Fair enough. Can you explain the skepticism?

One approach to estimate the time when computer brains become comparable to human ones is to compare computational power. However, a human brain does a lot more than fuelling the mind, it controls and steers the body, and keeps its own house. The brain has 100 000 000 000 neurons, each of which has 10 000 neighbors and emits signals 100 times a second. That's 10^17 Flops, or 100 Petaflops.

How far are we off from that? The supercomputer HPE Cray EX235a has already reached about ten times that. 1194 Petaflops, according to the TOP500 list. So folks, behold, repent, the Singularity is nigh.

Regards,

Joachim

@Rafael Dean Brown:

I think it would be good to provide an exact and all-agreed definition for what a natural general intelligence is, which is only vaguely circinscribed and caregorized by the known intelligence theories (Gardner's, Albert's, Garrett's, Albus's, Fagan's, etc.) and which equally and simultaneously concerns human individuals, teams/groups of humans, collectives/societies of humans, as well as the entire mankind, as a whole. In other words, what are we talking about?

Anyway, thanks a lot for this interesting question!

Best regards,

I.H.

Imre Horvath That is exactly the problem underlying this debate. We do not have an agreed definition for natural general intelligence (IQ certainly is not the only measure), and we also lack an agreed definition for artificial intelligence.

The definition for AGI being put forward by a few experts in the field (roughly summarized by me) is that AGI is an AI that can perform all human tasks better than all humans. This is a moving goal as the definition by Turing was certainly a far lower bar:

Joachim Pimiskern Just to be clear, is it your view that AGI is already here but not announced, or that it imminent. According to your calculation, is AGI then a matter of compute as some in OpenAI have suggested?

Rafael Dean Brown

I fully agree with you: IQ tests are limited and poor indicators of intelligence as a complex phenomenon and the used IQ indices are unable to measure intrinsic variable aspects of intelligence like emotional intelligence and social intelligence. An IQ test is reflexive, it only measures how well an individual can take an IQ test and anything else is just prediction.

The second part of your response to me also warrants attention: You wrote: "AGI is an AI that can perform all human tasks better than all humans". OK, so this is a superintelligence. Literally, AGI would be only equal with human intelligence. But let us think further: There have been many efforts reported on towards understanding, modeling and replication of parts of natural and learnt intelligent behaviour of human individuals as computational intelligence. However, I am aware of much less attempts related to (i.e. to do the same with) group/team intelligence (whatever it means), and almost no information concerning the intelligence of collectives/societies of humans and that of the entire mankind.

Best regards,

I.H.

Rafael Dean Brown In my opinion, there's already enough computational power available in single computer systems to write software that can think like a human can. So AGI is merely a software, not a hardware problem. I'm not saying that AGI is already reached. The Turing test is only a metaphor. It is not the common goal of AI to make a system that is able to convince a jury that it is a human.

ChatGPT currently seems to be less smart than it could be as each 'chat' is forgotten when it is closed. A while ago OpenAI suggested to give ChatGPT a longer-lasting history of a person's interactions with it, but I think it raised privacy concerns. On the other hand, OpenAI intends to make a robot, and according to a theory, mind requires a body. If this theory holds true, we are right before another big boost of AI.

Imre Horvath At one point in the not so distant past, I would agree with you that the given definition of AGI is more appropriate for superintelligence. As I said, it is a moving goalpost. Here is OpenAI’s definition of AGI as stated in its mission: “Our mission is to ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity.” https://openai.com/blog/planning-for-agi-and-beyond

If the definition of AGI is one that equals human intelligence- this benchmark has been surpassed or at nearly human level by many AI models on a number of human intelligence measures (from IQ, bar exam, to GLUE, etc), though mostly on standardized language based tests since we lack a general human intelligence benchmark.

(see https://www.linkedin.com/pulse/5-charts-showing-large-ai-model-progress-vs-human-benchmarks-hall; https://sites.rutgers.edu/critical-ai/wp-content/uploads/sites/586/2022/01/Bender_AI-and-Everything-in-the-Whole-Wide-World-Benchmark.pdf)

So, with your definition of AGI, would you say we are near AGI with current generative AI models according to available standardized intelligence tests?

Rafael Dean Brown

Though these two terms are not used in the literature, I often argue with the terms "fragmented replicated intelligence" and "holistically replicated intelligence" in discussion. For me, fragmented intelligence is anything that has been done so far in the various sub-domains (approaching to 100) of replicating human intelligence or developing systems that are able to solve practical problems with such replicated intelligence (such as Go). The success in these fields is the results of (i) having clever guys for programming, (ii) having a reasonably well circumscribable problem, and (iii) having high capacity/performance computers. If these three are concurrently given as conditions and resources, then it is evidentially possible to achieve better results than that can be achieved by human brains due to their inherent limitations. However, for me, it is still in and remains in the realm of artificial narrow intelligence.

If the implementation of artificial generic intelligence is the question, then comes my second phrase, "holistically replicated intelligence", into the scene. This helps me to cast light on the most fundamental problem, which is unmanageable complexity. This characterizes the problematics of AGI. We do not know how to combine the various (distinct) domains of AI research and development (e.g., how artificial vision, speech generation, humanoid motion, context dependent reasoning, heuristic problem solving, interpretation of norms, and so forth come together into a holistic whole, not to mention abstraction, awareness, consciousness, emotions, and the other scientifically and computationally wicked things.) Probably, we need not only software, cyberware, and brainware for this, but also bodyware (wet organic bodies). Even if we would have a conceptual/epistemological/methodological framework for an all-inclusive synthesis, the computation of this would need such resources (including money) that will not be available for academic research. Beyond these, there is a third issue. Let us suppose that the previous two challenges are successfully taken and the "holistically replicated intelligence" is out-there as a fully-fledged manifestation of AGI. What shall we use it for? How shall we benefit from it? There is not only an ontological, but also a teleological issue related to this …

In my view, it is better to think of reproduced intelligence and intelligentization as we have thought of machines and mechanization. The latter things were considered as enablers in the physical domains of human operations. Therefore, let us try to see the former things as enablers in the cognitive domains of human operations, under proper human control and supervision. Obviously and inevitably, the danger is there because as machines could be used as weapons, intelligence can also be used as weapons. I hope strict regulations will come soon.

Imre Horvath I certainly agree with you that strict regulation is much needed. The EU’s AI Act seems to fall short in my view. Thanks for contributing for deep insights.