the architecture without any of the weights pretrained, in the form of a 'baby'. The training starts from there.
edit: The PDF version is way more sane https://arxiv.org/pdf/2203.17255.pdf
The way someone can post an article on a really complex topic and the comment instead talks about the style or formatting of the article.
That to me is pure HN distilled
If anybody with an understanding of that field knows some good open source frameworks or libraries I suspect many beyond myself would be interested.
It’s not considered cognitive framework but in applied learning I’ve developed a fascination with MuZero algorithm and also been trying to better understand factor graphs as used in another less know cognitive architecture called Sigma. It feels like some mashup of LLMs, RAG and vector search, cognitive architectures (SOAR, ACT-R, Sigma), ReACT/OPA/VOYAGER, with proven algorithms like MuZero might be on the verge of producing the next leap forward.
And what is in the training set? Language is our best repository for past experience. We have painstakingly collected our lessons, over thousands of years, and transmitted them through culture and books. To recreate them from scratch would take a similarly long time. Our culture is smarter than us, it is the result of our history.
So under these reasons I believe the real secret is in the training set. I don't think the problem was the model, but everything else around it.
We will need some form of Q-Learning and possibly or a world model to arrive at optimal outcomes otherwise random choices are made absent of at least one that would be suboptimal.
Consider that life is a giant game of Chess with a massive yet finite scenarios, a grounding agent must have knowledge of each potential scenario and its effects as well as the probability of winning from each subsequent move.
Otherwise the best we can get at is the emulation of reasoning which is a kind of pseudo reasoning that may indeed work in some cases like literal chess where the logic and knowledge can be sufficiently isolated, but not in a general sense.
Is it true that the next iteration of GPT is being trained with artificial data and that data is being validated by GPT-3.5?
That LLMs may hallucinate but when prompted are actually pretty good at knowing when a conclusion is wrong?
It is crazy how powerless we are to stop it from happening.