I modified the title for accuracy. The original title misleads, slightly, IMO: "When Will AI Exceed Human Performance? Evidence from AI Experts." I swapped AI out and replaced it with ML.
The paper itself uses the acronym HLMI (high level machine intelligence). Quoting:
"High-level machine intelligence (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers."
So a collection of machines could accomplish HLMI, without needing any single machine to do it alone.
Also, "Survey of ML researchers" is more accurate than the actual article title, "Evidence from AI Experts". The survey result isn't evidence of anything beyond what current opinion is. It sure isn't evidence for when this will actually happen.
I disagree with your modification. They are definitely talking about actual AI. Computers that are actually generally intelligent and can do every task humans can do.
The only bit of the book Superintelligence that I liked was the historical bit at the start, where it described the problem AI researchers have: as soon as they achieve a milestone in AI, the goalposts shift and that milestone is no longer considered an AI marker.
The simplest example was "can make a human believe they're talking to another human", which was achieved with ELIZA. Before then, what ELIZA could do was "AI". After, it was "just this thing, you know"...
Amusingly, the median response for 'AI researcher' is almost 40 years after 'all human tasks'. I am not sure that those being surveyed shared a common understanding of what was being asked.
I mean, AI research seems like it will be the last to go by definition: as long as there is anything else around to be taken over by AI then that is a justification that the AI researchers aren't done right?
But AI research goes when the AIs are capable of doing the AI research, not when every last AI question has been researched. Why will that take nearly 40 years longer than math research?
I don't think this is true. If you develop a meta-AI and it doesn't immediately solve all other fields, the humans won't just suddenly stop. Maybe the Meta-AI ends up solving all of the rest of the fields later and the humans turned out to be useless in retrospect, but after the Meta-AI is created and before all fields are solved there will necessarily still be humans doing work since they can't actually know that "solve AI for all fields" needs no more human assistance until it is.
I learned from working with lawyers, salespeople, marketers, managers, that I tend to understand most the difficulty of those areas I know best, and discount other areas until I've done or studied them. If researchers are asked the difficulty of automating areas outside their expertise, this may be true of them too. Their common area of expertise is of course AI itself.
I like this, but I feel it's a little optimistic (or pessimistic depending on your view). Isn't asking ML researchers when AI will dominate human performance a bit like asking a barber if you need a haircut?
One aspect of qualification is domain knowledge, which experts certainly have. Another aspect of qualification is calibration, which can only be proved & adjusted over time with a track record. A number of academic studies of prediction markets and other forecasting systems have shown that well-calibrated non-experts, with no skin in the game, often do better than actual experts, who often have poor track records as a result of incentives (or selection) to hype and extremize.[1]
Philip Tetlock has written on this topic for years. Two of his books are Expert Political Judgment and Superforecasting.
Edit: So to directly answer your question, rather than AI experts, I'd prefer technology experts (AI or otherwise) with a track record of well-calibrated predictions.
Perhaps labour economists, demographers, or sociologists?
I don't believe ML researchers are unqualified -- it's more of a potential incentive problem. I don't think it's unreasonable to suggest that people involved with / employed by a technology may have a tendency to exaggerate its benefits.
Ideally I'd trust the numbers here more if there was at least a cross-section of knowledgeable people being surveyed across a few different disciplines (or at least more than one).
A historian of science, business professor, or futurist, might provide perspective on when people automating previous trades estimated they'd be automated; how long until they were; how long previous innovations spent in various stages of translation; and how this compares to what is known so far about the AI pipeline.
It has less to do with title than systematized knowledge of models, but people with those titles are more likely to have invested in acquiring this knowledge and these models.
I'm trying to see the analogy ... if I ask a barber if I need a hair cut and they say "Of course you do." they may be lying because they want the $15 (well, I'm bald so ...)
How does the ML researcher gain from lying "In year X" if they don't believe that to be true? It's a tenuous connection.
Ha, that's fair, I take your point. I was being a little glib. I didn't mean to suggest conscious malice on the barber's part -- only that the barber may be incented to provide more haircuts than are strictly necessary.
Perhaps the analogy works better if the barber is a friendly, honest person, who takes professional umbrage when they see other people with longer hair :)
I think explaining your own actions in games is a weird thing to ask for. It requires "aboutness" (that you're thinking about the problem). Aboutness is a really inefficient way to handle problems, but it's handy because we can apply it to all new situations, because we have general intelligence. Conversely, when humans have trained hard at a task, they generally lose aboutness, like an ANN. Things are done on instinct, feeling etc. In short, the NN has been trained, and general intelligence is no longer required to do the task. Indeed, it's been superseded.
More damningly for this kind of survey: Aboutness for a single task is not the same as general intelligence. And it's general intelligence that we want.
Drones can travel pretty far. And they are a lot less energy efficient than a walking robot, because they need to constantly expend energy to fight gravity.
I agree we're still a few orders of magnitude behind on the myriad of technologies that will enable the terminator like robots... however, the exponential progress we've been making in computer science (e.g. machine learning) is making the likelihood of discovering these critical bits very realistic.
These studies are always interesting, but I don't think they have much more scientific validity than, say, asking a bunch of religious fundamentalist preachers when the second coming of Jesus is going to be. No one knows how difficult it's going to be, and while we've overcome a lot of challenges, I'm positive there are many more to overcome before we get to human-level intelligence in computers. Whatever that means.
What would it mean for a prediction to have scientific validity? No one can know the future of course. But if you are going to try to make a prediction, surveying a bunch of experts is probably the best strategy. If nothing else, the wisdom of crowds phenomenon shows the average of a bunch of wild guesses is often surprisingly close to the correct answer.
All I mean is that these predictions have to be taken with a grain of salt. I feel like there's an inherent assumption here that the scientific and political landscape will remain unchanged and progress will continue at or exceed it's current pace. There are plenty of things that could prevent this from being true -- nuclear war, economic/political collapse, natural disaster, to name a few. I do hope they're accurate -- mainly because I think it would be cool to see human-level AI in my lifetime -- but at the same time I remain skeptical
Only 20% of respondents expect "Chance of global technological progress dramatically increases after HLMI" happening 2 years after HLMI is achieved, while 80% picks the other choice, "30 years after". (Table S4)
Here is the definition of HLMI from the survey:
"High-level machine intelligence (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers."
It appears to me that if machines or software, which can be replicated billions of times in the span of two years, can do every task better and cheaper than humans, it is akin to having 100+ times more active researchers working on R&D with much higher bandwidth of communications among them than human researchers do.
It is true that we might be limited by computer hardware availability, but given that the median time of HLMI predictions is 45 years from 2016, we are unlikely to be limited by hardware then.
Another possibility is that most predictors believe they will be limited by the speed of physical experiments, my answer is that smart simulations should allow HLMI to perform many experiments without waiting for their real-world results. A recent paper from OpenAI has shown us that learning in simulations can be effectively transferred to solving real-world tasks. (https://blog.openai.com/robots-that-learn/) In 45 years, the quality and scope of simulations would be far better than in 2016.
It's sort of a weird definition for HLMI since it involves cost. I.e., if you can replace an engineer but leasing the hardware, power etc. costs more than the engineer's salary it doesn't count?
This is an important point, because ignoring that it could be the case that the first HLMI requires significant computational power to run - definitely to train, I'd expect, at least. So you can have a 5x human agent on some super computer but you have to build 1000 of those before you start talking about significant impact.
These people are not subject matter experts in these fields...
An interesting question would to have them consider the location and the IP environment. Will 10% of the public have their laundry folded by AI in the east or west first? Will it be wrapped up in patents?
Experts are known to be bad predictors of the future outcome of their fields. Many times these predictions obey the Gaes-Marreau law.
In the case of AI, according to one particular study, something similar happens: expert predictions are contradictory, indistinguishable from both non-expert predictions and past failed predictions.
It's 'Maes–Garreau law'. Kelly made up a 'law' based on a few AI predictions falling into a certain range; but you can see from the linked paper (which uses like 20x more predictions) that it's not really that accurate a law and there's more that goes into AI forecasts than just +X years. (For example, China vs the West in OP, which is interesting and makes sense thinking about it, but hadn't occurred to me.)
Yes, it's not actually a law, but a (perhaps cynical) heuristic.
Possible interpretations:
* People do not want to sound too optimistic but they hope to see some improvements before they die;
* People hope they are dead before someone proves them wrong;
* People deal with temporal magnitudes in the order of the human lifespan and close multiples.
It's funny how surgeon is listed as the farthest out application of in the abstract. I think surgery is in fact the easiest of all the listed jobs in an AI sense, but it might depend more on advances in robotics.
Doesn't machine performance already exceed human performance in a number of areas? As was just demonstrated this week when AI beat the world's beat Go player?
> Defeat the best Go players, training only on as many games as the best Go players have played. For reference, DeepMind’s AlphaGo has probably played a hundred million games of self-play, while Lee Sedol has probably played 50,000 games in his life.
But it's still much less than 100 million games. If I assume that playing a Go game (or looking at a record and understanding the moves taken) takes one hour (which is a wild guess since I'm not familiar with the game), 100 million games take nearly 11408 years.
In addition to what someone else who replied to you said about number of games, I also was listening to something recently talking about how much energy had to go into building one device that could beat humans at one task. Today and for the near future it's not a sustainable model at-scale.
If you asked aerospace engineers what they thought of the future in 1960 they would've said we'd have Mars colonies and asteroid mining would've revolutionized our economies.
I'm developing a theory that America's thoughts on what technology is capable of swings wildly between two poles (possibly with every generation?): A strong luddite streak that pooh-pooh's technology, followed by a ridiculous fantasy that technology can do everything. We're firmly entrenched in cycle "B" right now; their Martian colonies is our true AI.
We think everything is just around the corner because so much has changed over the last few decades, without realizing that those changes have only come in certain areas, while the rest of the technical world is proceeding along at a much slower, more methodical pace.
I saw an ELI5 post the other day from someone on Reddit asking what air traffic controllers did, that software couldn't do better. I actually had to sit for a moment and ponder the person (almost certainly a youth, admittedly) who thinks we're already at the point on the futurism curve where the task of safely coordinating air (or any!) traffic is better done by a computer than a person. They just couldn't wrap their heads around the idea that a group of trained people, in 2017, with advanced software and visualization tools at their disposal, might be better at that than a computer acting on its own.
The example fits elegantly because I do think AR is in our future (and our present) and I'm absolutely thrilled about what it's going to bring to the world. But the idea that we're going to replace (waste) the meat computer in our heads - let alone that we can - within the next few... What, years? Decades?
Anything in that timeline seems ridiculous to me, and not because I can't imagine such an incredible and future, but because I know how the technology works, I know how far away it is from replacing (not augmenting, which again is today and I think has a rich future ahead of it) human brains and senses. Yes, automation is going to replace all our jobs, and also the sun is going to burn up someday. We need to prepare for both - arguably the former more adroitly than the latter - at a pace that makes sense both for humanity today and humanity in the future.
51 comments
[ 2.6 ms ] story [ 109 ms ] threadThe paper itself uses the acronym HLMI (high level machine intelligence). Quoting:
"High-level machine intelligence (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers."
So a collection of machines could accomplish HLMI, without needing any single machine to do it alone.
The simplest example was "can make a human believe they're talking to another human", which was achieved with ELIZA. Before then, what ELIZA could do was "AI". After, it was "just this thing, you know"...
Philip Tetlock has written on this topic for years. Two of his books are Expert Political Judgment and Superforecasting.
[1]: https://en.wikipedia.org/wiki/The_Good_Judgment_Project
Edit: So to directly answer your question, rather than AI experts, I'd prefer technology experts (AI or otherwise) with a track record of well-calibrated predictions.
I don't believe ML researchers are unqualified -- it's more of a potential incentive problem. I don't think it's unreasonable to suggest that people involved with / employed by a technology may have a tendency to exaggerate its benefits.
Ideally I'd trust the numbers here more if there was at least a cross-section of knowledgeable people being surveyed across a few different disciplines (or at least more than one).
It has less to do with title than systematized knowledge of models, but people with those titles are more likely to have invested in acquiring this knowledge and these models.
How does the ML researcher gain from lying "In year X" if they don't believe that to be true? It's a tenuous connection.
Perhaps the analogy works better if the barber is a friendly, honest person, who takes professional umbrage when they see other people with longer hair :)
More damningly for this kind of survey: Aboutness for a single task is not the same as general intelligence. And it's general intelligence that we want.
They thought machine translation would be solved in 5 years in the 60s, too. I'm vastly more skeptical.
Only 20% of respondents expect "Chance of global technological progress dramatically increases after HLMI" happening 2 years after HLMI is achieved, while 80% picks the other choice, "30 years after". (Table S4)
Here is the definition of HLMI from the survey: "High-level machine intelligence (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers."
It appears to me that if machines or software, which can be replicated billions of times in the span of two years, can do every task better and cheaper than humans, it is akin to having 100+ times more active researchers working on R&D with much higher bandwidth of communications among them than human researchers do.
It is true that we might be limited by computer hardware availability, but given that the median time of HLMI predictions is 45 years from 2016, we are unlikely to be limited by hardware then.
Another possibility is that most predictors believe they will be limited by the speed of physical experiments, my answer is that smart simulations should allow HLMI to perform many experiments without waiting for their real-world results. A recent paper from OpenAI has shown us that learning in simulations can be effectively transferred to solving real-world tasks. (https://blog.openai.com/robots-that-learn/) In 45 years, the quality and scope of simulations would be far better than in 2016.
This is an important point, because ignoring that it could be the case that the first HLMI requires significant computational power to run - definitely to train, I'd expect, at least. So you can have a 5x human agent on some super computer but you have to build 1000 of those before you start talking about significant impact.
An interesting question would to have them consider the location and the IP environment. Will 10% of the public have their laundry folded by AI in the east or west first? Will it be wrapped up in patents?
In the case of AI, according to one particular study, something similar happens: expert predictions are contradictory, indistinguishable from both non-expert predictions and past failed predictions.
https://intelligence.org/files/PredictingAI.pdf
Could you please link me to something that explains it?
Possible interpretations: * People do not want to sound too optimistic but they hope to see some improvements before they die; * People hope they are dead before someone proves them wrong; * People deal with temporal magnitudes in the order of the human lifespan and close multiples.
The last one, and then the first, are my guesses.
> Defeat the best Go players, training only on as many games as the best Go players have played. For reference, DeepMind’s AlphaGo has probably played a hundred million games of self-play, while Lee Sedol has probably played 50,000 games in his life.
We think everything is just around the corner because so much has changed over the last few decades, without realizing that those changes have only come in certain areas, while the rest of the technical world is proceeding along at a much slower, more methodical pace.
I saw an ELI5 post the other day from someone on Reddit asking what air traffic controllers did, that software couldn't do better. I actually had to sit for a moment and ponder the person (almost certainly a youth, admittedly) who thinks we're already at the point on the futurism curve where the task of safely coordinating air (or any!) traffic is better done by a computer than a person. They just couldn't wrap their heads around the idea that a group of trained people, in 2017, with advanced software and visualization tools at their disposal, might be better at that than a computer acting on its own.
The example fits elegantly because I do think AR is in our future (and our present) and I'm absolutely thrilled about what it's going to bring to the world. But the idea that we're going to replace (waste) the meat computer in our heads - let alone that we can - within the next few... What, years? Decades?
Anything in that timeline seems ridiculous to me, and not because I can't imagine such an incredible and future, but because I know how the technology works, I know how far away it is from replacing (not augmenting, which again is today and I think has a rich future ahead of it) human brains and senses. Yes, automation is going to replace all our jobs, and also the sun is going to burn up someday. We need to prepare for both - arguably the former more adroitly than the latter - at a pace that makes sense both for humanity today and humanity in the future.