The AI Revolution Hasn’t Happened Yet
medium.com
medium.com
> Thus, just as humans built buildings and bridges before there was civil engineering, humans are proceeding with the building of societal-scale, inference and decision making systems that involve machines, humans and the environment. Just as early buildings and bridges sometimes fell to the ground — in unforeseen ways and with tragic consequences — many of our early societal scale inference and decision making systems are already exposing serious conceptual flaws.
Well, with a lot of assumptions, if believe they hold in practice, we have a lot of powerful theory for regression analysis, and we didn't get that theory by "reverse engineering intelligence". We got hypothesis tests, confidence intervals, prediction intervals, etc.
So, more generally, we can proceed mathematically: State some assumptions and then use those to prove some theorems. We want the assumptions to be justified for our practical applications and we want the consequences of the theorems to be powerful for the applications. These steps have been followed often enough in math before, back to Euclid's plane geometry, the Pythagorean theorem, trigonometry, spherical triangles, the area of a circle, the volume of a sphere, the wave equation, ellipses, analytic geometry, calculus, differential equations, the stiffness of space frames, etc. For the current applications with no theory, maybe we need to stir up some more theorems and proofs.
Then I mentioned that from Euclid through calculus to wave equations, there's more math from theorems and proofs we can and sometimes do apply. And we can stir up still more applicable math.
My goal was just suggesting how to do better with real, valuable applications to important real world problems. We have such examples from US national security -- the A bomb, the H bomb, GPS, stealth, phased array radar, adaptive beam forming sonar, etc.
By analogy, I was suggesting a better socket wrench or numerically controlled milling machine and not a self-driving car. I was not suggesting anything that we could regard as intelligent, not even as intelligent as a field mouse.
IMHO, powerful math with valuable applications is now very doable. We can schedule equipment maintenance, airline crews, airline fleets, workers, trucks, etc. We can have computers do the data manipulations specified by the math and have the Internet move the data.
But for also having systems that are as intelligent as a field mouse, kitten, puppy, octopus, song bird, even walk as well as a cockroach, that's harder.
At one time I worked in some AI based on some MIT AI work; I wrote software, worked with GM Research, gave a paper at the Stanford AAAI IAAI conference, as sole author or co-author published a list of papers. Here I was not suggesting anything that has anything to do with any of the MIT AI work I've known about.
For people at MIT who have done work more like I have in mind, I can think of D. Bertsekas and M. Athans.
Athans was in deterministic optimal control, and the OP mentioned that that field did "back propagation" long ago. Athans is, was, a good applied mathematician. When I was at FedEx, I chatted with him in his office on how to find how best to climb, cruise, and descend airplanes. He told me a cute story about how an F-4 could get minimum time to climb, say, IIRC, to 100,000 feet: Climb up to just 5,000 feet or so, go into a dive, get supersonic where actually the drag was less, and then go nearly vertical, all supersonic, directly to 100,000 feet.
For neural networks, IIRC there is some nice math that shows how general those can be for representing functions, and for some parts of stochastic optimal control Bertsekas proposed such a use for neural networks. There he, as usual, was being mathematical.
But don't you think we also need more powerful, but less mathematically tractable, formalisms like probabilistic programs [1]?
My view is that for high end applications of computing now, drawing from, building on, Çinlar or Bertsekas is about the best we can do and, due to current computing and the Internet, suddenly terrific.
But my view of something real in AI, say, as good as a kitty cat, ..., human will much more directly use very different approaches; that if in the basic core programming there is some math in there, then it will be darned simple.
So, my current view is that the good approaches to such AI will make direct use of little or no pure or applied math. Instead, my guess is that animal ... human intelligence is just some dared clever programming, rediscovered and re-refined so far many times here on earth. From the many times, my guess is that there is basically one quite simple way to do it.
My guess is that the sensory inputs, first, feed data that becomes in the brain essentially nouns: Floor, rock, water, etc. Early on the data on the nouns is quite crude, but later with more experience gets refined. E.g., a kitty cat quickly learns that floors are solid to stand and run on, and some are shaky and might result in a fall. Then with more input and experience, some verbs are combined with some of the nouns. The strength of the combining is mostly just from experience; yes, we could write out some simple strength updating algebra. But to be cautious, the learning is deliberately slow: E.g., not everything round on the floor is good to eat.
There is a continual process to simplify this data, i.e., a form of data compression, into causality, e.g., learn about gravity. The learning is good enough to identify the concept of gravity as the cause that makes things fall and to reject irrelevant data like just what are falling from, a table, a window seat, the top of a BBQ pit, a tree limb, the second floor landing, etc. Also reject night, day, hot, cold, and other irrelevant variables -- that's smarter than current multi-variate curve fitting that has a tough time appraising what variables are likely irrelevant.
If I were going to program AI, that would be the framework I would use. I regard the learning as close to a bootstrap operation -- the first learning is very simple and crude but permits gathering more data, refining that learning, and doing more learning. To get some guesses on the details, watch various baby animals and humans as they learn.
I see no real role for math or anything I've heard of in current ML/AI, and I don't think it's much like rules in expert systems. And my guess is that the amount of memory needed is shockingly small and the basic processing, surprisingly simple.
Reminds me of google's solution to their "racist" photos app problem: they just blacklisted certain search terms and images.
We have heuristics not laws.
What's the theoretical minimum number of transistors to build a real time hardware MPEG-4 decoder @4k using Intel's 14nm process.
I'd wager less than 1% of businesses outside of SV even have a clue where to begin, or what to use it for, my employer included. "Do we hire some AI guys?"
We'll need to crack the 1%-using-it mark for me to consider the revolution "begun"...
I also think you would be surprised at how many businesses both do, and really want and value, software - but don't have anyone convincing to do it for them
A random example: my SO often works with high-res photos of their products, and sometimes this involves modifying them to suit the random requirements of other businesses. Something I currently solve for her by running ImageMagick one-liners if she needs more than a few photos modified. A person with basic familiarity of shell scripting and CLI tools could improve their workflow and keep improving it on-demand, as the requirements change.
Without that, working as a programmer in such disorganized team, you will mostly spend your time working "against" them. It's like you are trying to glue things together and people will just go and and rip those glued together things apart constantly.
Unless you have some sort of evidence supporting your claim, I guess we'll never know how surprised I would be by the reality you're describing.
Does the AI which can help with literally everything a business does exist? That sounds rather general purpose and extremely open-ended.
Business people make no sense.
Me: "It's finished, here are the areas of strength and weakness, and here's where we can deploy the system for maximum effectiveness."
Business: "We're thinking about the best way to deploy this."
Me: "This is how you deploy it."
Business: "We'll think about and get back to you. Don't do anything until we tell you."
Me: "..."
Business: "..."
2. Data entry. We took a picture of this customer's utility bill / bank statement / receipt / whatever. Now do we give it to human to identify relevant fields and manually type them into a spreadsheet, or do we have a computer automatically extract the business-relevant data? Or, heck, maybe that's too complex, but can we at least have a computer help--automatically filter out bad images, do perspective correction, highlight areas of interest, etc.? (This sort of thing is actually used in, e.g., digitizing census records; we don't trust handwriting OCR to be good enough on its own, but we trust it to automatically highlight relevant fields, in order, and provide a first-draft guess at the transcription to assist the human transcribers).
I could probably come up with a few more if I thought about it for a while, but those are the areas I've actually worked on recently.
My employer does supply chain optimization, because many "average" businesses have supply chains that need to be optimized. We have specialized AI tools based on deep learning and differential programming, and we have human supply chain experts that can fit those tools to the specifics of a customer's situation.
I would say that despite our best marketing efforts, many of our customers don't know that we're doing this "AI" thing they've been hearing about.
The anti-pattern with "Do we hire some AI guys?" is that, if you have to teach those guys how a supply chain works, you've already lost. Pick someone who is already specialized in the problem you're trying to solve.
this article by evan miller does a good job of explaining why: https://www.evanmiller.org/predictive-analytics.html
At current trajectory, we're not headed towards a general intelligence. Progress has been made, but there are big gaps. Smart home devices are a great case in point. They are somewhat flexible in the voice commands they accept. Specific phrasing and pronunciation are not necessarily required. Their responses and speech, however, are all pre-programmed and templated by humans.
Edit: There is potential for more breakthroughs in the future, but I am not seeing them on the horizon at the moment.
* Wavenet, now productionized at Google as text to speech
* Alpha[Go]Zero
* Neural Machine Translation, on production at Google
For more perspective: https://arxiv.org/abs/1801.00631
Machines are the tools we make to aid the above.
1. We don't know all the mechanisms that a brain employs to achieve intelligence. We see billions of interconnected neurons and we assumes "Yea, this might be generating intelligence".
2. We don't know if we are already at some fundamental limits of intelligence. For example, you can see may instances in nature where a pattern emerged that maximizes some sort of efficiency. (Like Honeycomb pattern). So, the end result of this will be that, even if we transfer the process by which our intelligence work to a machine, it will have the same performance as an average human brain...
Maybe throwing more power at the current solutions won't ever make the progress we want, we need to find our car, so to speak. And a lot of people are working on that.
I'm not a researcher either, so it's difficult for me to articulate exactly what problems I see, but (shameless self plug) I did write an article a little while ago attempting to: https://medium.com/@danShumway/modern-ai-techniques-arent-wo...
I tend to be fairly dismissive of inference because what you end up with is a highly specialized algorithm rather than something that can easily continue to adapt. I suspect inference would probably fall into Jordan's category of "things that we call AI but probably shouldn't."
But that's not to dismiss how important fast/cheap inference has been in allowing companies to actually build things with AI.
Of course this is defending a strawman. I don't know anyone that said it was just a computing power issue. In fact I think most people vastly underestimate the role of computing power. Even algorithmic improvements are enabled by computers letting researchers do experiments that would have been impossible before. And by doing lots of experiments they gain intuition about the problem, that they wouldn't develop in a vacuum.
In 50 years people will laugh about how poorly these things were named. Just like we now laugh about what symbol we chose for a source of electrons in a circuit : "+". Whoops.
I mean, all these "structures" and algorithms that people create in software, each a little different than the last, each with different performance/result trade offs. But nobody actually understands the fundamentals as to what is behind some of the very impressive results we are seeing (due to the technology allowing us to throw computing horsepower at it).
the little problem here is that it took 4 billion years of computing and a computer the size of a planet to come up with the nifty machines that are our brains. So unless you have brought a lot of tea and biscuits I think we should really think twice if just throwing more training and power at overly generalised algorithms is a good and realistic path forward.
It's not so much that the idea of "throw more things at it" is impossible, it's that it's a questionable path towards human intelligence. If you just want human intelligence without any further understanding of how to make it there are cheaper ways already
That is not impossible to simulate, even if it includes the entire lineage of the humans. Also, nature prefers generalized algorithms as well, starting from DNA.
I believe 3 orders of magnitude more processing power we would achieve amazing results in a decade; not AGI type of results but very close to it from our perspective.
Part of the reason is that I now believe you can simply "bruteforce" some problems with existing ML algorithms (like thousands of layers deep neural nets) but more importanly, one(not me) could test new ML algorithms that are not feasible now (I don't have any examples) and people would be able to itterate much faster in developing such algorithms.
There are a few areas where performance is good enough to be practically useful. One of those (facial recognition and tracking) is currently causing a “revolution” in China. I’m just not sure it’s the kind of revolution we want.
Our remarkable bodies weren't designed by engineers, they survive now because, earlier, so many didn't. Are we smarter than evolution, because we like to think we are? Time will tell.
I don't know that cars will happen soon(5-10 years), but I wouldn't want to bet one way or the other past that. We don't need perfection, we just need to equal humans, and humans are actually pretty bad, we are just used to it.
Current “AI” is basically function approximation and nothing else. And humans do everything they do in a 20W power envelope.
While I'm unsure whether the on board computer of your autonomous car will be able to leverage past experience, I thought it was a forgone conclusion that the telemetry from all the cars on the road will be used to iteratively improve the core model. Which would then be dispersed as an os upgrade, effectively teaching your individual unit from the past experience from all the units on the road so far.
There are types of neural networks (and other algorithms) that work literally like that. Just because a simple deep perceptron does not work like that does not mean no network does.
This one is just off the top of my head, it's quite recent but the memory part is based on old prior work. https://medium.com/applied-data-science/how-to-build-your-ow...
UPD: without even digging deep in the different types of networks, even AlphaGo(/Zero) works like that.
We rewrite our memories to fit our mental schemas, to the point where someone describing what they saw is more likely wrong than right, even in dramatic ways . (see a stabbing? Did the man in the suit or the man in rags commit the stabbing?). We suffer change blindness, confirmation bias, prejudice. We rationalize and justify to a ridiculous degree, and in the few cases we become aware of this, that awareness does not allow us to change the behaviors. If someone is wrong, the worst way to get them to change their stance is to show them they are wrong.
We're born helpless and spend a strong percentage of our lives learning how to not die. We transfer information inefficiently and inaccurately, with every generation biologically starting from scratch. We spend 1/3 of our lives unconscious (in addition to that helpless period), and almost 2 decades becoming ready to function independently, at which point most people have only a few years before they dedicate an even larger portion of their life to bootstrapping the next generation.
The Turing test exists because we can't even define what we are describing as obvious (and as I mentioned previously, humans fail the turing test often). Almost everyone that drives has been in some form of a car accident, the overwhelming majority of which were caused by human error. We burn plants so we can inhale the (toxic) vapors, we overestimate rare risks and underestimate inevitable ones, we drink poison for fun, and enjoy it because it reduces our thought processes, we gamble money with the intent of winning more money when it is well known the odds of winning are terrible. We entertain ourselves with habits that target innate thinking fallacies and call it "gamification". We ignore issues that we have confidence will arrive, and then react with panic when they do arrive because we've made no preparations. We declare human life to be so precious we don't want to end the potential, even to the extent of stopping people from preventing that potential, but don't take action to support that life once it is born. We look at a list of flaws like this and shrug it off. We oversimplify, stereotype, and categorize even when errors in those systems are pointed out to us. We don't like being wrong SO MUCH we'd often rather continue being wrong than accept that we were. We eat foods that are unhealthy in unhealthy quantities, and produce and purchase foods that directly encourage those habits. We have short attention spans and short (and inaccurate) memories.
Comparing current AI approaches and human thought is apples and oranges, but to mock AI efforts as function approximation ignores how much function approximation we do. We function, and the diversity of tasks we function at is indeed amazing. The complexity and adaptability of the human species is awe-inspiring. But doing amazing things is still not the same as doing them _well_.
I don't say this to claim humans are terrible. I'm pointing out that we are poor judges of quality and that any system following different fundamental restrictions will have different emergent behaviors. I expect that a car that can drive more safely and more consistently than a human is both a complex problem and much easier than most assume. Driving _well_ is harder, but driving better than a human? Not nearly as hard. What percentage of drivers do you think consider themselves to be "above average"?
Driving better than a human from vision alone is extremely hard. Driving better than a human in an area for which you don’t have a 3d capture is extremely hard. Driving better than a human when it’s raining or snowing is extremely hard, etc, etc. Don’t be so eager to discount humans.
Assuming you mean that machines can’t do cognition at present, why do you think we won’t solve this problem in the next 20 years?
Would you care to cite any sources for that? Sounds like something a couch expert would say. Have you talked to every Waymo engineer?
Wake me up when they’re testing L5 in Alaska in winter, using a car with no steering wheel. Then I might consider trusting my life to it.
Sure, it isn't cars in every environment without having seen the territory, but it is an incredibly useful product that can be extended into more cities as the technology improves. And plenty of cities around the world have similar conditions to what is being used in Arizona.
Low speed, AI driven electric buses on limited routes are already all over the place too.
Driverless trucks on certain runs is likely too within a decade.
Of course we are. We went from "natural state" to spaceflight in mere couple thousand years, of which the most meaningful were last 300.
We design, build and test things on a scale that's orders of magnitude shorter than gene-driven biological evolution. Still, that doesn't mean everything gets perfected instantly, and we humans are quite an impatient bunch (no surprise here, given our short lifespans).
Context aware language processing!
When do we want it?
When do we want what?
Please don't go BS like "Google is an AI app" etc. I'd expect an app where AI is the indispensable component. You can make a decent search engine/camera/phone/car/microwave without AI.
https://thenextweb.com/syndication/2018/04/16/drones-will-so...
Plug: I work at a company (https://www.pachyderm.com) whose product is designed precisely to track data provenance across pipelines and through a company's larger data-processing operation for this reason
If you write a blog post describing how critical that is to practical data science efficacy with some examples I bet you'll end up in a bunch of VP inboxes.
Do we have the computing power to come anywhere close to what we need, will we have that computing power any time soon without a major breakthrough?
- Camera that make better looking pictures
- Better looking movies / videogames
- Better medical diagnostics.
- Face recognition
- City surveillance; finding criminals, lost children.
As I recall, the article in WIRED from like 20 years ago, talked about how he missed his dad, and how in the future you would be able to take a room full of all his dad's old crap, and AI would be able to recompile his dad into dad 2.0 in the cloud!
(he didn't use these terms but that was the gist of the ridiculous and absurd WIRED article interview with him.)
Why do I bring up Ray? It's because circus acts like Ray and the Singularity University, sold a bunch of people certain goods, which were VASTLY overestimated in terms of delivery times, and in terms of the good themselves.
First of all, many folks don't want to admit this, but WE MAY NEVER BE ABLE TO CREATE A SYNTHETIC CONSCIOUSNESS! I think we will eventually, but it's definitely not a certainty. It just may not be possible for a completely synthetic 'consciousness' to exist.
This type of hype has become almost religious in nature, just like the folks who really think Moore's Law is an actual scientific law and not just an observation and prediction based on very limited data.
In conclusion, you'll get your AI just like those folks in the 50's, or was it 60's? ...well anyway, back when they had those world's fairs and predicted flying cars.
We're just not there yet. We don't even have basic foundations to build this yet. There is always a possibility of a brilliant individual, who may leapfrog humanity, but for right now, it's just vaporware.
AND YES. Even IBM's Watson and Google's GO winning "AI" are just vaporware machine learning and big data sets, it's not AI even a little bit.
Sorry folks, I love futurizing, and inventing words ;) , just like everyone else, but... if AI is in the stadium, we aren't even in the parking lot.
---
And now, after my rant, I shall read this article since it looks really interesting. I'm surprised it's on Medium.
Long story short, my claim is that the AI revolution will happen when the AI agent will be able to abstract on a task experience and transfer that to a different task AND abstract away that and transfer that to the set of tasks and so on.
Sort of like the Alpha Zero would next be able to fly helicopters or wash dishes or cook a meal or whatever. Not anytime soon.
"How dare you claim IBM's Watson isn't BREAKTHROUGH?!"
"How dare you not want the same buggy useless software which powers siri and cortana to now be used to drive cars?!"
What are you cynical reality based folks talking about! Saudia Arabia granted citizenship to it's first robot. Clearly you hate science! If AI wasn't real yet, why would they do that!?
Lifelike 'Sophia' Robot Granted Citizenship to Saudi Arabia https://www.livescience.com/60815-saudi-arabia-citizen-robot...
Even if you don't believe they are good enough now, to say that they are NEVER going to be good enough is ridiculous.
Of course it will be possible to create synthetic consciousness eventually, why wouldn't it be? It may take 10 years or 10,000 years, but if you think it will never happen for the rest of human existence, that is nonsense. If it already exists in nature, then there is absolutely no reason why it can't be done synthetically.
---
“Watson is a joke,” Chamath Palihapitiya, an influential tech investor who founded the VC firm Social Capital, said on CNBC in May.
A Reality Check for IBM’s AI Ambitions - MIT Technology Review
https://www.technologyreview.com/s/607965/a-reality-check-fo...
Why Everyone Is Hating on IBM Watson—Including the People Who Helped Make It
https://gizmodo.com/why-everyone-is-hating-on-watson-includi...
Is IBM Watson A 'Joke'?
https://www.forbes.com/sites/jasonbloomberg/2017/07/02/is-ib...
---
The final article, is the most telling, in its revelations. Of course, you must speak the language to understand IBM's gartner magic quadrant-esqe based reasoning: Their argument is because a bunch of businesses they convinced to sign up have 'Watson' stamped on their engagement contracts, it's PROOF that Watson is breakthrough technology, catapulting research into a new era of big data and machine learning....lol
I'm not saying some of the things that IBM claims it does, cannot be currently done, but I no doubt at all the IBM is not doing it, and Watson is most certainly behind closed doors a complete joke.
For example, you can't say that DeepMind is vaporware. AlphaGo is real, and it works. Self-driving cars are real, and they work, etc.
"But the episode troubled me, particularly after a back-of-the-envelope calculation convinced me that many thousands of people had gotten that diagnosis that same day worldwide, that many of them had opted for amniocentesis, and that a number of babies had died needlessly."
http://bair.berkeley.edu/blog/?refresh=1
I mean one of their papers on the stuntmen was on the front page of HN recently.
They are most certainly on the forefront.
It's a bit like the early days of the internet, it's not always the most practical solution, but it definitely works.
Unless you have a reason to compare it to the "early days of the internet", don't. A technology being new and unknown doesn't make it like the internet.