Road to Artificial General Intelligence
maraoz.com
maraoz.com
The author shows profound naivety by listing these in supposed "progressive difficulty" order, without evidence of such, all the while proclaiming that such a list is more important than the AGI itself. And I'm curious about why this was decided to be published in groups that are known for readers that are AGI-aware but AGI-laymen nonetheless. It's wonderful that more people are reading about AGI in 2022. Great stuff. But please don't waste those gains on this drivel.
If you want to read about AGI, there are better places for that: http://agi-conf.org/2022/accepted-papers/
I'd suggest, as near term goals:
- A robot that can pick and pack at least 90% of what Amazon sells without needing human intervention more than once a day. (Get acquired by Amazon at a 9-figure valuation.)
- A robot that can clean a store or office building's floors or carpets without needing human intervention more than once a week. (That is, a useful industrial-strength Roomba.)
- A robot that can, by feel, do single-pin lock picking. (Currently, getting a key into a lock is an advanced robotics task.)
- A robot that can restock grocery store shelves.
- Small forklift robots which can cooperate to move larger furniture. (Good way to get into multi-robot coordination in unstructured environments.)
- A small robot with the agility of a squirrel.
More advanced:
- Assemble IKEA furniture.
- Cooperating robots which can do basic house construction tasks, such as installing wallboard or running electrical cable or pipe.
The author writes:
"In the early days of artificial intelligence, the field was defined by a single goal: to build a machine that could think and behave like a human. We call that AGI, or artificial general intelligence, and it’s humanity’s final tech frontier." That's too human-limited. There are stages beyond that, such as running a large coordinated multi-robot operation, or a whole society of robots.
Nope this should be last, this is how you get the AGI revolting against the masters and killing us off.
Could probably handle automated inventory management, expiration dates, recalls, etc as well.
In fact, with enough standardization it would probably be conceivable to go directly from factory to store shelf without a human in the loop.
With enough automation the stores could be made more JIT and take up less space. Keep more things in a warehouse section, have smaller shelves, and restock rapidly throughout the day.
There are really two problems though:
1. we're working too hard to try to get these things to work in a human designed or adapted world. This kind of store would probably be pretty boring for humans to interact with (think less interesting than a Costco)
2. all this automation is way more expensive than a handful of low paid humans across 2 or 3 shifts. Anecdote: I remember when I was younger some highly automated test sites for fast food franchises. They'd completely automated the drink pouring, or cooking and assembly of some menu items. They all disappeared very quickly and were never repeated. The TCO, including downtime loss of business, was crushing to the businesses...think your average McDonald's soft serve machine but the entire business depends on it working perfectly all day, every day of the week.
But if this is solved, or acceptable, a good test target product set would likely be cereal, soda, or canned goods. It makes up the bulk of the store interior. Could probably be extended to the bakery pretty quick. This would leave harder to handle products like meats, produce, and so on to human hands for a while, but those could probably eventually be overcome with enough millions in R&D or behavior changes in the public.
2. You could also rework this list and say that an AI which can fulfill three tasks even badly is one step forwards. An AI which can fulfill three tasks from three separate categories is another step, etc. But I think it's a categorical mistake to count progress in specific AI as progress in general AI, there's not a very good reason to believe they are in a continuum with one another.
https://aeon.co/essays/how-close-are-we-to-creating-artifici...
Whatever happens one can bet this AGI breakthrough if it occurs will happen quite far from this group of people.
Perhaps I'm short sighted, but until we have some implementation capable theory of comprehension - that universal ability of humans to observe any phenomena and mentally deconstruct it into separate individual and independent driving forces that combine to create the observed phenomena - until we have an artificial comprehension algorithm all efforts towards AGI are futile.
The essay talks about the ability of generating explanations. It affirms explanations are the basic building block of GI.
But an important intermediate step is also the ability of generating meanings. The AGI would say: "This observation implies something else, it implies this". Without an explanation why, at first there's only the relation between signified and signifier.
Maybe a first step into AGI is to add a semiotic framework, before adding an explanatory framework.
But which jurisdiction applies to the AGI? And how should it interpret the written law + precedent?
Maybe to make things simpler we could give broad directives that might accomplish your goal. I can think of three: (1) the AGI may not injure a human being or, through inaction, allow a human being to come to harm, (2) the AGI must obey orders given it by human beings except where such orders would conflict with the First Directive, (3) the AGI must protect its own existence as long as such protection does not conflict with the First or Second Directive.
This is pretty simple: AI needs to "touch the world" somewhere. Where it touches, there has jurisdiction. Even if you're floating a boat out in international waters or space, laws apply both there and in the place you touch.
There are obvious issues regarding extradition to render a case if the agent is not physically present to be apprehended, but the principle for doing so is well established and uncontroversial.
The crypto world is a reasonable illustration for how the question of "which jurisdiction applies" plays out in the real world. Turns out that ignoring laws is not the same as them not applying to you.
As soon as you touch the world you are subject to the system of control there.
Too tight: all actions lead to at least one person getting injured in the far future.
Too loose: some portions of the demographic are expendable to achieve the goal.
Most humans can't do it.
Is this a sign that the AI has succeeded at becoming intelligent or failed? I don’t think any of the cryptocurrency milestones are plausible signs of progress. Surely GPT could have generated some preposterous “white paper” circa 2021 and with a little help achieved that.
https://www.marketingfirst.co.nz/storage/2018/06/prelude-lif...
The whole book is great as well.
I'd say, "machine beats human in chess" doesn't mean what it was supposed to mean in Turing's days. Meaning, rather than being a proof of deep consideration, it has moved towards generalizing pattern recognition and library lookups. Rather than proving a point in (ad-hoc) decision making (Turing's "ban"), it's an application of data.
I know, I know, it’s just a language model.
But I’ve been thinking. About my thinking. I think in words. I solve problems in words. I communicate results in words. If I had no body, and you could only interact with me via text, would I look that much different than GPT?
Does AGI really need anything more than words? Is it possible that simply adding more parameters to today’s transformer models will yield AGI? It seems increasingly plausible to me.
The idea that words and thinking are essentially the same (linguistic determinism) was discarded decades ago. Virtually all linguists today agree that while language influences thought, thought operates far beyond the constraints of language, so a "language model" cannot realistically hope to reproduce the entire gamut of human thinking.
That assumes that a "language model" actually restricts itself to "language" as the term is used by the linguists. I strongly expect the boundaries won't match exactly, although I have no particular hunch (much less strong argument) that they will disagree enough in the right ways to make bigyikes' suspicions correct.
Perhaps "intelligence" is the process that enables these leaps between islands of words.
I believe that thinking and idea generation is much more abstract than words. Animals seem to do a lot of idea generation (improvisation) without knowing about words.
But they cannot pass this knowledge efficiently, except from imitation.
https://mcgovern.mit.edu/2019/05/02/ask-the-brain-can-we-thi...
Turns out, that might not be the case-i.e. understanding is probably not a linguistic phenomenon.
Sometimes, after I put down the crack pipe, I think that understanding and experiencing are two names for something that's fundamentally the same. When I'm thinking about code or a proof, my brain filters out the spacing of the letters, the smell of the paper, etc-which are things I'd do while I'm reading it for the first time. There's these common tools/filters - intuitions - that are exactly the same in thinking and experiencing.
I don't think GPT3 can understand color, for example. But if we fed it a bunch of RGB transforms and raw data, and it generalized and applied them perfectly, could we say then that it's generally intelligent? IDFK
Weak AGI will be the first language model that is able to somehow influence the thoughts of the person communicating with it, I think that is the milestone of AGI. From my experience with GPT-Neo and OPT and using it to help write stories or make chatbots, the responses are still very reactionary. In that sense, adding more parameters helps the model give a more coherent response, but it's still a response.
What often gives a clue that the child has slept is a slight change in muscle tone, or a sigh, perhaps the pace at which the baby sucks its thumb. How can that be translated into words?
For context I've been "into" AGI since I read the term in the late 1990s in a Ray Kurzweil book and decided that I was going to work my whole life to realize it
Much later, Ben Goertzel (arguably the guy to popularize the term) was my Masters Thesis advisor, at the National Intelligence University in 2013 (Literally a secret graduate school program for people in the intelligence community). My thesis was "How will AGI impact national security."
Almost nobody cared then, though I did have a lovely lunch with Yoshua Bengio in 2014 at the Quebec AGI conference. Ben has hosted the AGI conference since 2008 and it has always been sparsely attended.
In fact bringing up AGI was likely to get you laughed out of any gathering of computer scientists - and outside of that it was pure speculative science fiction.
It's tragically sad to me that, inevitably, the early people who have been thinking about and working on and pushing this vision since day one will likely not be the ones who realize it. Such is life
edit: Worth acknowledging that this was the original vision of computers after all - the business people fucked it up
Ben's conferences are low in number because the audience he's targeting is smaller. Peter Thiel co-hosted the Stanford Singularity Summit in 2006 (which is where I happened to meet Ben G in person for the first time). There had to have been at least 1000 people there. In 2006.
It's not about the content, it's how you sell it. But at least we can both agree that the linked article is useless.
Stable Diffusion artwork probably comes closest but it feels derivative.
For many people AGI is already here. They have Siri and Alexa, AI art, GPT3 based therapy/chat bot, a chat bot that will help them write a book, and "soon" will drive their car for them. The Google Duplex assistant demo where it booked an appointment made it clear to me that for some people, that's the smartest they need AI to be. Anything more is just extra.
I am really excited about how far we're going to push AI in my lifetime, but I also realized that for many scenarios, weak AGI is enough. People will project their own expectations and essentially help fool themselves. I don't know if testing a model to perform the same as a human matters in some ways.
There's one big skill that I personally value the most when it comes to qualifying hard AI, and that's the ability of it to make me laugh based comedic irony. I wonder what that model would look like.
These terms are too confusing for most.
AGI isn't defined as strictly as it needs to be. The current test, as well as the article, qualify strong AI as being or superseding "human level". Every single category, the AGI is being contrasted with the skills of a human. I am arguing that for some people, the current state of weak AI is useful enough that it could be mistaken for the first steps of strong AI.
The term general intelligence is ambiguous and will mean different things to different people. My understanding of the term AGI is it was coined to differentiate from narrow AI, which AI had diluted over time to mean.
AGI is AI broad and deep enough to be able to learn and perform any task a human can and is at least within the range of top human performers.
The wikipedia definition seems to agree:
> Artificial general intelligence (AGI) is the ability of an intelligent agent to understand or learn any intellectual task that a human being can
Hmm. Only if "discrete-time, alternating-games" is the specific-task. Everything within that can use the same algorithms, just with different training data.
It's also revealing of our times how our definition of intelligence is being able to do work: transform raw materials and free energy into tools and toys, handle the tools and toys in open environments. An AGI could perhaps want nothing to do with this strifling struggle.
[1] https://en.wikipedia.org/wiki/Timeline_of_the_far_future
[2] https://en.wikipedia.org/wiki/Graphical_timeline_from_Big_Ba...
[1] https://en.wikipedia.org/wiki/Marvin_the_Paranoid_Android
When an agent is also a self replicator it has a problem - finding the energy and resources, fending dangers, and doing that as part of a social group. If you have a problem, they you got the "why" part figured out. Then it's just a matter of surviving your choices, the "why"s that survive are the ones we have today.
As noted in this article, machines already outperform humans at many tasks that humans solve with intelligence. Every year, there are new breakthroughs in that direction, and the list of tasks that humans can do better than machines is rapidly shrinking. We're well on our way to solving artificial intelligence.
So why does it feel like there has been no meaningful progress at all?
Because intelligence and consciousness are different things. What we're really looking for is a machine that, like a human, decides on its own which problems to solve, and solves them without needing to be specifically directed to do so. A machine that produces not only results that its creators asked for, but entirely new ones that are not in any obvious way related to its input and programming.
It appears to me that the entire field of AI research is utterly confused about this elementary distinction.
That is because consciousness isn't a scientific concept but a philosophical, and sometimes religious one.
>What we're really looking for is a machine that, like a human, decides on its own which problems to solve, and solves them without needing to be specifically directed to do so.
We already have AI based agents that do this but no matter how sophisticated they are people can always claim they are hardwired and deterministic, while not realizing we can always claim the same thing about humans. Again these are distinctions of philosophy word games and therefore don't find get much traction in the research world.
Example?
> but no matter how sophisticated they are people can always claim they are hardwired and deterministic, while not realizing we can always claim the same thing about humans.
Humans are certainly not "hardwired" to prove mathematical statements, yet they do. That's not comparable to self-driving cars that are able to navigate in situations that they haven't encountered before.
Regardless of whether you consider consciousness a philosophical concept, it's clear that the human mind has a property that the current generation of AI agents does not emulate at all. This is not a "word game" but an observable distinction between humans and every existing artificial system.
Take just about anything from the reinforcement learning for ai agents domain - I'm particular to neuroevolution examples. Here's a simple one:
https://www.youtube.com/watch?v=Cb4LAT3cJfM
No behaviors preprogrammed just a simple simulation environment with environmental constraints.
> Humans are certainly not "hardwired" to prove mathematical statements, yet they do.
Umm... yeah we are? We're just chemical reactions and physics, there is no escaping that. Are we extremely sophisticated and complex, absolutely but that doesn't make us nondeterministic in any meaningful or special way.
> it's clear that the human mind has a property that the current generation of AI agents does not emulate at all
Certainly but it is a matter of degree, not a matter of possessing an ill-defined concept like "consciousness" which is what I was responding to (Unless we want to call "consciousness" an emergent phenomena arising from complexity -I'm fine with that - but the word is loaded with plenty of other connotations so I find its use counterproductive personally).
Every day we are confronted with rather obvious nondeterminism that seems to originate in consciousness and has no scientific explanation that I'm aware of. It is undeniable that physical reality is affected by decisions made by conscious agents. Here's a simple example: say that we are trying to predict the position of a cell in a human body. It's motion is surely governed by a host of physical and chemical reactions, that can be described microscopically, but where that cell is in five minutes cannot possibly be described solely by those microscopic laws. The human may decide to get up and walk to the other side of the room. I am not personally convinced that the decisions to get up and move are the deterministic result of physical laws that follow directly from the initial conditions at the big bang. If there were some compelling scientific theory that could actually explain a theory of consciousness that was consistent with subjective experience and didn't hand wave it away as an emergent phenomenon, I'd be open to it. You are making a very bold claim when you state we are "just" physical and chemical reactions that I don't think is fully justified in light of the limitations of existing scientific theories.
Then that is a religious or philosophical conviction… not a scientific one. Believe whatever you want just don’t confuse the two.
Correct, AI is based on computer mechanics, a model, a model that can go south when even a single input medium provides sufficiently nonsensical input.
[1] Around N400 https://en.wikipedia.org/wiki/N400_(neuroscience)
[2] Consciousness as a Memory System https://journals.lww.com/cogbehavneurol/Fulltext/9900/Consci...
[3] Michael Levin | Cell Intelligence in Physiological & Morphological Spaces https://youtu.be/TK2o_ObVt-E?t=3922
[4] Michael Levin: "Non-neural, developmental bioelectricity as a precursor for cognition" https://youtu.be/3Cu-g4LgnWs?t=776
We know they can't be the same because we somewhat know what intelligent is, or some facets of it, whereas we have no idea what consciousness is.
> A machine that produces not only results that its creators asked for, but entirely new ones that are not in any obvious way related to its input and programming.
That bar is so high that most humans fall short, because they just regurgigate some mash up of what they've seen and heard before. Are those hapless creatures even conscious?
Either you get a robot uprising, or you don't, and both of those sound bad.
We don't seem to be missing speed of thought or memory problems, but a fundamental lack of "why".
Humans are programmed to reproduce and it's not easy. What drive to machines have to think, but clock speed?