I think your "of course" is glossing over too much. It is very, very far from a given to me that GAI is required in order to solve self-driving cars well. I think it's reasonably possible we'll have good self-driving cars within 15 years (50% sounds about right to me for the longer end of that time horizon), but I highly, highly doubt we'll have GAI in anywhere close to that timeframe.
GAI is a pipe dream. There is no technology that we can even conceive—let alone build in the next 10-15 years—of which passes as GAI. Our current AI is based on statistical inference, GAI will need to be able to do far more than infer from data and interpolate actions.
No we will never have GAI.
---
I know never is a long time, but I can comfortably predict this because there is not only technological problems with GAI, but also philosophical problems. The way we define intelligence is such that if we find or invent something that combats human intelligence we immediately redefine it to exclude that aspect.
I'm trying to understand the worldview that leads to what you believe. As far as I can tell, you're assuming there is a supernatural aspect to the human mind.
According to my worldview, humans are just robots built out of DNA, proteins, etc. The idea that machines couldn't do everything a human can do (and more) doesn't make any sense to me.
I’m sorry but I don’t have that world view. I must not have been clear enough (english is not my first language; and I’m not even super clear in my native language).
I actually don’t believe in the superiority of the human mind, I think the word intelligence is actually a scientific distraction. Humans are remarkably—but not uniquely—adapted at communicating with each others and manipulating our environment for our benefit (most of the time; but sometimes to our determent, e.g. by causing a climate disaster). My statement about intelligence was not to share my believe about humanity, but a statement about human believes about our humanity (a theory of mind of sorts). I believe that humans are so human centric that we will always evaluate intelligence in human terms. As such nothing will ever supersede it as we will simply alter our former believes in order to keep our mind at the center of the universe.
Now as for machines that can do everything that humans can do (and more), I still consider this a pipe dream, even if we disregard the word intelligence. This is not because I believe the human mind is superior to machine, it is more because we haven’t conceived of how we would do this. Statistical inference has its limits, even with all the computing power in the world, it will never do what specialized machines like our brains have evolved over millions of years to do (well maybe if we allow our machines to freely evolve and procreate and set them loose in a dynamic self sustaining environment; but that wont provide any benefits for us).
I also have my doubts about the utility of AGI. People state that full self driving cars is easier with AGI, but I haven’t seen any proof of this statement (nor for the utility of AGI driven cars for that matter; we’ve had self driving trains for decades without AGI). With traditional machine learning models, there is tons of utility in the data analysis in all sorts of fields (just like there was with factor analysis and linear regression before it), but with AGI I fail to see what we will get from a vague notion of can do everything humans can do, which doesn’t apply to traditional machines that are specifically programmed for that task. In short, I don’t even think there is a market for AGI.
I may have read too much into your original post. I think I get what you're saying now.
We apply the concept of intelligence to other species, so I don't see why we couldn't apply it to machines. If people want to redefine intelligence so that humans stay on top, then they are wrong to do so. There is undeniably an objective component to it, even if there are many cultural factors at play as well. If you don't like the word "intelligence", replace it with "competence", "productivity", "performance", etc. and the same argument will apply.
You raised two other issues: the possibility of AGI and the utility of AGI.
I think we have good reasons to believe that it's possible. Three possible future scenarios: (1) we get to the point that we can simulate brains to a high degree of accuracy in software, (2) we get to the point that we can build artificial brains out of neurons made of silicon, (3) we get to the point that synthetic biology is advanced enough that we can engineer new kinds of intelligent organisms. These might be science fiction by today's standards, but they show that there is no issue with conceivability. As an aside, the human brain runs on relatively low power consumption and operations per second, so computational power isn't a bottleneck. We have enough compute, we just don't have the right algorithms yet.
As for the utility of AGI, I strongly disagree that there is no market for it. By definition, an AGI can do anything that a human can do. There is already a job market for humans. As soon as AGI can compete with humans in domains like management, engineering, scientific research, education, etc. I don't see why there won't be a huge economic incentive to adopt it.
> As an aside, the human brain runs on relatively low power consumption and operations per second, so computational power isn't a bottleneck.
You hit the nail on the head here (pun intended). Brains (and other adoptive biochemical processes which reinforce behavior) are way more ingenious then to be described as a simple computer. It is not just neurons firing in patterns with synapses which can reinforce certain pathways. No, there is a whole biome in our bodies, from our guts, our muscles, through hormonal systems and yes, nervous system. Our behavior is also influenced by bacteria and other microorganisms that doesn’t even share our DNA (but sometimes actually interacts with our DNA). This complex machinery has evolved over millions of years, and I don’t think there is any way for us to simulate it. Even if we look outwards, our behavior is just as much influenced by each other as it is by our brains. So for a successful simulation you’d need to “raise” your machine in a similar manner, allowing interactions with a whole community (or more) that treats it as equal. Our brains are merely a tool which encodes this interaction as learned behavior. Just building a brain will miss all of that and give us a lousy model.
In statistics we learn that: “No models are correct, but some models are useful”. I’ve always taken this to mean that there is not always utility in a 100% accurate model. I think this applies for AGI. If we create a machine that can do everything that a human can, and better, what can we use it for? You claim increased competence, productivity, and performance, but we already have machines that do that. And we humans have found a way to work pretty well with those. We even have fully automatic vehicles that transport millions of people every day (e.g. in Vancouver B.C. and Copenhagen). It doesn’t need to be better then humans in everything, just accelerating and decelerating in the right places, it doesn’t need to know the airspeed velocity of an unladen swallow, it will only need to know if something is blocking the tracks. Traditional statistical inference will make these systems better and cheaper, and they will benefit humanity, there is no need for them to have generalized knowledge beyond their immediate utility, so they will never be given one.
I agree with you in the sense that I think many problems don't require general intelligence once you have a solution. However, to come to those solutions it took decades or centuries of intelligent humans working on those problems. The potential benefit of AGI isn't in automating the problems we already know how to solve (although it might help refine our solutions). It's in accelerating progress in domains where there are unsolved problems.
Sometimes we get hung up on comparing AI to humans. I wouldn't want an AI that does math and physics in the same way that humans do. Our brains aren't very good at those things. I'm more interested in the potential for AI to find new, more effective ways of working in these areas.
The simple difference is humans have the biological goal to replicate their genes. That is the root source of intelligence. Software does not because it has not evolved. If you want real AI then start with simulating life in a simulated environment as complex as the real world for many thousands or millions of generations.
Humans may be like robots but their programing is not statistical pattern matching or Boolean logic. Rather it is the intelligence selected for by evolution to achieve the goals of an individual organism.
How does this not make sense?
Intelligence does not just happen absent goal directed evolution.
As far as I can tell, your points don't apply to what I said. I'll quote myself:
>The idea that machines couldn't do everything a human can do (and more) doesn't make any sense to me.
In other words, it doesn't make sense to me when I hear impossibility arguments about AGI that appeal to something special about humans. I'm not making a claim about software and hardware in 2023. I'm saying in principle, there is nothing stopping us from building artificial systems that are just as complex and capable as evolved biological organisms, other than our lack of knowledge of how to build such systems.
>Software does not because it has not evolved
Software does evolve. To get evolution you need three things: replication, mutation, and selection pressure. We replicate, mutate, and select software all the time. I don't think you need a large scale simulation to get evolutionary dynamics.
>Humans may be like robots but their programing is not statistical pattern matching or Boolean logic
That might be true, but I'm also skeptical. Part of the problem is we don't really understand intelligence very well. If we really understood what makes humans capable of intelligent behavior, we would be able to replicate it in artificial systems. I don't see a fundamental barrier other than having the right algorithms.
You might be thinking about evolutionary learning (or more broadly evolutionary algorithms), which is a machine learning method different from—the more common and more successful—artificial neural networks. Several implementations actually exist, and I think some are even quite successful, but none that I’m aware of have been shown to have AGI capabilities.
The question of if an AI can be regarded as a sentient being has already been debated ad nauseam in science fiction (e.g. Data in Star Trek).
All that's needed is someone to achieve GAI at a supercomputer level. From then on, it's just a matter of scaling it down. Your iPhone has more compute power than a supercomputer 30 years ago, after all.
And what's also a way to "cheat" is fusing computers and organic matter. Brain-in-a-vat style, or, for yet another Star Trek reference, Voyager's organic compute network (which IIRC even suffered a pathogen infection at one point). And that is definitely possible in the next 10-15 years. Give Musk's Neuralink enough apes to euthanize and it will get done.
GAI is not limited by computing power, but theory. Yes we will be able to make better inference by feeding more data into our models, supervised learning will be able to work with thousands of parameters, etc. This will lead to all sorts of new discoveries in many fields, including biochemistry, astrophysics, neurology, etc. It will also bring us a bunch of new fun toys to play with including better and more capable robots, and it will also lead to better engineered social systems (including traffic engineering; but also public transit, and—yes—self driving vehicles—but we already have self driving trains, so it is not as big of a deal as you think it is).
However this misses my point about AGI. It is not about computing capabilities but theory. AGI will have to evaluate situations that has never arisen before without having data that directly directs it. I don’t see how statistical inference can do that, and as such we will need an entirely new methods of training it. Like I said, we haven’t even conceived how an AGI technology works, there are no algorithms which will work in theory even with unlimited computing power. Contrast this, for example, with quantum computing, which has a bunch of algorithms demonstrating it’s capabilities. With AGI all we have is “perhaps our cars become better at driving them selves”. I’m not buying it.
Your iPhone has more compute power than a supercomputer 30 years ago, after all.
In 30 years we will not have an iPhone-sized device with the power of today's supercomputers. Transistor scaling is in its final stages. We'll be lucky to get 2-3 more process nodes and that'll be it.
Even if we find a novel semiconductor that can scale 1000x better then silicon transistors (say DNA semiconductors), then we are still limited by classical computing. Meaning our most intelligent systems are still just good old statistical inference (I say still, but statistical inference is really powerful, and getting 1000s of orders of magnitude better at statistical inference means a world of amazing discoveries).
Perhaps quantum computing will enable new kinds of inference, then we will need to scale superconductors like we did transistors in the past 75 years, that is possible (not in the next 10-15 years though; 30-40 years is more realistic). However I don’t think this will lead to anything more then amazing statistical inference. People theorized artificial neural networks way back in the 1940s and we already had algorithms for it in the 1970s. I’m not aware of any algorithms, not even theoretical ones that gives us AGI, on any computing device (not even quantum computers).
> the way we define intelligence is such that if we find or invent something that combats human intelligence we immediately redefine it to exclude that aspect.
And what makes you so sure the opposite won't happen?
E.g. GAI will be achieved earlier by excluding some aspect of human intelligence that some people deem essential.
The whole "problem" is that we can't agree on what constitutes intelligence.
I predict that it won't be long before AI systems can score 200 on standard IQ tests.
Have you tried any of the state-of-the-art out there?
ChatGPT is easily circumvented to violate all its own guardrails with a few instructions.
Knowledge without understanding.
It clearly doesn't understand it's own rules well enough to realize when it is asked to violate the rule "don't do C" that being asked to do A->B->C is the same as being asked to do C.
We need true GAI to do this, of course, and the question is: will a "real" GAI be willing to drive a car? How much will we have to pay it?