In 2016, a prominent computer scientist–a pioneer in artificial intelligence, he would be a winner of the 2018 Turing Award–said:
We should stop training radiologists now, it’s just completely obvious within five years deep learning is going to do better than radiologists.
Hasn’t worked out that way. AI can be a useful supplement to a human radiologist, but I don’t think it’s being used anywhere on an exclusive, human-radiologist-replacing basis.
Just as well that the training of radiologists wasn’t shut down.
It is often unwise to make radical changes based on the opinions of experts who are proponents of particular technologies. (Of course there are cases where such radical changes are called for–the aviation experts who in the 1920s and 1930s foresaw a major role for aviation in naval warfare, for example.) But in the case of robotics/AI at the present time, I think over-claiming is generally more likely than under-claiming.
Lets define this as many are confused. AI, Deep Learning, etc is just machine learning.
Now machines can learn at a very high level, as the data they can hoover up, along with appropriate algorithms, to deal with that data, is far beyond anything an actual mind can do. Its dumb though, real dumb. ;)
The said a decade ago that self-driving cars were a year or two away. No one who actually works with “machine learning” believed it.
Any one year old can still outperform the most advanced algorithm in identifying a cat, and it’s not even close.
The first tell that it was BS was that the person making the prognostication was a computer scientist (had to think if I should use scare quotes for scientist), which is to say someone that knows nothing about radiology, rather than a radiologist. Every time computers attack some new domain, progress seems quick at first as the simplest cases usually yield to brute force. It then, just as quickly stalls out with progress gradually slowing to a halt.
My memory is that the first place they applied computers was to mammograms and used fairly standard image processing tools to highlight anomalies that were then passed to radiologists for final judgement. The computer, unsurprisingly, was much better at discerning very subtle artifacts than a person but still required the judgement of knowledge and experience before cutting someone open or even poking them with a needle.
It doesn’t seem that things are much further along now. If you want an easily seen example of AI fail, just look at how search engine optimization has about ruined Google and every other search engine. Try to find something as specific as a particular size bolt on Amazon. Both places where you’d think the incentive would be strong to keep things relevant to the person with the money.
One could easily say that AI methods could reduce work needed to do some tasks by like 99%, but advocates have to say it’ll be 100% and they set themselves up to fail. It’s so dumb.
Keep in mind an AI holds both the chess, and go championships. Not trivial at all.
Speaking of Naval Aviation, tomorrow is Midway Day.
HUZZAH for LTCMDR Dick Best!
“Keep in mind an AI holds both the chess, and go championships. Not trivial at all.”
Actually, exceedingly trivial.
A game is an artificial environment with a highly constrained universe of choices, a limited number of allowable operations, and a fixed definite goal. As Brian and MCS pointed out, the computer ‘player’ is simply brute-force forecasting the optimum move given the conditions on the game board, the rules of the game, and quite likely strategies previously developed by humans. It was just a matter of time and Moore’s Law until there was a machine with enough horsepower to do that fast enough to approximate a human player, who gets up from the game and performs a dozen tasks no computer can yet do.
Both chess and go use brute force search algorithms. And have for many decades. Zero “intelligence” involved. AI based chess and GO programs never even made into the 1980’s as competitive players. Lisp or Prolog wont get you very far.
There was some actual encoded basic intelligence in AI software in the 1970’s to 1990’s but the fundamental problems of knowledge engineering made them marginal in the real world. Conceptual modeling and Pierce ontologies looked very promising the late 1990’s but the area soon veered off into worthless academic trivia.
There is zero intelligent or cognitive insight in AI software based on machine learning (neural nets Mark II). Its just a new way of encoding the mathematical results of massive search / pattern matching algorithms. Thats all. Its brute force math based on huge advances in GPU computational power.
As to any deeper meaning to the results, none. One of my kids is doing a PhD in an area of pure math that overlap with ML. One of the project applications is a variation of something I did the old fashioned way, feature extraction from images, over 20 years ago. We got better results in less time doing it the old fashioned way. A sophisticated problem domain model and some fancy signal processing math. Running on hardware several orders of magnitude less powerful than needed for the ML software.
Every AI academic who makes grandiose claims for his work or area is at best a fool, at worst a total charlatan. A statement that has be true for the last 70 years. As its obvious they haven’t a clue as to what intelligence, knowledge representation, problem solving, domain expertise, or reasoning with knowledge actually involve.
The Greeks asked all the fundamental questions 2500 years ago and so far no one has come up with any adequate answer. Although Pierce and his followers came up with the most pragmatic workaround so far.
I’ve been working on automated target detection/classification for a couple decades now. Whenever new algorithms don’t “work”, which is always, the reply is always “we need more training data”, then things go back to very basic approaches that actually work, then some new algorithm comes along that gets everyone excited, rinse repeat, etc., on a several year cycle. Seems like the timescale involved is approximately the PhD length time.
SF author James P. Hogan had a book where he noted that the distinction that AI lacks — because we have zero idea how We Do It — is “common sense”.
The example he gave is, “The cat has fleas. We want to get rid of the fleas.”
Computer solution: Heat kills fleas, throw cat in furnace.
Now, even a 6yo should generally reject this, because heat also kills cats, and that’s not a valid solution. That’s “common sense”.
But unless you TELL the computer not to do that, it doesn’t know better — because it lacks even a rudimentary component of “common sense”. Where and how the hell that develops in our intelligence/learning process is something we have no clue about.
And no AI/Machine Learning, etc., will be able to largely, much less fully, replace a human until this is resolved.
And, since we don’t grasp common sense AT ALL, it’s an open ended problem with no possible estimate of resolution. True AI is indefinitely postponed.
}}} It doesn’t seem that things are much further along now. If you want an easily seen example of AI fail, just look at how search engine optimization has about ruined Google and every other search engine.
Ca. 2001, back long before Google was even the “main player” in search engines, there was a search engine called “Infind”. I have zero idea how it worked, and it disappeared about 2005 or so… but in all the times I used it, most of the time the link I sought was on the first page, and it was never ever NOT on the second page if not the first.
No idea if the explosion of pages would have weakened its functionality, or any other problem, but it would be nice to seek out that algorithm and find out more about how it worked.
I don’t see how the fact that SEO has “ruined” google is due to AI…that’s simply due to people gaming the google rules–when links were what determined ranking (which was google’s amazing breakthrough, those of us who used pre-google search engines should remember how much insanely better google was than previous attempts…), people started making tons of phony links to bump to the top, so google tried to change things to eliminate that, by looking at stuff like how long people spend on a page and how much of it you look at, which is why recipe pages all have tons of text at the top before you get to the actual recipe, etc.
Brian….”google tried to change things to eliminate that, by looking at stuff like how long people spend on a page and how much of it you look at, which is why recipe pages all have tons of text at the top before you get to the actual recipe, etc.” Also lots of photographs. Just about impossible to find a recipe that just tells you how to make it.
Probably right that this is motivated in part by SEO, but also, I think, by the pervasive cultural attitude that everything has to be a Story and a Big Deal.
David: Well, I was told by a guy who did SEO professionally who was advising me on a website I had a few years ago that that’s the reason recipe pages are all like that, so I think everyone has to do it, then the contest becomes how to make all that filler more interesting than all your competitors…
As to radiology, many simple x-rays, like those in emergency rooms looking for fractures, etc, are read in India. Radiologists don’t like to get up late at night.
There are quite a few AI applications which are now pretty useful: for example, voice transcription on my iPhone now works surprisingly well. Google Translate doesn’t yield results that are eloquent, but they are usually enough to get the sense of a document.
It’s been said that once any AI application actually works well, it’s no longer considered AI…
}}} It’s been said that once any AI application actually works well, it’s no longer considered AI…
Easy peasy: Do the results pass the Turing test?
If a knowledgeable person cannot easily tell the difference between what is done by the AI and what is done by a human (e.g., for language translation, can a dual speaker tell the difference between a human translated document and a machine-translated one?) then it’s no longer “AI”, it’s just “I”.
For most things — even translation — we ain’t there just yet.
Mike K.,
I read about a radiology group that had a villa in Italy where they rotated members on a monthly basis. This would seem to be a better and more accountable way than Bombay. More enjoyable for the radiologists and almost certainly more expensive.
It’s probably a bit unfair to blame AI for poor search results. Google doesn’t really care if I find what I’m looking for, they just want me to click on a paid link or two and Amazon seems to have adopted the strategy of grocery stores changing layouts every few weeks to encourage wandering around. The joke is on Amazon since I almost always avoid them if I’m looking for something specific just to save time and aggravation. Of course, as the saying goes, if you aren’t paying, you’re the the product. Oddly, I don’t remember getting anything from Amazon for free.
I was thinking more about the chimera of natural language interfaces that were supposed to revolutionize everything from search to spread sheets to programming itself. None of that came to pass. In programming, it’s gotten even worse. As if either the presence or absence of semi-colons, commas and brackets sending your computer off into the weeds wasn’t enough, the latest “cure” makes white space do the same thing. Spread sheets just add another half dozen layers of menus.
“Google doesn’t really care if I find what I’m looking for, they just want me to click on a paid link or two”
Well, but they do in the sense that if I don’t find what I’m looking for, I’ll soon go elsewhere and they won’t be able to sell those ads…
I’ve recently been using the Brave search engine, which apparently is not base on google, unlike most of the other options…and it is terrible. Just awful. It’s like being back in the pre-google days when you were very unlikely to find what you were looking for at all.
This article that talks about the problem of teaching comprehension rather than just reading as rote memorization of individual words.
https://nypost.com/2022/06/01/expert-idiocy-on-teaching-kids-to-read-is-beyond-comprehension/
A good example of how word processors seem to have stopped evolving at simple spell check. I think the longest I’ve ever been able to tolerate any sort of grammar check is about ten minutes. They always seem to want to substitute the syntax and vocabulary of a 12 year old for mine when the “suggestions” don’t totally reverse my meaning. The spell checking isn’t that great, I’ll be impressed when a spell check flags “fro” unless I’m writing about obscure wood working tools. How hard could it be to have a list of obscure words that seldom are used in normal discourse to point out as well.
A game is an artificial environment with a highly constrained universe of choices, a limited number of allowable operations, and a fixed definite goal.
This statement applies to school as well.
There are quite a few AI applications which are now pretty useful: for example, voice transcription on my iPhone now works surprisingly well.
AI works less well for competitive activities. Competition leads to arms races that tend to erode any player’s advantage.
For example, Internet search technology is better than ever, yet it seems more difficult to get accurate information online than it did in the past. Everyone games the search systems; fake information is the norm. “User reviews” of products and services used to be reliable alternatives to false/exaggerated advertising claims. Now, so many reviews are fake or are slanted in an effort to gain traffic and clicks, that in many cases it’s difficult to learn from them anything useful about particular products or services.
I was first on line with Windows 3.11, Mosaic for my browser and either a 9600 or 14.4K modem. Before Alta Vista, you found URL’s in magazines or other sites. You had to type them in with all the backslashes and all. Things have gotten better.
Actually, that’s wrong, I first connected to a GE time share computer at 110 baud through an acoustic coupler with a Frieden communications terminal (dumb terminal) that cost more than $8,000 in 1968. Connect time was $10 an hour and compute time was $0.05 a second. Things have gotten much much better.
MCS…”Try to find something as specific as a particular size bolt on Amazon”
If you search for ‘bolts’ at Grainger, they give you a form that lets you select thread size, material, finish, etc….whereas if you search for ‘relays’, the criteria on the form are coil volts, contact rating, etc. I don’t know whether they require vendors to specify the values in machine-readable format, or if they have a large staff going through the vendor catalogs and pulling out the relevant information.
David,
I’d been using Grainger since the ’60’s, as had my father and his father before him. Their print catalogs were outstanding with a great deal of specialized information, all well organized and indexed so you could find that one special motor or whatever that you needed. I consider their web site all but unusable.
As an example, type “mcmaster.com” into the URL line of your browser and look at the difference that organization makes compared to “search”. Admittedly, I haven’t been on Grainger for a year or so and usually only when I had a specific part number I was needing so maybe the general experience is better now than it has been for the last ten years. McMaster-Carr and Grainger are in somewhat different businesses as well and I still find Grainger useful when I need to replace a motor but I almost always have to find it somewhere else first.
Parametric searches are great until they aren’t. As an example take single phase motors or just about any component that runs on “house current”. The voltage might be specified as 110, 115, 120, 125 or 130 volts. If there’s any rhyme or reason, I’ve never seen it in 50 years. They all are intended to operate on U.S. line voltage but the manufacturers insist on all of those ratings. and the same happens at higher voltages, especially when a motor is rated to operate here and in Europe. If I know exactly what I want down to the part number, no problem, if I’m just looking for something close enough to work, It’s always a struggle to figure out just what parameters will give me a possible part.
>
There are quite a few AI applications which are now pretty useful: for example, voice transcription on my iPhone now works surprisingly well. Google Translate doesn’t yield results that are eloquent, but they are usually enough to get the sense of a document
<
Not AI. Just brute force pattern matching offloaded onto a very powerful server. For untrained voice recognition the algorithms really haven't changed in decades. At least since the mid 90's last time I worked on one. But the processing power available now means that you can get a very high accuracy rates without the need for training to a particular voice. The training pass of 20 / 30 years ago was to give the recognition software a sporting chance of getting it kinda right with the local processing power available at the time.
As for language translation. That was one of the epic fails for Artificial Intelligence and traditional language theory. All language translation software based on formal linguistic theory models failed terribly. Then someone had the bright idea of running the largest corpus of accurately translated bilingual text through some very basic statistical analysis and pattern matching. The transcripts of the Canadian Parliament in English and French. The translation software based on this data-set proved to be not only accurate but very usable. Unlike all previous translation software.
Then the internet came alone providing a huge data set of native language texts and enough basic translations to get something usable up and running. If the translated text amount is great enough the translation can be very accurate. There is something very magical about watching the NRK TV news in Norwegian which broadcasts a simulation closed caption in Norwegian in a form that Chrome can feed into Google Translate for simulations translation into English. So you get a simulation translation in English at the bottom of the screen.
Not really AI, just some fancy stats and even fancier math and some very serious processing power at both ends. Still very impressive no matter how you look at it.
“Just brute force pattern matching offloaded onto a very powerful server.”
That’s as a web app. Its amusing how much you feel the need to put down AI. Anything beyond a noble being doing it himself, by hand, is shat upon.
I told you it was dumb, but dumb does not matter if you have enough sample points. However you would like to characterize it, AI is the future of war, among other things. When your OODA loop runs in microseconds, the ability of humans to even make a difference, goes away. By the time you notice its happening, its happened. Hook that into hyper sonic weapons and its over before it began. ;)
AI is the future of war, among other things. When your OODA loop runs in microseconds, the ability of humans to even make a difference, goes away.
What’s the military equivalent of a fat-finger error?
DARPA has run competitions for several years now where AI “pilots” invariably wipe the floor with humans. That’s actually a much, much, much easier problem than “is this thing a tank” since we have far more intimate knowledge about how airplane mechanics work than the human visual and neural systems.
Computers will always have the edge in reaction time. The human nervous system operates at the leisurely rate of about 100 miles per hour compared to electronics which work their way down from the speed of light, call it 100,000 time faster. Eventually I expect that strategies will be found that exploit weaknesses in the algorithms.
The place where there’s real money to be made and no lack of voluminous data perfectly suited for crunching is Wall Street. I’m pretty sure it’s not for lack of trying, yet I’ve only heard of time arbitrage as a success. Again, exploiting speed and brute force rather than anything resembling intelligence. If anyone ever managed to beat the market this way, it would probably collapse as soon as it was generally known. Whatever utility the stock market has left financing productive enterprise would quickly be subsumed in a bot war.
How long would Major League Baseball last if they replaced human pitchers with machines or the the NFL replaced running backs with MRV’s?
If I recall correctly, the massive edge the pilot AI had in recent competitions is it pushes the performance of the “plane” far closer to the extreme tolerance than humans are willing to do. Something that could be used to train pilots better, I suppose. Fact is that the AI won’t ever get tired, stressed, etc., and it’s only a matter of time before pilots are all automated, certainly for military purposes. (And recall from above I’m by no means an AI maximalist, I just think dogfighting is more like chess than “is this a cat”…)
A missile will always be able to out maneuver any airplane, period. In air to air combat, if you’re close enough to see the other guy, you’re doing it wrong. Whoever shoots first wins. It’ll come down to whichever side has sensors better at penetrating the other side’s “stealth”. Once you get rid of the pilot, there’s really no difference between aircraft and a missile. A cruise missile is just another name for a one-way drone. No reason you can’t extend the concept to the air war. Except, that is, for all the Air Force generals with visions of the Read Barron dancing through their heads.
Well it’s been obvious since vectored thrust was well demonstrated decades ago that the human pilot is the massive limiting factor in airplane performance but the fact that the Air Force is run by fighter jocks means they refuse to accept that reality.
At some point the USAF will face some enemy that uses drone swarms to wipe out an entire squadron, but they may get lucky in that there’s no obvious adversary to develop it…China’s the obvious suspect but they seem happy to use economic power rather than military to achieve their goals, and one-child is going to end their ambitions before they’re realized anyway…
China seems as obsessed as we are with the fighter treadmill. Which is not to say that we might not be surprised by cheap Chinese drone swarms supplied to some other enemy just as Russia was surprised by Turkish drones in Ukraine. The drones just aren’t as sexy.
It might not have anything to do with the US. My money is on a scenario where Chinese jets, ships, and tanks get shredded by clouds of (comparatively) cheap Taiwanese drones.
But it doesn’t have to be China/Taiwan. Any of the non-nuclear powers would be wise to see if they could buy/build better defense capability by skipping the traditional – and horrifically expensive – systems (late-model fighter jets, aircraft carriers, etc.); and building lots and lots of drones.
I’ve been saying this for years.
“Big War”, the way we’ve been fighting it? It’s ripe for a huge shift. The Ukrainians are pointing the way.
I think the future is going to be less “Big Army” and more “Dragon’s Teeth”, with a small cadre of professional soldiers as leavening for the levee en masse out in the population. You can’t afford to keep everyone under arms, but you can cycle them through some sort of national service program to get them accustomed to things, and then seed small arsenals out in the countryside for them to use.
Couple that with weaponized drones and all the rest? War ain’t going to be the same, at all.
The delicious irony of this is the fact that the people who pioneered and actually taught all the third-world types to go after rear-area logistics, coupled with front-line attacks on attrited combat forces? That would be the Soviets, who simultaneously were teaching Africans how to do what they were having done to them in Afghanistan. It’s like their military wasn’t (and, isn’t…) talking to each other, at all.
The solution to most larger human war devices — tanks, aircraft carriers, etc — was demonstrated about 50 odd years ago with Project Thor.
They’ve used the “must be the size of telephone polls” as an excuse (presumably) to not pursue Thor ever since. Offhand, a large crowbar is quite doable and probably sufficiently effective in larger numbers.
It’s possible this is a secret no one knows about, but if it ain’t, someone, somewhere, is going to be massively surprised when the next war heats up sufficiently — because someone with space capability IS going to put it into play.
The big drawback with Thor is that there’s no way to tell at launch that it isn’t an ICBM. A distinctly non-trivial ambiguity. Similarly, If launched from orbit, there’s no real way to tell it isn’t a re-entering warhead.