Rise of AI Conference 2018 · Berlin

Building Artificial General Intelligence

Chris Boos · CEO, Arago AG

At Rise of AI 2018 in Berlin, Chris Boos, CEO of Arago, argues that current AI is far from strong AI and pushes back on what he calls AI hype. He states plainly that machines do not understand anything, that comparing neural networks to the human brain is misleading given the gap in scale (about a million nodes versus roughly 84 billion neurons) and missing chemical and quantum layers, and that deep learning is only one part of AI, not AI itself. He distinguishes narrow AI, applied to single problems; science-fiction-style strong AI, which he says no one seriously building AI knows how to start; and generalized AI, applicable to more than one problem with the same engine. He describes three underlying algorithm types, machine learning, natural language processing and machine reasoning, tying each to an evolutionary analogue, and argues companies should first explicitly communicate the knowledge they already have to a machine before relying on costly trial-and-error learning, since data by itself can never fully explain the world.

Key takeaways

  • Boos argues comparing neural networks to the brain is misleading: about a million nodes versus roughly 84 billion neurons, missing chemical and quantum layers.
  • He distinguishes narrow AI (single-problem), science-fiction strong AI (no credible path yet), and generalized AI (applicable across problems with one engine).
  • He estimates natural language processing systems are at about 1 percent of what real human dialogue requires.
  • Using Arago's Civilization-playing AI, he shows machine-learned site-selection logic mirrored a real-estate customer's decision process, illustrating unrealized knowledge-transfer potential.
  • Boos argues companies should first explicitly communicate existing knowledge to a machine, saying this has roughly a 95 percent energy advantage over pure trial-and-error learning.

Detailed notes

Deep learning is not AI

Boos opens by rejecting several claims common in AI marketing. He states that machines do not understand anything: a machine compares incoming data to a pre-created structure, and expecting machine "understanding" to behave like human understanding leads to wrong decisions. He also rejects the framing that neural networks are "modeled after the human brain," pointing out that a large neural network has about a million nodes against roughly 84 billion neurons in an adult human brain, that a neural network needs about half a nuclear power plant to run versus 20 watts for a brain, and that the brain is also a chemical and, per some neuroscience research, possibly quantum system that current AI does not replicate. Finally, he states deep learning is only part of AI, and warns that treating it as the whole field risks a new AI winter.

Three types of AI

Boos describes three categories. Narrow AI applies one algorithm or small set of algorithms to a single niche problem; his example is a company that trained a neural network to increase tomato yield in greenhouses by 20 percent, which fails if applied to a different crop in the same greenhouse. Science-fiction AI, the "strong" version capable of something like genuine love, has, in his view, no credible development path yet, leaving ethics squarely a human responsibility. Generalized AI sits between the two: applicable to more than one problem using the same type of engine and data, valuable because AI's real purpose, he argues, is not just insight but execution and automation, which industrialization already made possible by breaking work into explainable pieces.

The three core algorithms and their limits

Boos frames machine learning as automated pattern recognition tied to a positive action, comparable to animal instinct, and says it has a natural data limit: there is never enough data to fully describe the present. Natural language processing, which he calls the compressed output of evolution, lets humans transmit experience without trial and error, but he estimates current systems are at about 1 percent of what real dialogue requires. Machine reasoning, including case-based reasoning, lets systems weigh future options the way humans do mentally, but maintaining its rule bases is costly, and he says once maintenance costs exceed the benefit outside a narrow domain, it should not be used.

Knowledge transfer and Moravec's paradox

Boos names three central AGI problems: Moravec's paradox, that tasks easy for humans (evolved over billions of years) are hard for machines and vice versa; "little data," since humans make confident decisions from too little information while AI confidence is mathematically dependent on sufficient data; and knowledge transfer, illustrated by an example where Arago's AI learned city-placement logic playing Civilization that mirrored a real-estate customer's site-selection process, showing transferable patterns exist but current data is not connected enough to transfer automatically.

Why companies should teach machines what they already know

Boos argues data should be owned by whoever produces it because it is a tradable asset used to explain the world to machines, not only for predictive analytics, which he calls a rebranding of decades-old linear regression. He contrasts two ways of teaching a machine: observation and trial and error, which is data-heavy, versus direct communication of existing knowledge. He says he is surprised most companies default to the trial-and-error route first, since companies exist because they already know something; explicitly transferring that known knowledge before learning the rest by observation, he says, has roughly a 95 percent energy advantage over machine learning alone.

Quotes

Machines do not understand anything, like zero. There's no understanding in a machine.

Chris Boos

If you want to build an AI, anyone can do this, you need algorithms, data, you need talent, and you need compute power.

Chris Boos

Questions answered in this talk

Is deep learning the same thing as AI?

According to Chris Boos at Rise of AI 2018, deep learning is only part of AI, not the whole field, and treating it as a single universal solution risks causing a new AI winter.

What are the three types of AI Chris Boos distinguishes?

Chris Boos describes narrow AI, applied to one problem at a time; science-fiction-style strong AI, which he says has no credible development path yet; and generalized AI, applicable to multiple problems with the same engine and data.

Why does Chris Boos say companies should communicate their existing knowledge to a machine instead of relying only on machine learning?

Boos argues at Rise of AI 2018 that companies already possess valuable knowledge, and explicitly transferring it to a machine before relying on trial-and-error learning has roughly a 95 percent energy advantage over machine learning alone.

Mentioned

deep learning, narrow AI, generalized AI, machine reasoning, natural language processing, AI winter, Moravec's paradox, knowledge transfer, case-based reasoning

Keep exploring

More from Chris Boos

More from Rise of AI Conference 2018

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 4,074 words

Show the full transcript

so I'm gonna have a fun talk because Peter said all the important things that I can say all the fun things generally I maybe not used strong AI as a word is a question of where are we somewhere between these fairly dumb narrow approaches that are great at optimizing and we're and the very strong AI that we just heard and I could not agree more simply does not exist and won't exist for quite a while to start this out and you've probably hear this from everybody who's been doing AI for a while it's kind of annoying in the AI hype all the stuff that is being said about AI so let's clean up with some of the hype stuff I'm gonna be more aggressive Peter was very polite I'm gonna tell you machines do not understand anything like zero there's no understanding in a machine what a machine can really do is compare incoming data to a pre created understanding mechanism or structure the problem with people pushing that into marketing and you know exactly

that the marketing record is on when somebody tells you my machine understands blah blah blah the problem with that is that we as human acumen --zz actually do have understanding and then we draw conclusions from that if the Machine understands something it should also know this and that and the other and we expect the machine to do things that it just will not do so this is one of the most misleading things that makes people take completely wrong decisions and not just have wrong expectations and be disappointed but really go down the wrong path the second thing that is out there and I find personally very amusing is that all of a sudden everything is modeled after the human brain right just because we have neural networks and we named them that that way we call this his model after the human brains guys seriously if you ever say an AI is modeled after the human brain please also go out to your former biology teacher and tell them that the amoeba is modeled after

the human body because it's like basically the same thing a big neural network has about a million nodes the brain of an average person after puberty has about 84 billion nodes the the the rural that Newark needs about half a nuclear power plant to run and we can do it on 20 watts okay if you look at me I don't get that but most of you do right so also there's a chemical system if you go to the after party of this event try it out go to the bar have a few beers and see what it does to your brain there is with every chemical combination and there is an infinite number of chemical country configurations your brain gets a new configuration there's a whole dimension we're missing so there's definitely no where were so much at the beginning of this that we should not talk about it and then for about five years there's research from the neuro scientists out there that say that is probably also some quantum mechanics going on in

the brain meaning something that injects probability so even if we can build a neural network as more low halls in 2029 that has the size of the brain we're still missing the chemical system and the quantum mechanical system we are truly far away from the brain and an amoeba does not function like the human body either and the third thing and I think this was already out there a few times this morning deep learning is not AI well deep learning is part of AI but there are more algorithms out there and by pretending that there is a single algorithm out there it's not just a bad way of solving problems it also means that we're definitely in for a new AI winter if we believe that we can solve everything with one problem I mean try that with your life try solving everything with a problem someone is going to be very pissed at you either your boss or your girlfriend right so there are currently in perception three types of AI right one being

the narrow eyes this is the stuff where someone takes typically a single algorithm or a small set of algorithms and applies them to a niche small problem these are great right because they achieve great things they really optimize these optimization algorithms are way better than what we had before from statistics because they can live with more free variables which means in you have to explain less to them or they can deal with more complex problems but they deal with one problem at a time and my favorite narrow AI I keep telling this because I love it so much is a company that trained a neural network to improve tomato growth in greenhouses you connect your fertilizer you water your light proof you have on the train network twenty percent more tomatoes in your greenhouse if you happen to live in a place where there's not so much Sun and the tomatoes don't taste that great don't have this stupid idea to start growing beans in the same greenhouse because the AI is gonna kill all

the beans because them stupid things don't turn red right this is the problem and everybody knows this if you apply one solution that works on one problem really well and take it somewhere else it probably screws up on the other but I have to say no one should be afraid of using those narrow AI systems why not because all of us have been hunting for efficiencies in our personal lives in our companies for a long time and why not apply them there's no black magic to this this is not black mirrors this is just optimization on the other hand of the system we have the science fictiony ice this is probably what the strong guy a robot and I'm gonna go down the positive lane here a robot that actually says I love you and knows what that means and actually means it I can promise to you no one who is seriously building artificial intelligence studying that has the slightest effing clue how to start that project this is good and bad news the

good news is that ethical morality just stays with us humans and the bad news is it stays with us humans we actually have to deal with it there is no thing out there that's gonna take the question of ethics for us the whole debate about ethics is just a question that we we we want to push off being good people to a few programmers or a strange being that does not exist there is no digital being yet in the middle somewhere you have more generalized AI I chose general and it's always difficult to to these terms because it's definitely not strong but it's applicable it this basically means it's applicable to more than one problem with the same type of engine and the same type of data why would you want those well because even in today's world a eyes are not just about gaining insights from data a lot of people think in AI is good to find some new insight that's not in what's supposed to do it's supposed to act is supposed

to execute and especially in our old economy it makes a lot of sense to it to act because you can automate a lot of processes why can you automate a lot of processes because it's basically impossible to explain the entire world to an AI otherwise we would have this strong AI it is impossible so we can only explain small pieces to an AI and then actually let them run them well what is it that we have done in companies right we have terrorized the things to do because we believe in an industrial nation know that no person can actually understand the end product and we've put everything through processes we have nicely compartmentalized everything that we do and this is why industrialization as a precursor to AI means that we have finally cut down the economic system in a way where X X X of accessible to machines now we're going to implement it we shouldn't be so surprised if you want to build an AI like anyone can do this as I said no

black magic you need algorithms data you need talent that's kind of important yeah and you compute power well compute power now comes out of the socket and you just get it from Amazon Google whoever algorithms anybody can download ail Rhodes they're just available they're fantastic algorithms out there you don't have to find your new one there are great ones out there data a lot of people have data most people have a total misconception what to do with data and then you have the problem of people because AI building AI typically is not about one single speciality that people who do this need a broad spectrum and that's where it gets tricky because that means that you either need a large team of a lot of people with a lot of different skills or you need people who are fairly experienced and both are hard to build if you look at these single technologies and I speak three and I'm gonna explain these three each one of them singularity has a problem right so for example

machine learning which we see in most narrower eyes today has a simple problem with just data there is not enough data to explain the world to a machine and if you don't believe this think about it if you need a 10 year data set to teach something to a machine you've now taught the average of these 10 years to the machine the now might be very different there is never ever enough data to describe now this is why pure machine learning has a very natural limit then you have natural language processing we see this in all the chat BOTS and so on language is kind of the end product of evolution like 4 billion years of testing and trial and error went into this and all of the natural language stuff you see and I absolutely love the Siri video it just shows you it's a simulation still it's a total simulation it works super well when the machine knows what you're trying to do if you have a number where people call to call

in sick and they use different words like I'm sick I'm under the weather I'm not feeling well and so yeah this works but if you want to have a general conversation this does not work at all so there is a total boundary to this and then you have the machine reasoning including case based reasoning and so on this is a fantastic system because we're able to plan into the future and then way optionality against each other the problem is maintaining these rule base --is is a tough job and the tough job means that you actually have to live around with a lot of effort to doing it and then oftentimes if you are trying to be walk out of a narrow space with an expert system you are most likely going to out perform maintenance as a compared to benefit so there's more maintenance than benefit and then you don't do it let's have a look about how how these algorithms roughly work because everybody is talking about it and I feel that that a

lot of people don't know what's inside of it and and maybe we can have a little chat about this so number one is if you have lots of incoming data from sensory organs like cameras or lighters or eyes and ears and noses and you put them in a processor and the job of that processor is to find a pattern in this data and then associate a positive action with that this is basically what we do in machine learning and this is also in the biola a world of biology what we call instincts right and in in an animal we would say if it sees a tiger and runs away that's a very good it has recognized the pattern tiger and it knows that it has to run away why because all its companions that did not have this reflex of running away were eaten that is back propagation in nature basically these are rules right if tiger then run but these rules can be encoded and compressed with us it's very simple right I can

explain to you how math works I definitely can't explain to you how I can make my heart beat because that went from the brain to the nervous system to the nerves around the heart right the longer you live and the longer you process these things the more they get ingrained into your system and this is exactly what happens with neural networks basically here you have machine learning written down as finding patterns and data and associating a positive action with that and then compressing it into some mathematical structure now the best of these systems they are able to conceptualize is always about outpacing the evolution outpacing this isn't purely evolutionary process right everything that doesn't go for the survival function dies outpaced the evolution for example with concepts concepts means that if you have associated the action of eating with the pattern of Haring and then you say like hoorah I've got a lot of things that look like Haring let's associate eating with all of them and then all of a sudden you only have

the exceptions whereas don't eat the Japanese round fish because you die same mechanism but as there are more edible fish than poisonous fish you have now outpaced evolution because you didn't have to go through the 70,000 or whatever repetitions to learn which fish you can eat for each fish you made an assumption a lot of higher animals have done this what we as humans have done something fantastic we have created language we are able to put all that experience that we gathered into words and pass it on to someone else and they don't have to trial and error it at all or have children in order to propagate that knowledge we can actually just tell someone okay for those of you with children you probably know that just telling someone normally doesn't help but the amount of repetition goes down dramatically so when you tell your child is don't touch that it's hard but it doesn't touch it seventy thousand times one will do right so um this is what we can do with language

and it's quite amazing and the knowledge what has not been so compressed that we can't express it anymore all of this expressible knowledge the conscious knowledge that we can symbolize through language is accessible to us and we can pass it on and we can we have outpaced evolution dramatically this is what natural language processing is supposed to do and if you want to like my personal estimation we're about at 1% like we're nowhere close to what is actually possible in dialogue and so on in technique then I said this is what makes humans different like the homo species as compared to other animals on the planet then there's the third type of algorithm which we said was reasoning if we want to say this nicely we have outpaced evolution again by having fantasy or imagination if you want to say that a little more scientifically we are able to plan the future in our heads and weigh options against each other and make a good choice we actually do this in our heads everybody has

about 700 words per minute in their heads making constantly an argument of what to do and what not to do geniuses have a little more other people might have a little less but basically that it is and now you can inject goals that are beyond survival inject them meaning you don't set them yourself like this does not include consciousness at all this does not include self-awareness or anything like this but we've now explained the ability to learn by trial and error the ability to communicate something that you have learned and the ability to weigh in a special situation something that you've learned against each other these are three algorithms I hope we can agree that if you want to be a little more general than just applying like one thing to one solution which you can do well you probably need these three type of algorithms here's what we can do today in terms of of more generalizing you can use machine reasoning to compartmentalize the space because people give you knowledge but let you're

in unfortunately people that give you knowledge typically will not give you the full explanation of logic like as a computer scientist you want a precondition a post condition provable space right this does not exist in people people will assume that there is context and according to this like it's actually weird that we don't have more fights is because we are so good in interpreting context we basically use these type of algorithms to compartmentalize the space and reduce the amount of data you need for machine learning problems and then you use the learning to deal with ambiguity or unexpected events and stuff like that you can use a chat bot to extract knowledge from people so you avoid having to learn everything just by observation and you can use a semantic database to store all the experience that you already had so that your next step in reasoning on your next step in machine learning has it easier and this is the actual point the economic effect as compared of a narrow AI application and a

more general AI application is that the amount of time you need to put the machine to a new problem basically goes down logarithmically it's potentially like a reverse exponentially this is the actual application we cannot go further than that today but that's pretty good right so if your first project takes three months and the next one takes one and a half months and the next one takes and so on right you can actually get going pretty quickly on different problems but it's still exactly what Peter just said it is a single problem at a time approach and because we can refer to all the stuff that is already be in the system from the other problems we just make that like making these problems accessible and automatable faster so I think that we have a lot of questions about these ai's we have to deal with with three big problems that they eyes totally have three big problems the problem is not finding the right data set and so on there are three problems that

you have to think about when you're thinking about generally I one is more of X paradox meaning that it is easy whatever is easy for a human is hard for a machine and whatever is hard for a human is probably easy for a machine this is a simple evolutionary fact because whatever is easy to us we've been practicing for four billion years even the crocodile could walk right and whatever is hard for us like let's say non Euclidean geometry we've only known this for seventy years most people get a headache from that this time is easier to catch up for a machine this is a true problem because we expect machines to do stuff that is easy for us as well the second question is little data as people we're constantly making decisions actually pretty good decisions with way too little information that's what a eyes have to do as well and they have to feel confident about it which is very hard to do because in the end confidence is math and if you

simply don't have enough things to fill into your equations how can you solve them and the third thing that would really make a general AI take a big leap is knowledge transfer if you can learn something in a game and and you know that with our AI we for example played civilization we've the machine has learned a process of how to choose the spot for building we can post data analyze we can see that the decision process that is taken is exactly the same process that a real estate customer of ours has taught the machine to which objects of real estate to look into to make the decision but the data is far from being connected enough so this knowledge could transfer automatically yeah so we can see that this is possible in the future but it's definitely not there yet data is important and you have to pay attention that your data stays with you I'm actually a favorite of those people that go with the EU and say we need data protection but

we need data protection because data is a tradable asset because it's a way to explain the world to machines the problem that is not part of the data is the oil of the 21st century religion that is being pushed around is that data is a tradable asset and you need to share and trade it because a company like BMW has all the data that the company likes Calvin where it needs to build a better screw it has very little data likes to zero on how to know what a car to do next this is why it makes total sense if you produce data it should be yours because you can trade it in the future data ladies and gentlemen is not just there to do predictive analytics actually predictive analytics is just the fancy name for linear regression it was people were too embarrassed to actually say like we're doing these things that we know for 30 years right so they've renamed it data is there to explain the world to a machine and in

order to do that you need structured data and if you want to extract structure for non structure you need a lot of data and there's two ways to do this right you can collect the data by sitting on the shoulder of a consumer and watch whatever they doing you get a very superficial picture but you get a picture of the entire world or you can sit inside industry you get a very deep picture but a very narrow one because each company only does a very narrow thing this is why data needs to be interchanged there's two ways of teaching a machine there's two ways of getting knowledge into machine you can have observation and trial and error which is the data heavier approach or you have communication with the Machine you explicitly give exact an exact whatever information to the machine I'm very surprised that most companies try the trial and error and observation route for the first time because most companies exist because they already know something why can't they communicate what they already

know to a machine first transferring knowledge you already have and then learning what you don't have or don't express by trial and error and observation has an energy advantage of 95 percent like machine learning actually is not very energy efficient if we think about AI we shouldn't think about insights into data we should think about our experience if a machine looks at us and tries to replicate what we do or if we teach something explicit to a machine we make a machine use our experience maybe it can recombine the experience in different ways making our experience executable it's about execution it's not just about insights thank you very much

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.