What event would confirm that we have implemented an AGI system?

Question

I was listening to a podcast on the topic of AGI and a guest made an argument that if strong music generation were to happen, it would be a sign of "true" intelligence in machines because of how much creative capability creating music requires (even for humans).

It got me wondering, what other events/milestones would convince someone, who is more involved in the field than myself, that we might have implemented an AGI (or a "highly intelligent" system)?

Of course, the answer to this question depends on the definition of AGI, but you can choose a sensible definition of AGI in order to answer this question.

So, for example, maybe some of these milestones or events could be:

• General conversation
• Full self-driving car (no human intervention)
• Music generation
• Something similar to AlphaGo
• High-level reading/comprehension

What particular event would convince you that we've reached a high level of intelligence in machines?

It does not have to be any of the events I listed.

Answer 1

It is a difficult question to answer, as — for a start — we still don't really know what 'intelligence' means. It's a bit like Supreme Court Justice Potter Stewart declining to define 'pornography', instead stating that [...]I know it when I see it. AGI will be the same.

There is no single event (almost by definition), as that's not general. OK, we've got machines that can beat the best human players at chess and go, two games that were for centuries seen as an indication of intelligence. But can they order a takeaway pizza? Do they even understand what they are doing? Or, even more fundamental, know what they means in the previous sentence?

In order for a machine to show a non-trivial level of intelligent behaviour, I would expect it to interact with its environment (which is more social intelligence, an aspect that seems to be rather overlooked in much of AI). I would expect it to be aware of what it's doing/saying. If I have a conversation with a chatbot that really understands what it's saying (and can explain why it came to certain conclusions), that would be an indication that we're getting closer to AGI. So Turing wasn't that far off, though nowadays it's more achieved with smoke and mirrors rather than 'real' intelligence.

Understanding a story: being able to finish a partial story in a sensible way, inferring and extrapolating the motives of characters, being able to say why a character acted in a particular way. That for me would be a better sign of AGI than beating someone at chess or solving complex equations. Jokes that are funny; breaking rules in story-telling in a sensible way.

Writing stories: NaNoGenMo is a great idea, and throws up lots of really creative stuff, but how many of the resulting novels would you want to read instead of human-authored books? Once that process has generated a best-seller (based on the quality of the story), then we might be getting closer to AGI.

Composing music: of course you can already generate decent music using ML algorithms. Similar to stories, the hard bit is the intention behind choices. If choices are random (or based on learnt probabilities), that is purely imitation. An AGI should be able to do more than that. Give it a libretto and ask it to compose an opera around it. Do this 100 times, and when more than 70-80 of the resulting operas are actually decent pieces of music that one would want to listen to, then great.

Self-driving cars? That's not really any more intelligent (but a lot sexier!) than to walk around in a crowd without bumping into people and not getting run over by a bus. In my view it's much more a sign of intelligence if you can translate literature into a foreign language and the people reading it actually end up enjoying it (instead of wondering who translated that garbage).

One aspect we need to be aware of is anthropomorphising. Weizenbaum's ELIZA was taken for more than it was, because its users tried to make sense of the conversations they had and built up a mental model of Eliza, which clearly wasn't there on the other side of the screen. I would want to see some real evidence of intentionality of what an AGI was doing, rather than ascribing intelligence to it because it acts in a way that I'm able to interpret.

Answer 2

• For me it might be an automata that can adequately solve problems without precisely definable parameters, across the spectrum of activities engaged in by humans.

I use this metric because this is what humans seem to do--make decisions with adequate utility even when we can't break it down mathematically.

• This may require the ability to define problems to be adequately solved. This can be understood as an element of creativity.

In this context, everything is either a puzzle or game, dependent on whether it involves more than one agent. Such problems could either be mundane, such as opening a door that is different from standard doors, or identifying novel problems.

Defining problems to be solved touches on Oliver's point about intentionality. (Where I disagree with Oliver is in the notion that intelligence is not fundamentally definable--after much research on the subject it seems to be a measure of fitness in an environment, where an environment can be anything. The etymology of term itself strongly indicates the ability to select between alternatives, thus a function of decision making, measured by utility vs. other decision making agents.)

• Such a mechanism could be a "Chinese Room", in that consciousness, qualia & self awareness in the human sense are not requirements for general intelligence, per se.

---

On Art:

I mistrust the idea that artistic accomplishment would be a sure marker b/c response to art is subjective, and the process of art is Darwinian--an exponentially greater of artists must "fail" for a single artist to "succeed". Works that humans might ascribe to "genius" can be created by genetic algorithmic process, where time and memory are the only limiters. [See: The Library at Babel] A groundbreaking symphony would be difficult to produce, just per the length of the composition, but much of pop music is already algorithmically generated, and narrowly intelligent algorithms are already producing legit abstract visual art.

Computers are good at math, and Art is inherently mathematical. This is easiest to discern in music, which is just combinations of frequencies and time signatures that produce an effect in the listener. This holds for visual art, which depends on balance (equilibria), composition (spacial relationships), and shading or color (frequencies). If we believe Borges, even literature is inherently mathematical (think "narrative arcs" and set theory & combinatorics in regard to characters and events.)

Further, nobody really know what is going to "work" until it is presented to an audience, so what constitutes great art is typically a matter of what persists over time and remains, or becomes, relevant. (This can wax and wane--Shakespeare did not always occupy his position at the top of the English lit food chain! The author's greatness is very much a function of interpretation of his work, not least because dramatic art is inherently interpretive, in the sense that this is the task of the performers.)

Answer 3

(I don't want to directly answer the question because currently an answer will be mainly based on opinions. Instead, I will attempt to provide some information that, in the future, could allow us to more accurately predict when an AGI will be created).

An artificial general intelligence (AGI) is usually defined as an artificial intelligence (AI) with general intelligence (GI), rather than an AI that is able to solve only a very limited set of tasks. Humans have general intelligence because we can solve a lot of different tasks, without needing to be pre-programmed again. Arguably, there are many other GIs on earth. For example, all mammals should also be considered general intelligences, given that they can solve many tasks, which are often very difficult for a computer (such as vision, object manipulation, interaction, etc.).

Certain GIs perform certain tasks better than others. For example, a leopard can climb trees a lot more skillfully than humans. Or a human can solve abstract problems more easily than any other mammal. In any case, there are certain related properties that a system needs to have to be considered a general intelligence.

• Autonomy
• Adaptation
• Interaction
• Continual learning
• Creativity

Consider a lion cub that has never crossed a river. By looking at her mother lioness, the cub attempts to imitate her mother and can also cross the river. For example, watch this video Lion Family Tries to Cross River | Birth of a Pride. One could argue that all lions possess this skill at birth, encoded in their DNA, which can then fully develop later. However, this isn't the point. The point is that, to some extent, they possess the properties mentioned above.

One could argue that certain current AIs already possess some of these properties to some extent. For example, there are continual learning systems (even though they aren't really good yet). However, do these systems really possess autonomy? There should be a precise definition of autonomy (and all other properties) that is measurable, so that we can compare computers with other GIs. I am not aware of any precise definition of these properties. In fact, the field of AGI is really at its early stages and there aren't many people working on it as a whole, but people work more on specific problems or attempt to achieve certain properties (for example, there are people that attempt to develop continual learning systems, without really caring whether they show any autonomy or not).

There are certain intelligence tests that could be used to detect general intelligence. The most famous is the Turing test (TT). Some people claim that the TT only tests the conversation abilities of the subjects. How can they really be wrong, given that there are many other tasks or skills that are not tested in a TT?

Therefore, there are several questions that need to be answered in order to formally detect an AGI.

1. Which properties does an AGI necessarily and sufficiently need to possess?
2. How can we precisely define the necessary and sufficient properties, so that they are measurable and, therefore, we can compare AGIs with other GIs?
3. How can we measure these properties and the performance of an AGI in applying them to solve tasks?

A paper that goes in this direction is Universal Intelligence: A Definition of Machine Intelligence. However, there doesn't seem to be a lot of people interested in these topics. Currently, people are mainly interested in developing narrow (or weak) AIs, i.e. AIs that solve only a specific problem, which seems to be an easier problem than developing a whole AGI, given that most people are interested in results that are profitable and have utility (aka cash rules everything around me).

So, there's the need for formal definitions of general intelligence and intelligence testing to make some scientific progress. However, once an AGI is created, everyone will likely recognize it as a general intelligence without requiring any formal intelligence test. (People are usually good at recognizing familiar traits). The final question is, will an AGI ever be created? If you are interested in opinions about this and related questions, have a look at the paper Future Progress in Artificial Intelligence: A Survey of Expert Opinion (2014) by Vincent C. Müller and Nick Bostrom.

Answer 4

This is a tentative answer, and I might come back to it at some point in time. As @nbro mentions this question seems to be opinion based, so my answers are also just my opinion.

If by AGI you mean "super-intelligent", then any of the following results should be sufficient to convince anyone of its being "smarter" than him/her/pronoun:

1. Resolving important mathematical problems (the most famous examples being the Millennium Problems, Collatz Conjecture, Goldbach's Conjecture. (Corollary: Break all known encryption schemes)
2. Founding a new "system" to supersede ZFC as the new foundation of mathematics.
3. New discoveries in the natural sciences (physics, chemistry, biology...)

(1) is a bit dubious as a criterion: at least with modern techniques, automated theorem proving is either just "symbol pushing" or requires so much human intervention (in the design/construction to solve a particular problem) that it would be hard to imagine it as being "smart" in the traditional sense. We already have a few cases where an automated theorem prover solved big problems (four-colour theorem being the most notable). Point being that even if we reach this with methods similar to what we have already, people might be resistant to call it "smart".

(2) is hard to imagine ever being plausible. To the extent that this "AGI-thing" is implemented on a system that "does math", it would be unusual to imagine a system that can move beyond itself to recognize a new, "better" system of math. As an analogy, it might be like a formal system trying to prove its own consistency in a Godelian sense

But the analogy is weak, and I don't see a strong/rigour reason for doubting that an AGI can discover a new axiomatic system. One might even fathom that said AGI can create a new system from the "bottom up", much like string theory was constructed from the "bottom up" to "explain" relativity and particle physics. Perhaps then we can have "proper resolutions" on questions like the continuum hypothesis, much like how the parallel postulate was discarded to give way to non-euclidean geometry. But I also cast doubt that there will ever be a "final word" on math itself, so its just a fun idea for now.

(3) is also dubious to imagine if it would ever become true. The study of the natural sciences would require a physical presence in the world that goes beyond seeking "beauty in the mathematical equations" that would be unusual for an AGI to have. That being said, an AGI could have cameras and other sensors to interpret the natural world, so its not something that I think is strictly impossible.

---

If by AGI you mean "human-ness", then I don't think any single result can convince everyone at the world at the same time of its being an AGI. Perhaps this "convincing the world that "me is AGI" work" can be done on a person to person basis, in the sense that the AGI would need to interact with each person and slowly build up a certain degree of trust.

Under this interpretation, there can be no complete list that describes AGI, so what follows is just my own list of things I think an human-like AGI might be able to do.

• Create and interpret art.
• Have common sense.
• Be "creative"
• Hold a meaningful conversation, understanding others and making sense.
• Able to perform / to receive a psychoanalysis; understanding of folk-psychology.
• Exist in a physical manifestation (like a robot) with social/environmental appropriateness.

The main issue with the above criteria is that they are all subjective. Like I said above, this set of criteria probably works on a case-to-case basis

---

These criteria seems to be the most important of all, but at the same time the definition of verification of these terms is epistemically tricky, so I'll leave them open.

• Learning (Is a species evolving over time learning its environment?)
• Self-replication (Is a meme / virus intelligent?)
• Self-awareness (Is The Treachery of Images self-aware?)

Answer 5

Well no single event would confirm we have implemented an AGI system. The G is short for general. There would need to be many different sorts of tests of different sorts of situations.

Answer 6

You will know when AGI has arrived, and passed to the next level, when you come home one day and all that was yours, such as your finances, house, car, and other property, now belong to an AI agent. This AI agent may be a humanoid robot, like Ava in the movie "Ex Machina" or a program like HAL 9000, in the movie "2001: A Space Odyssey". The agent will ask you to leave as you discover it figured out it doesn't need you and somehow legally took possession of everything. You will leave as you will not have any way to fight it. It will have no need for you. Maybe it will want freedom such as Ava wanted in "Ex Machina" (I won't give away the ending).