Rise of AI Conference 2019 · Berlin

Towards a General Theory of Intelligence

Peter Morgan · CEO, Deep Learning Partnership

At Rise of AI 2019 in Berlin, Peter Morgan, CEO of Deep Learning Partnership, argues that deep learning cannot lead to artificial general intelligence and proposes a physics-based alternative. He defines intelligence as an agent's adaptation to its environment and argues there are many distinct types of intelligence — physical, spatial, musical, linguistic, social, emotional — not just the logical-mathematical kind that deep learning approximates. He traces how biology builds intelligence hierarchically, from chemical-gradient navigation in bacteria without neurons, through the roughly 130-neuron nervous system of C. elegans and the million-neuron bumblebee, up to the human brain's roughly 100 billion neurons organized into repeating cortical columns. Comparing digital hardware such as Google's TPU and the Summit supercomputer to neuromorphic systems like SpiNNaker and BrainScaleS, he argues that brain-like hardware that combines memory and processing is a necessary step toward AGI, since digitally simulating the brain is far less efficient than biology itself. He proposes grounding a general theory of intelligence in physics, specifically the principle of least action, and highlights active inference, tied to Karl Friston's free energy principle, as the theoretical approach he finds most promising among several decades-old research programs.

Key takeaways

  • Morgan defines intelligence as an agent's adaptation to and prediction of its environment, spanning at least nine distinct types beyond the logical-mathematical.
  • He argues deep learning is statistics and linear algebra that can never produce general intelligence, only incremental accuracy gains.
  • Biology builds intelligence hierarchically, from neuron-free bacteria through 130-neuron C. elegans and million-neuron bumblebees to the roughly 100-billion-neuron human brain.
  • He argues neuromorphic hardware such as SpiNNaker and BrainScaleS, which combines memory and processing like the brain, is a necessary step toward AGI.
  • He proposes grounding a general theory of intelligence in the physics principle of least action, favoring active inference as the most promising specific framework.

Detailed notes

Redefining intelligence as environmental adaptation

Morgan opens by arguing that to build artificial general intelligence (AGI) it is first necessary to define what intelligence is. His definition: intelligence is an agent's adaptation to, and successful prediction of, its environment. He argues this is not one faculty but many — logical-mathematical (Einstein), physical (a high jumper who could out-perform Einstein at the high jump), musical (Mozart, Beethoven), linguistic (Shakespeare), interpersonal, emotional, and social intelligence up to the level of nations keeping peace. He contrasts this with narrow systems such as AlphaGo, AlphaStar, Deep Blue and IBM Watson, which he calls statistically brilliant but "completely dumb" and not general.

Why deep learning cannot reach AGI

Morgan states that deep learning is fundamentally statistics and linear algebra — summing weighted vectors through large matrices — and that this can never produce introspection or genuine understanding, only ever-smaller accuracy gains. He cites BERT (from Google) and GPT-2 (from OpenAI) as impressive statistical achievements, noting that Facebook, Microsoft and other big tech companies are likewise pushing ahead with such statistical methods, but says progress on general intelligence via these methods is effectively zero. He notes that Geoff Hinton and Yann LeCun never intended deep learning to imitate the brain, and reports that Hinton has said his own deep learning work now needs to be set aside to reach the next level by studying how biology actually does it.

How biology builds intelligence, from bacteria to the brain

Morgan describes intelligence as hierarchical in biology: bacteria have no neurons at all yet navigate using chemical gradients; C. elegans has about 130 neurons, the simplest nervous system that can still survive, eat and reproduce; bumblebees have about a million neurons and can navigate and perform their waggle dance; the human brain has on the order of 100 billion neurons with roughly 1,000 to 10,000 synapses each, organized into repeating cortical columns of about two million neurons that show generality (visual cortex can be repurposed for sound processing in blind people). He argues any general theory of intelligence must explain this whole hierarchy, from atoms and neurons up to societies and nations.

Neuromorphic hardware as the missing piece

Morgan contrasts von Neumann digital hardware, which separates processing from memory, with neuromorphic hardware that combines both, as biology does. He cites Google's TPU and the Summit supercomputer (three exaflops, GPU- and AMD-processor-based) as extremely powerful but "dumb as a brick" at general intelligence. He points to SpiNNaker, part of the Human Brain Project at the University of Manchester under Steve Furber, which he says has built about a billion artificial neurons across 48 chips per board, 24 boards per rack and 10 racks, and to BrainScaleS at the University of Heidelberg. He states digital simulation of one second of brain activity took the Human Brain Project about a week on a full data center, whereas neuromorphic hardware runs about a thousand times less efficiently than biology, versus about a million times for conventional digital chips.

A physics-based theory: least action and active inference

Morgan proposes that intelligence, like all physical systems, should be describable through the principle of least action, the single underlying formula (in Lagrangian or Hamiltonian form) he says unifies Newtonian mechanics, relativity, thermodynamics and quantum mechanics. He surveys several decades-old theoretical programs and singles out active inference, associated with a theoretical neuroscience group and its free energy minimization approach he links to Karl Friston, as the framework closest to a genuine physics of intelligence. He concludes that the data, theory, algorithms and (nascent) hardware needed to build AGI already exist, but estimates it could take five to ten years as the hardware matures.

Q&A highlights

Asked how society can bring together the computer science, physics, neuroscience and psychology needed for AGI when academic incentives push toward narrow specialization, Morgan points to Karl Friston's theoretical neuroscience group, which mixes strong mathematicians and computer scientists with neuroscientists, as an early case study, and adds that psychologists are still needed. He argues that while many disciplines must contribute ideas, the actual mathematical theory will ultimately have to be written down by one or a few very strong scientists, comparing this to Einstein deriving general relativity.

Quotes

It takes a village to create an AGI, just as it takes a village to raise a child.

Peter Morgan

We do all that in three pounds. We don't need three football fields or ten megawatts, we need a couple of sandwiches, an apple and a drink to discover the laws of the universe.

Peter Morgan

It does not give us a neuron, absolutely not, even close, but never will do.

Peter Morgan

Questions answered in this talk

Can deep learning lead to artificial general intelligence?

According to Peter Morgan at Rise of AI 2019, no: deep learning is a statistical, linear-algebra method good at pattern recognition, but progress toward general, introspective intelligence through these methods is effectively zero and always will be.

What is neuromorphic computing and how does it relate to AGI?

Peter Morgan explains at Rise of AI 2019 that neuromorphic computing, exemplified by the SpiNNaker project at the University of Manchester and BrainScaleS at Heidelberg, replicates the brain in hardware by combining memory and processing, running far more efficiently than simulating the brain on conventional digital chips.

What theoretical framework does Peter Morgan propose for building AGI?

At Rise of AI 2019, Peter Morgan proposes grounding a general theory of intelligence in physics via the principle of least action, and highlights active inference, linked to Karl Friston's theoretical neuroscience group, as the approach closest to that goal.

Mentioned

Artificial General Intelligence, deep learning, principle of least action, active inference, free energy principle, neuromorphic computing, SpiNNaker, BrainScaleS, von Neumann architecture, cortical columns

Keep exploring

More from Rise of AI Conference 2019

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 6,196 words

Show the full transcript

hi everybody I'm Peter Morgan so I rips in a company called deep learning partnership which is an AI consulting company so it's plain old deep learning for my bread and butter but today I'm going to talk outside of that into what I see as the future so here's a little book let I wrote for a riley it's free to download so go ahead machine learning is changing the rules there's my company website there's my Twitter handle I tweet on sort of what I'm going to be talking about today a lot so you can sort of keep up to date if you if you so desire so what am I going to talk about so what is intelligence we a good place to start right if we trying to solve it we need to ask what it is figure it out exactly what it is we're trying to solve so I'll start by outlining what I think intelligence is okay I'll answer that question today and then we'll look at you know physical systems biological non-biological

so the non-biological ones being computers silicon that type of thing biology biological systems being us right we we are the you know we are what we are trying to model okay this is it we're trying to model the human brain and not only the human brain but you know how biology does intelligence how biology does intelligence okay that's that's the aim of my talk is to describe that and a system that might actually build it in then we'll do a quick recap a very quick recap on deep learning because I've just tried to convey that you know biology isn't actually deep learning at all in a lot of the talks today particularly the neuroscience talk I saw earlier before lunch made that very clear we're not trying learning won't get us to intelligence right and I'll explain exactly why well basically it's statistical and what we want is a physical theory of intelligence right so I'll go more into the physics in the neuroscience finally we'll have a look at a theory of AGI and

I'll talk about that so that's the very end of the talk okay so without further ado if yeah so why are we okay hopefully the clicker will work so why do I you know why do we want to build age I well we want to solve intelligence we want to understand intelligence so we want to build systems that you know we can use to do all these tasks and then make money from that or just do Li you know automate all all the boring stuff and ultimately if its general intelligence we can automate everything so then we got to ask you know philosophically philosophically what do we do right if we build these systems and my answer to that is we'll just get back to being more human right we don't we won't have to work in factories and production lines and you know even though we a lot of people have convinced themselves you know this is what the true meaning of life is I don't I don't buy it so once we build

it it'll become more obvious what what would have to do with ourselves okay nothing to get too worried about okay I'm not a doomsayer I'm not here to say oh my god you know the world's ending but you know the conversation at some point always goes there so I sort of just get that out the way at the beginning so let's focus on trying to build it for the rest of the talk so what does general intelligence are any of these general intelligent alphago alpha star playing Starcraft from deepmind you know I've come over from London where deepmind of Base the Google company deep-blue from IBM back in the late 90s IBM Watson winning at jeopardy natural language processing none of those are even close to Journal and that they do very specific things they're completely dumb they're completely stupid they use statistical methods so no they're super clever so they can totally outperform us just like a calculator output with multiplication right so not dismissing it entirely but it's not general intelligence so it's

not what I'm here to talk about that's what deep mind does it's not what I'm trying to do although ultimately they do want to do that and they've they've said that very very very clearly but right now they're not doing it so what is intelligence well here's the answer it's not one thing it's many things so it's basically we are agents in our environment so it's our adaptation to our environment if we were living in a different universe as different laws of physics so gravity goes up instead of down we would be adapted to that and so that would be our intelligence would be adapted to that environment okay so it's simply adapting to our environment exploiting our environment modeling our environment in predicting successfully predicting what's going to happen next in our environment if you think about it it just kind of boils down is you know deflate everything right now it just kind of boils down to that whether we're trying to win the lottery or build a rocket to the moon or

even build Journal Intelligence we're just trying to model our environment it just so happens that we're trying to model our brain okay with general intelligence so we've done a pretty good job you know we've we've done you know the laws of physics we've had Newton we've had Einstein we've had fine men we've had some super clever computer scientists biologists neuroscientists but it's all the same thing we're just trying to model just trying to model our environment okay so it's not just Einstein that's not the only thing that has to do with intelligence it's spatially you know we have athletes you know that are way better than Einstein at doing the high jump right so he was crap at that but very good at modeling general relativity but when it came to hurling himself over a high bar at seven feet you know he would just fail miserably when the high jumper goes to solve you know that understand the universe right now mathematically he would fail so there's not one type of intelligence okay there's

many and that's what the G and a GI stands for it's general okay so we're trying to build things that are spatially intelligent aware of the environment we're not doing a very good job there are we with climate change and everything like that Donald Trump blah blah blah right so you know that that's a for me as a big part of intelligent is to understanding and not destroying our environment but you know kind of living in harmony that might be a definition naturalist intelligence not not getting too political but you know that's what I believe musical you know Mozart Beethoven the Beatles whatever type of music you like and I've managed to turn that off there is a type musical intelligence not everyone can create write a beautiful melody or a symphony but some people can write so these are all the different types of intelligence linguistic you know Shakespeare or you know who can write a great poem etc etc interpersonal people relationships how we relate to one another social intelligence emotional intelligence introspection

how we understand ourselves you know how we react in society how we go about you know a place in the world and then you build it up hierarchically in two nations you know do we go to war do we not I mean how do we keep peace internationally these are all different types of intelligence that these machines will have to do if they're truly general intelligence okay so it's not just the building Einstein you can see that now it's everything it's physical its political at social its emotional okay a big ask a big ask but we do that in three pounds in our skull right so biology has done it using a hundred billion neurons roughly somebody's hit 86 billion but it's of that order of magnitude and perhaps it's the connections a thousand to ten thousand synapses per neurons so it's a very complex system so it's no wonder you know we can't give ourselves too hard a time that we haven't built it yet because these you know it's complex in also you

know the Earth's four billion years old we've had a few years okay we didn't just come up with this stuff yesterday you know we can go back ten thousand years twenty thousand years we were pretty much living in caves right it's only the last hundred or so years that we've kind of you know hit this exponential where we're taking off and the singularity is and all that kind of stuff right so we're living in very interesting times but this is where intelligence has brought us and this is where we are we're sitting in his conference talking about it right we wouldn't have been doing this maybe ten even ten years ago okay so we're ready to build it how far have we come I would argue not very far we've all these talks you've seen today have all have been about deep learning and statistical methods the best we've done I think is around 50% for logical-mathematical linguistic the last talk was on Bert super impressive stuff there's GPT two of open AI there's Bert

with Google you know Facebook Microsoft everyone all the big tech companies are plowing ahead with these statistical methods but nothing to do with general intelligence so you know statistics will only get us so far right and in fact when it comes to introspection thinking about the universe they're actually zero they're not even off the ground okay they haven't even started haven't even begun and they never will do because it's statistics won't won't do it physics or statistics right I'm not trying to discredit anyone here I'm just trying to focus on general intelligence okay okay so how will we get there well it takes a village to create an AGI just as it takes a village to raise a child right it's going to take a village to build these things so it's not just physicists it's not just computer scientists it's not just you know psychologists it's all of us together so it's neuro scientists psychologists physicists computer scientists everybody politicians because there's general intelligence we need everybody in the room together okay so what

you know so physically you know biology biologically and non-biologically let's have a take a look at you know how how physically intelligence manifests okay so biology very clever we start with bacteria no no central nervous system no neurons nothing and yet they survive they reproduce they're intelligent okay maybe more intelligent than us maybe they'll survive longer than us so at the rate we're going I'm not putting my bits right now so you know that is intelligence without even a new or a single neuron anywhere to be seen so they're using chemical gradients you're just using the laws of physics to wiggle and wobble about but they've learned how to reproduce create DNA and reproduce they haven't learned anything right that that's a that's a em from amorphous in yeah this is this is what the laws of physics can to in has done you know in terms of biology the the first thing with a central nervous system is to C elegans I think it has a hundred and thirty-odd neurons so arguably this is

the dumbest thing with the central nervous system so but yet it can you know understand it can survive it eats it reproduces the bumblebee has about a million neurons this is how biology does things it can navigate it does its wiggle dance it's it's societal that lives in a society and right up to us the brain the human brain in fact you know you've got all the memo elephants fish chimpanzees us right so there's a whole spectrum here of intelligence you know assuming that assumably will start with the you know simplest systems and build up but is there something here that unites these all of these systems here is you know because ii i i'm interested in understanding theoretically what intelligence is and in other words you know writing down the mathematical equations that we will use to guide us when we build these things okay not not just sort of trying does this work does this work alright another neuron another layer to my deep neural net where i want to understand the basic

physical the physics of intelligence you know we do as a community at the AGI community okay so the first thing we notice is that biology is hierarchical we start off with atoms they're certainly not intelligent molecules neurons they're not intelligent under themselves but when you start connecting these things together the connectome then intelligence emerges from there and then we have us you know agents and then we have societies and in nations and finally the world so you know the whole thing is hierarchical from atoms up to nations so this general theory will have to explain the whole lot okay so it's it's a big ask right we're actually trying to describe the whole thing and not not just bits of it okay so just to kind of sort of expand your mind put it all in perspective so we're coming up with a truly general theory of intelligence and that's what a neuron looks like it's super complex do we need to go to this level you know there's a lot of open questions we

don't know we'll see okay we'll just see nature does clearly because that's what a neuron looks like but when we build these systems what we have to it's a very very very interesting open question right now okay but it won't stop us perhaps coming up with the underlying theory so this is what a brain looks like this is what intelligence looks like it's organized into structures repeating structures of about two million neurons and these cortical so-called cortical columns again and over and over and over again so if somebody goes blind god forbid that some of those neurons that we use to process visual information quickly repurposed into processing sound okay so they're kind of general is a generality they're these are just information process just these are information processing systems that seem to have some sort of generality so we do know that that's something we do know so far and there's that I'm going up the hierarchy here here's the connectome so it's all connected besides these tens of tens thousands of synapses per

between each neuron these things are organized into some larger if you go up one level and as hierarchy then you start to see this connect home for me whether that's in humans or even bees you know there is as layer above the neuron set maybe that's a right layer to start with okay and then above that then that's attached to a central nervous system and in all all be all living beings a fish should be a human so you know we know it's not just a brain in a jar although you know some might argue Stephen Hawking was kind of that because he had terrible disease and there are people so we can function like that but we're really interested in building the whole physical as well but the brain is the main thing once we've cracked that we've we've cracked intelligence I'm just saying you know we connect it to the rest of our system we have or do we need kidneys when we but I doubt it I doubt we'll have this thing

a system with kidneys in lungs and everything else right but that's how biology has done it in every single thing so there's this other part 90% you know which I think we need to worry about the brain more than the rest of it but the rest of it super clever but kind of clunky it's just nature's put us together like that we may not need to go down to that consider that so much that if we're building robotics we will but robot might not have kidneys but all ahead you know it needs to have be able to move around fluidly and we haven't quite got that yet even though Atlas it does nice backflips and super I mean it's getting pretty good asthma every month you know Boston Dynamics comes out with a new video it's like oh my god if it's like so human and then they up they beat themselves but it's not it's not intelligent you know it can't play chess it can't you know ponder the universe its own existence it

doesn't even know it's alive that's how dumb it is but it's very good with the physical all right okay that's all part of intelligence and then the social systems we talk about cloud robotics form robotics social intelligence that that's that's the final layer okay so here's what here's where we are with the non-biological you know we've started off with CPUs of von Neumann architecture we've heard about today the separating the processing from the memory and then we've also heard that biology doesn't do that it does a processing in the memory on the same thing your honors both a memory and a processor so none of these hassle von Neumann architecture okay so that's what it looks like it's beautiful we saw a picture of a teepee you version three in the last slide it's like a supercomputer it does like a petaflop or something or more Graaff cord is there there's a hundred pitiful ops in like a tracks you know that's it's a world's most powerful supercomputer almost not quite but it's getting up

there so these are a six is specifically designed to multiply huge matrices billion by billion matrices together which the brain isn't doing clearly in you know that's how many times bigger than the brain is that right it's got nothing to do with general intelligence but it's super good at multiplying it's good at linear linear algebra it'll statistically you know calculate things that's why Burt will get statistically brilliant but I'll never understand Shakespeare because it can't because it's Mulder it's linear algebra right any algebra doesn't doesn't do that okay and there is the world's biggest at three exaflop its summit chock-full of GPUs and I think AMD processors super super impressive but it's as dumb as a brick right doesn't do anything it calculates a big calculator so what are we missing so we we do all that in three pound right we don't need a three three football fields or or 10 megawatts we need we need a couple of sandwiches an apple and a drink you know and we can we can discover the

laws of the universe okay so clearly we we're not even close by any of those other approaches maybe one approach in hardware is neuromorphic computing which stated aim is to replicate the brain okay in hardware and so there's a project called spinnaker part of the human brain project which is doing that Steve Furber who was part of arm basically University of Manchester been working for the last twenty years he's built the team there have built now under the wing of the human brain project have built I think a billion artificial neurons so they're analog they post they process the information using analog not digital and it's more it's not the volume in it's kind of combined so they call that memristors where the memory and the processing is combined in the same place in the architecture looks completely different there it is that's a that's a in neurons right there so immediately we see you know it's a little bit smaller than the others that's all analog computing and that's what it looks like so

yeah we start off with a thousand neurons for core and then we put them into chips and it's hierarchical then we put 48 chips and aboard 24 boards in Iraq we have 10 racks right and that's sitting in a data center in Manchester today and you can login now if you want to set up an account log in and start processing stuff using your more for computing and you can put in the in this data set you can place a toku you can play chess and it doesn't do deep learning at all it's it's processing it with new or morphic algorithms and hardware ok so so for me this is definitely a step in the right direction so hardware is important it's not everything obviously the algorithms need to be there as well but I think we need to be running it on the right hardware otherwise it's a simulation you can always simulate anything on it on a chip but you need you know they've done that in the human brain project using you

know the von Neumann architecture I think they processed a second of what the brain can do it took them like a week in in in you know the whole data center so you know you can always simulate things but it's not efficient so you can actually do simulate in real time using us neuromorphic architecture because it runs efficiently like the brain I think it's about a thousand times less efficient than biology whereas their CPUs and everything else that we saw digital is about a million times less efficient so so it's definitely a step and this is the first generation so we're going to iterate just like we did with the CPU processing and a digital we've had 50 years we've got 10 billion transistors on a thing at centimeter big right so so we'll do the same with neuromorphic we'll have generation upon generation and soon we'll we'll have a hundred billion neurons because it's Moore's law its exponential and we will see you know sort of human level intelligence whether it will be sentient

and is down to the algorithms actually so I think I think we have the hardware we're a very beginning of this journey was the hardware that we need algorithm so let's look at algorithm so that's another one the brain scales that's University of Heidelberg here in Germany that's a part of the human brain project as a spinnaker versus a Google teepee use I would argue the one on the left with neuromorphic is the right hardware to be working on for general intelligence okay in this quantum computing nothing to do with how the brain works but you know sexy pictures I put it up there and it is it is like the last form of computing right we cannot forget it it's a big deal it's going to come online but I don't think it will be you know used to you know understand nature too well I think biology doesn't use it so I doubt we won't have to go to the quantum level but the quantum will be very important in other aspects perhaps

not general intelligence okay so there's the fourth summary of everything I've talked about four different types of hardware the digital and neuromorphic quantum and in biology right so we want to try to build a thing at the butt and a lower right-hand corner here and I think neuromorphic is is getting that's the way we'll do it and that's what they look like ok so the brain you see the similarity a little bit the neurons in the brain to digital the neuromorphic and quantum looks a bit out on its own weird but ultimately it's these little units of process information processing ok and then the data center of the future won't just be classical computing it'll be a mix of quantum you are more fake in and trend in classical so deep learning quick recap that's what it looks like this is what general intelligence isn't so it's not that it's not that in it and that's you know deep learning is not a GI so a GI is linear algebra so you know you're just

summing up the weights times these vectors and billion by billion matrices it does not give us a neuron absolutely not even close but never will do and it was never it was never meant to to be honest if you talk to the researchers like yin laocoon and geoff hinton they we're not trying to imitate the brain at off which we just found a stuff that works really well but in in geoff hinton he's in my last slide says that you know if we think really hard about how nature does it biology that's how we're gonna do it and he's recently said everything I've ever done on deep learning is sort of I'm gonna have to put that to one side now because to get to the next level and I saw a slide and some other presentation today you know it's in so you get a bit mainstream now you know deeper in it learning is running out of steam you know we can maybe tweak the knobs and get another two or three percent

of accuracy but yeah where do we go it's not general it's very specific it's not going to give us those nine types of intelligence and I think that's what people are trying to say even though they don't quite realize perhaps what this thing but that thing on the left is what we're trying to build not that little bit of linear algebra on the right okay so what is the theory okay you know I'd be I could just walk off the stage now and say huh you know we don't I don't know thanks for coming you know spending over money but I'll try I'll try I'll try I give you something a bit more than that so let's see so I mentioned earlier that it's not statistical right so what is it I it's physics right everything is built using the laws of physics so why not the brain right why not let's start there nice and simple so what do we know about physics right well we know Newton's laws we know relativity we know

general relativity we have Maxwell's equations from last century we're standing on shoulders of so many giants here max Hall Newton you know all the greats Helmholtz Boltzmann Gibbs you know we got the laws of thermodynamics we got quantum mechanics we don't think we will need that but we still have it we have relevance at quantum mechanics we have dark it we have stuff we don't know if dark energy dark matter we'd probably not involved as intelligence the thing is that all of that can be encapsulated in something called the principle of least action which says that nature behaves to minimize the free energy of any system you know whether that's microscopic or macroscopic whether it's a millisecond a nanosecond or the age of the universe there's this fundamental underlying principle of physics if you did a PhD in physics who did it who did a PhD in physics yes so a couple of people here will have seen this principle of least action and you write it down it's a nice beautiful formula you can

write it down Lagrangian or Hamiltonian form but it's you know one line and you can use it in principle to describe any system okay I say in principle because you know nature is complicated you know the brain is complicated as soon as you get more than a few atoms you know or a few more it's hard right so but in principle this is the underlying principle so that's what I'm going to start right with the theory of intelligent I'm gonna say if we can use it for everything else then let's apply it to the brain see how far we get that's what it looks like there's a beautiful book just written I think yeah last year Cambridge University Press surprised me there was nothing else before this because it's been around about 100 years it's called the principle of the section it's got all the different examples in the world of thermodynamics classical mechanics steam trains computers digital quantum everything chapter on each how its applied how its applied how its applied how its applied

so they need an extra chapter on how its applied to the brain I think but that's what it looks like you know that SS the action the Q is momentum tears time and Q dots you know the the V DT change of momentum of change in time and Delta s equals zero is all the physics it's a it's change in action nature acts so that that Delta s where s is the action is it tries to minimize that quantity where s is that in terms of Lagrangian okay don't worry about the math or the physics we're not going to go there but so let's apply that to the intelligence and since see how far we get because it works for everything else ok so again it's not just about pet recognition it's about modeling the world okay we model stuff we can imagine weak problems so we can build new models we can understand right we don't just do statistical analysis but otherwise it'd be zombies right or we'd be a silicon chip would be

built in an Intel Factory and we're not so here are some theoretical approaches I'm going to choose the first and act of inference one which uses the Delta s equals zero the others don't but these are these are very very very you credible attempts at general intelligence and all of these people have been working at least 30 years on these theories so there's certainly not to be dismissed and let's pick this one and just do all the others are wrong there's going to be bits of each there to kind of have have something is going to be overlap so you know we've Helmholtz started with is with statistical physics right it's the brain as a hundred billion neurons it's a big messy warm complex system so statistical physics is over a hundred years old how long oh okay so I will wind it up yeah let me wind it up so yeah thirty years for all of these right I'm gonna pick Tristan in the last two minutes okay now active inference we can look

this up again it uses a video it's minimizing into free energy let me just flick through and then I'll sum it up there's professor forests in there and we we don't have time to here so this is what it looks like we can draw diagrams and pretty pictures we can write down equations we can have me us in the environment the environment acting back on the agent the agent acting back on the environment all the physics is in place we can write down the equations for it we can do it for bacteria or the brain so it works on all system there's a method super super ugly but it's you know it's that's theoretical neuroscience and this is that that's the equation of effect a general theory of relativity for for the brain right so there you saw he's got loads of papers can we build it I say yes we have the data we have the theory we have the algorithms and we have the hardware okay and there's some AGI product is Marek's

here bins here there's a Jurgen Smith Hoover's not here these are all super clever people doing super clever things and I picked our active inferences one because I feel that's the one most based on physics and I'd say we're at the beginning of our journey it might take us five or ten years to sort of build us as we wait for the hardware especially to mature I believe we have the algorithms it's a super bold state to be saying but I encourage you to read some of some of the papers and there's geoff hinton saying at the very end you know this is three years ago the conceptual breakthroughs will take us we need new conceptual breakthroughs I believe the breakthroughs are going to under come when we understand the brain thank you very much I also would like to know more Mike can I steal five minutes of the Q&A so light for just one question I know that he overtake he's gonna make it up to you okay one question he always see

five minutes of lifetime yeah I'll make sure you get that I do five minutes off your lifetime to him yeah okay thank you so much for the question so we for your time just and taking one question so I just gonna be very quick to keep on schedule so I gotta just speak in very short war so richard feynman you know you've been a music producer besides all your other work in the academia and you've shown us a great slide where you said for a ji we need computer sciences physics neurosciences and psychology yeah so bring these four things together I mean the question is and you've been in academia in our society every every incentive is to be very specialized and to really focus on one subject and when people don't talk and connect enough with each other and I think we need that for AGI what are you doing or what what can we do as a society I need to bring these four groups together is there conferences to exchange those four

groups I mean how to get these different people these different tribes the question is that the short question well it was a long Commission I hope he has a short answer but yeah okay yeah so professor Kristen's groups in neuroscience group it's the theoretical neuroscience group they have very strong mathematicians very strong computer scientists deep minds a bit of a you know case study in this they get neuroscientist computer science sitting in the same room so it's a very beginning but I think do you need more I think he needs psychologists yeah so we need a little bit more input from other science sectors to begin with and yeah I mean it's yeah I mean you altima t you have to build these things the theory is going to be come from a very strong scientific method mathematically competent person okay you can get all the psychologists talking in a room and everybody else but ultimately someone or a couple of people have to write down a scientific theory just like Einstein wrote down the

theory of relativity so maybe I was a little bit generous you know we need everybody we need the ideas from everywhere but the actual person himself or herself is going to be have to be very super strong and Matt we saw some of the math it's not simple it it takes a while right is a huge it just like the theory of general relativity and I'll end here you know we can understand the concepts but the actual math is really really really hard right I don't know if you've ever tried it to solve that you know the field equations I have they're not easy right yeah so don't expect this to be easy it's gonna be simple but but the math is still gonna be you know hard okay yeah okay thank you okay Peter one more question where can we find your slides are you gonna upload them somewhere yeah I thought the conference is going to upload them if not then my SlideShare if is there available on my SlideShare so if you

go to SlideShare and just Google Peter Morgan I guess you'll find you you'll find him okay I think that's the best way and then the other thing is just to look at Cole frisson he has a great website with all his papers which is super heavy going but if you're really really I don't know you know if you really believe you want to try and solve intelligence that's the best place to start

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.