Rise of AI Conference 2016 · Berlin

Path to Singularity: How Authentic Are Thoughts?

Anish Mohammed

At Rise of AI 2016 in Berlin, Anish Mohammed questions whether human thought is really 'authentic' compared to machine cognition, and traces a possible path toward brain-computer interfaces, whole-brain emulation and eventual superintelligence. He argues the human brain, using 8 to 20% of the body's energy, is already extended by devices like the iPhone, a primitive brain-computer interface limited by the bandwidth of eyes and fingers, and describes retinal implants as an early, already-existing way of feeding electronic signal directly into the optic nerve. He cites a science study showing human memories are unreliable yet highly valued as authentic, and argues privacy of thought has never been debated in EU policy the way data privacy has. Moving to governance, he uses the DAO hack as a cautionary example for any future system of emulated minds ('EMs'), and explains trusted platform modules and zero-knowledge proofs, specifically direct anonymous attestation, as tools that could let entities prove properties about themselves without revealing information. He closes on ethics for machines, the trolley problem for self-driving cars, and an extended audience Q&A.

Key takeaways

  • Mohammed argues the iPhone already works as a primitive, low-bandwidth brain-computer interface that has not changed since 2007.
  • He cites a study finding human memories are fragile and often false, yet people still value them as authentic.
  • He warns that privacy of thought itself has never been debated in EU policy, unlike data privacy under GDPR.
  • He uses the real DAO hack as a cautionary example for any future governance system built for emulated minds.
  • He explains direct anonymous attestation and zero-knowledge proofs as ways to prove a fact without revealing how or what.

Detailed notes

The brain as a bottleneck and the iPhone as brain interface

Mohammed opens by questioning why humans consider their thoughts more 'authentic' than a machine's, and notes the brain consumes 8 to 20% of the body's total energy. He argues the iPhone already functions as a primitive brain-computer interface: before smartphones, people memorized directions, but now the phone is 'an extension of your brain', constrained by the low bandwidth of using eyes and fingers to interact with it. He says this interface has not changed even as phone functionality has grown enormously since 2007, and argues this fixed human interface will become the bottleneck for further human evolution unless people find a way to send signals directly in and out of the brain.

Brain implants, scanning and emulation

He describes retinal implants as an existing technology that feeds electronic signal into the optic nerve, which he calls a high-bandwidth bus into the visual cortex, arguing this already amounts to 'hacking your brain' to put data in, even though no equivalent method exists yet to get data out. He frames the next logical step as scanning a brain, understanding its connections, and creating an emulation, referencing Robin Hanson's work on emulated minds ('EMs') as looking at this from a societal rather than technological perspective, and states this approach seems computationally more attainable than some proposed alternatives on the path to a singularity.

Memory, authenticity and the missing debate on thought privacy

Mohammed cites a recent study, which he says was published in Science, showing human memories are fragile, biased and often unrelated to what actually happened, yet people still value the authenticity of their own memories highly even after being shown they are false; he notes court testimony relies heavily on such unreliable recollection. He then turns to privacy, noting his own background in privacy and security and Germany's relatively strong data protection culture ahead of GDPR, but argues that no EU Parliament debate has yet addressed privacy of thoughts themselves. He warns that once the skull is opened for two-way brain-computer communication, anyone with access to the connection could in principle know what a person is thinking, and that current cryptographic protections such as homomorphic operations remain largely theoretical.

From barter to blockchain: governing emulated minds

He sketches a history of transaction mechanisms from direct barter, to professional arbitration by lawyers, to centralized and then decentralized arbitration, arguing that if emulated minds ('EMs') exist and hold economic value, they will need governance mechanisms to protect them from each other and from other agents. He introduces public blockchain and smart contracts as one candidate mechanism, describing a spectrum from automation at the edges with humans at the center, to automation at the center with humans at the edges, the latter being a distributed autonomous organization (DAO). He references the real DAO hack as a cautionary example, noting researchers had warned against publishing the system beforehand, and uses it to motivate a more careful, trust-hardware-based approach.

Trusted platform modules and zero-knowledge proofs

Mohammed explains the Trusted Platform Module (TPM) as hardware that can hold keys and provide a baseline of verifiable trust, comparing it to a padlock on a house door. He highlights direct anonymous attestation as its most interesting capability: proving that an action was performed correctly without revealing how, illustrated with an Alice-and-Eve cave example and a zero-knowledge protocol for proving possession of a disease without revealing anything else to onlookers. He extends this to a phone verifying it is talking to the correct brain, and to emulated minds validating each other without knowing whether the other party is malicious, contrasting this zero-knowledge approach with symmetric-key schemes where anyone holding the key can open the box and see everything inside.

Ethics for machines, the trolley problem, and Q&A highlights

Turning to a side project done with friends, Mohammed asks how ethical and non-ethical behavior could be judged between emulated minds when no shared human understanding exists, proposing either a third-party entity or a cohort-based mechanism to make that judgment, alongside assurances of privacy and authenticity. He shows a simplified self-driving-car decision table (object in lane, erratic driver) as a toy example of translating human driving judgment into machine-readable rules. Asked whether he would buy a self-driving car that values another life over his own, he admits that despite considering himself an effective altruist, he is largely selfish and would not, and argues self-driving cars will only sell if buyers are convinced their own life is the priority. Discussing liability, he cites a past fatal radiotherapy overdose case that led to a lawsuit as a precedent suggesting legal test cases, not just philosophy, will shape how autonomous-vehicle ethics gets decided. Asked how to formally verify ethical decisions at scale, he says his team modeled decisions as a graph with high-probability traversal paths, but acknowledges the approach breaks down entirely once the state space becomes very large.

Quotes

Humans seem to value authenticity of their thought and authenticity of their memory way above everything else, even when it's proven to them that it is not really true.

Anish Mohammed

This will be the bottleneck for human evolution as such.

Anish Mohammed

Even though I consider myself to be an effective altruist, to a large degree, I'm selfish. I'm unlikely to do that.

Anish Mohammed

Questions answered in this talk

Is the iPhone a form of brain-computer interface?

According to Anish Mohammed at Rise of AI 2016, yes: the phone functions as a primitive brain-computer interface, an extension of the brain limited by the low bandwidth of using eyes and fingers, and its interface has stayed fixed even as its capabilities have grown enormously.

Are human memories reliable?

According to Anish Mohammed at Rise of AI 2016, a study he cites as published in Science found human memories are fragile, biased and often disconnected from what really happened, yet people continue to value their authenticity even when shown they are false.

What is direct anonymous attestation and why does it matter for privacy?

According to Anish Mohammed at Rise of AI 2016, direct anonymous attestation, enabled by trusted platform module hardware, lets one party prove it can do something without revealing how it did it, similar to a zero-knowledge proof, which he argues will be needed for entities like emulated minds to validate each other without exposing private information.

Mentioned

Whole brain emulation, Brain-computer interface, Retinal implant, GDPR, Distributed autonomous organization (DAO), Blockchain smart contracts, Trusted Platform Module (TPM), Direct anonymous attestation, Zero-knowledge proof, Trolley problem

Keep exploring

More from Rise of AI Conference 2016

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 6,064 words

Show the full transcript

So, I I would This is my first attempt at explaining this topic. So, I'll give it a try. Uh if at any point in time, if I become very geeky, put your hand up and stop me. So, uh it It's like I'm switching hats as I go along. The whole thought process was thinking about thinking uh as a human and then as a machine. And how do I transition from being a human to being a machine and being a Malaysian? And you know, we as humans, how do we really value our thoughts? Uh would we consider ourselves as authentic because as you recognized before, Fabian was kind of describing, you know, how the way we describe ourselves being different from machines is the originality.

And I'm just questioning the question of originality. So, I'm going to give it a try. And uh stop me uh whenever I go geeky and if there are things that doesn't very clear. So, there you go. Uh so, everybody has one of these. I'm assuming there's nobody in the audience who doesn't have one of these. And uh there's been a lot of work that has gone into it. Uh we've always been learning new things about this 1 and 1/2 kilo worth of stuff up in uh a skull. I mean, a brainy cage, right? It's a It's a cranial case that you actually have and the most valuable thing you can have. It takes up around 8% or 8 to 20% of your total energy that you consume.

Oh, yes. Okay, that's good. So, Why is it not going? Right. So, you know, over time, uh our understanding has been changing. You know, as more and more ability to understand the internals of a brain comes in, we have a better understanding of which parts of the brain gets lighted up when we see things. You know, the functional MRI have been very, very, very helpful in allowing us to understand, you know, how uh uh various parts of the brain really work. Now, you would ask me, why is it important, right? So, uh in general, what happens in our brain is a whole bunch of processing. And there's a bunch of meta-cognition that's actually happening. And we really uh don't have like very, very, very highly sophisticated models that we can actually uh take off the shelf and use it and uh pretty much say, you know, this is equivalent to a human brain yet.

So, it might happen. So, if you were to actually measure the ability of cognition and uh put a scale against it, I think some of us, at least in a social and empathetic scale and things like that, we might miss a few steps. So, this is not a linear one in the sense like you have to go through the whole path to get to wherever you need to be. But it is possible that we can assume a human to be a scale 10 and the so-called super consciousness to be something that evolves uh at some point of time. Uh you know, the different people have actually proposed different possibilities. One possibility is the fact that, you know, it could be alien as in you know, something could land on Earth and there emerges something.

Uh but in all likelihood, uh the likelihood would be we do something. We as humans do something and that will result in something that has that consciousness in that sense. Uh okay. Now, given that that's a possibility, you know, we we could probably think about, you know, we humans and uh you know, machines evolving along. The reason I thought I should mention this is like, you know, somewhere through this process, there's an interaction that's actually happening. And if you think about the iPhone, that's a good example I normally have. It actually changed the way you actually think. So, it's like normally, if you want to go somewhere, uh before the iPhone, you would have actually looked at the map and go, "Oh, I need to go there.

I need to understand how to get there." Now, it's an extension of your brain. You have an iPhone, you take it, or Android phone, you take it, and the moment you get stuck, Google Maps, or whatever maps that's in there. It's an extension of your brain, right? And actually, it's happening. It's what you have described there is a a brain computer interface, right? It's a primitive one. We're using our eyes and our fingers to actually do the interface. So, you have a real bandwidth restriction. The number of bytes you could send and receive is pretty low, right? And this is an enhancement actually actually we brought to our brains without even a hacking your interface at all.

Okay? And this is only going to improve as things improve, as Moore's law, you know, happens. As the, you know, as you can imagine, like 10 years ago in in 2007, the first time iPhone came out, the functionality that it had and the functionality it has now is very different. Right? And and the same interface is still there. We as humans still have the same old interface. Our interface is not changing. And this is a key thing that we all need to understand. And this will be the bottleneck for human evolution as such. Okay? If humans have to improve, we need to find a way to drill a hole through the brain you cage I was showing and look at that very interesting 1 and 1/2 kilos of stuff that's stuck inside and find a way to send signals in and take signals out.

Okay? So, uh before I go into that, I probably should say this, right? If you were to think about that improvement in human evolution, uh the first thing that will happen is like we plug something in uh without actually opening up our skull cage, like an iPhone. And then, at some point of time, we would actually open up our skull cage. I mean, it's already happening. So, there are retinal implants, right? So, in theory, a retinal implant is allowing you an interface from an electronic signal into impulses into your optic nerve. And optic nerve is a high bandwidth bus that goes to the base of the cortex, right? Oh, I should probably show a picture.

Okay, let me see. Sorry. I should have a picture so that I can actually Oh, yeah, hold on. Right, guys. Where's my pointer? Uh Which one? Pick the other one. Okay. Press the thing. Oh, yes, yes. So, uh over there or over here, that's where your optic nerve actually So, oh This is the cortex, right? And from your eyes, there's just two things that goes in and it crisscrosses across and goes into the optical cortex. And it actually supplies the back part of the brain, right? So, this is like a high bandwidth channel, a bus, if you were from a computer science perspective or a hardware perspective. You have a bus available to you. And if you have a retinal implant, pretty much you already hacked your brain.

You got a chance to put data into your brain. We haven't got a a a path to get it out yet, but there are ways of doing it. But, you know, I'm just describing what is actually happening. So, now the next logical step to happen, if you hack your brain and put stuff in, is to actually understand, you know, scan the brain properly, right? Understand what it does. Understand its connections. And then do an emulation. And this is a very interesting book from Robin Hanson. Uh if you have a chance, you should actually read it. It kind of looks at it not from a technological perspective. It actually looks at it from a societal perspective and not the political perspective.

And that is why it's it's a a more relevant and a lot more interesting. And for everybody who's interested in where humans are going, this is going to be one of the key steps before the so-called uh you know, real singularity might happen. It at least in my opinion. It's like this seems to be like a possibility more easier and attainable in that sense in terms of computational complexity than other solutions that are out there. Okay, let me try my next one. Okay, so uh there's an important thing that I thought I should mention this to you. Everybody thinks we as humans are unique because we have memories. And we have ability to create, you know, dream is a memory.

Uh you know, if I were to think backwards, yeah, of course, you when you wake up you have memory of something that you saw while you were asleep. That's why dream is a memory. And uh as you could recognize, you know, most of our memories are really fragile. It's very biased. And if you have any memory of anything happening, you should always cross-validate them with some data points that's available. Actually, this was a very recent study that was published in science, I think. And uh they had like a whole bunch of things that showed that, you know, whenever people have memories, it doesn't really relate to anything real. And uh and and this actually should make us question the real basis of human society.

So, if you were to be taken to court, the thing that happens is you'd be cross-examined, right? What is this happening here? Somebody's asking you questions about your memory. And somebody else's life is actually hanging on your recollection of something which is totally screwed up. Okay, so so and now I mean, this is why I I found this very interesting. So, humans seem to value authenticity of their thought and authenticity of their memory way above everything else, even when it's proven to them that it is not really true. Okay, so now this is a question of privacy. Every I mean given that I'm giving this talk in Germany, it's very interesting. Privacy, I mean all through Europe, there is GDPR coming up.

Uh okay, um my background is in privacy and security. So, I pretty much see Germany as one of the countries where there's quite a bit of, you know, protection for privacy of an individual in that sense. And I think about this, right? Now, the the the information that we are trying to protect is generally IP and things of that sort. sort. Imagine you you the privacy of your thoughts. There's no debate that has actually happened in any of the so-called uh EU Parliament or any other places to actually think about this. So, this is the thing that's going to happen when we actually have our brains opened up. The skull cages is protecting us from the external world.

When we drill a hole and connect the wires into it, if the communication goes backwards and forwards, we have a challenge there. Anybody who has access to the wire will know what you're thinking. I mean, at least that's a possibility, right? Now, what happens? And what happens is interesting. What happens is interesting in the sense like uh this both the cases that I was describing, A when we actually just have, you know, brain enhancement, and B when the case we actually have an emulation as such, both the cases, privacy is almost always lost unless we find really unique ways of doing this. The there are mathematical ways you can prove that you can do an operation and not allow anybody else to recognize it, but those are all theoretical, okay?

We haven't actually come up with ways by which we can actually do this in real yet. So, yeah, homomorphic operations, yes, you can do that, but it doesn't really mean anything. Okay, right. Okay, this bit is I should ignore it, but I had the slide. The last bit is a bit bit that is interesting for everybody. So, this is the bit that's most interesting. And there's a reason why this is most most interesting. Uh I think I should probably show you the slide rather. Okay. So, throughout history, this is how things have happened. So, when human started transacting, when they do barter, I I don't know if there are economists in this room. And if I say something rubbish, please correct me.

I assumed barter was one of the first set of transactions that happened. So, that's a direct synchronous transaction. You actually negotiate and you get your stuff done. And then professions came in, lawyers who make a lot of money. Sorry. Any lawyers in the audience? Apologies. So, they did arbitration. And then there were centralized arbitration mechanisms. And of course, you can actually have decentralized arbitration mechanisms. And this is the bit that's of a lot of interest to every one of us. If you were to have EMs flying around, right? One of the things that you actually want to happen is that the EMs are protected. Protected from themselves, protected against other EMs, and protected other other agents that are interested in the information or things that the EMs are have.

If EM has, in any way, shape, or form, any economic value, it will become uh what do you call an object of interest to people. I would say set of all agents. Could be malicious or not, we don't know. So, we need to have mechanisms by which we could allow governance to happen. So, this actually possibly offers a solution. Right, let me zoom in. Okay. And I'm pretty sure everybody knows this. Right? Is there anybody in this room who doesn't recognize it? Okay. Guys, it's called blockchain. The what The next speaker actually would talk a lot about it. I'm not going to talk anything about it. So, the the reason why blockchain is interesting is is the following, right?

So, it allows you to have a publicly available auditable data set. So, you could actually have the whole set of rules that you apply on your smart contract available to public. And the public could look at it and say, "Look, this is the rule. This is how it should be applied, and everybody should be applying it fairly, okay?" I'm I'm not going into permission blockchain or anything like that. I'm describing the public blockchain, right? And actually using that the basis of smart contracts. So, this is one way of looking at organizations and how humans and machines interact, right? So, you can look at automation at the edges, and you can look at you know, things at the center.

You can actually have humans at the center or automation at the center, right? So, if you actually have what I would describe as automation at the center and humans at the edges, you actually have what do you describe as a distributed autonomous organizations, right? And this is really interesting. This is really interesting in the sense like I know uh this has some slight connotations because of some events that happened, and I will talk about it in like two slides time. But if you assume it's very secure, let's make an assumption, right? Let's assume people do things properly. We do verifications, and things go according to plan. This should allow us I know that that people are they're laughing at it.

Yeah, I fully appreciate. I'm actually going to talk a bit about it and why I said those things. Uh because uh all the terms actually have their own meaning, and it it should be understood in the right context. It's not the first time in the world that verifiable systems were built by humans, and verifiable systems do work. And there's a whole bunch of mathematics that actually looks at that. So, I wonder why. So, anyways, uh the the concept of distributed autonomous organizations, the fact is that if we were to have a whole bunch of EMs, this would allow us to actually construct, you know, governance frameworks to allow those EMs to collaborate among themselves or with us as in the humans on the edge.

Okay? So, that's the reason I thought I should describe this. Um but other than that, hm, that's good enough. If I go Right. This is why some of the people in the audience were laughing. Okay? So, there was an attempt made to actually do this. So, first attempt was made a bunch of researchers said, "Guys, stop doing this. Don't publish." Uh well, nothing happened. And this happened. And well, you know, it's a case of actually building a system and having things not fully verified. So, I'm going to I have a like a bunch of slides kind of trying to describe how this could be done slightly differently assuming we are not going to use uh the parameters and the framework that the guys who did previously did.

Okay, this is no reflection on what they did and how they did. It's just the way that uh you know, people from the academic world who have actually been trying to implement things in trusted systems having doing it. Does anybody here recognize what this is? This is pretty bad, guys. It's actually written there. Can't believe you can't read it. It's a TPM, a trusted platform module. Okay. So, I'm assuming there are no hardware geeks in this room. Man, so bad. Right. Okay. So, this is the definition of trust, right? And the key thing is the feeling of certainty. I don't know what this means now. Uh but uh this is how it's been defined and the definition of trustworthy systems.

Okay? Okay. So, this is the key bit. A system's behavior can be validated in some convincing way. Okay, and this seems to be very tricky for everybody. And from what I understand, there have been systems previously that were around that have actually been gone through the process. Right. Okay, so there's been quite a bit of work that was done on this magical hardware called Trusted Platform Module. So, what it does is kind of uh I would say irrelevant, but I'll give you a flavor of what it does and why it's important. So, generally speaking, when you have trust, you need to trust somewhere or something to start with. There are ways by which you can actually have a totally bootstrap trust that's decentralized.

But the problem is once such a system is bootstrapped, if you want to go back and fix something, it's almost impossible. And that is exactly what the previous problem showed us. So, the DAO problem was you actually had like uh a kind of really decentralized, but you know, closer to decentralized, and you need to fix something. There's no way by which you can go into the system and fix it. So, you have to fork it to actually fix the problem. So, this is uh slightly lower How would I say? Slightly better situation or slightly less flexible situation. So, this is a piece of hardware that pretty much used to come with all the motherboards if you ask them to have it.

And then you will have a a bunch of keys in there. So, the way I would describe it is if you were to buy a door in a house, and the door in the house comes with a a padlock, and if you want, you can activate it and use a key to operate and lock it. That's the easiest way to describe it. So, you have a a level of security to your house, assuming your other doors are okay, and your windows are not the ones that people will be able to get into the house. So, right. Okay, I am not going to describe the whole thing. Okay, the most interesting bit about the TPM is this ability to do what you call direct anonymous attestation.

I shall try to describe it a bit more using this. Right, okay. So, who has heard of this before? Come on, guys. Put your hand up. I know you guys have. Right, okay. So, this is a way by which somebody proves to you about something before and after the the person proving to you that they have proved to you that they have the ability to do something, you don't know how they did it. Okay? Let me explain what that means. So, you have this Alice and Eve kind of thing where, you know, Alice is sat outside. Eve says to Alice, you know, uh we have this particular cave or this room. Uh we know for sure that there's no way by which I can go from side A to side B.

Uh you know, you can go in and check. But, what I will do is I will go in the side A and I'll emerge on the side B. Effectively, what this proves to uh Alice is the fact that Eve has a mechanism by which could do that, but doesn't show Alice how to do it. So, at the end of the day, Alice just knows one bit, essentially that the Eve has the ability to do this. This is very interesting in a sense like in a lot of very important when very important questions are asked, you want to have an answer being given, but you don't want the person who's actually asking the question and getting the answer know any more information than what you just told them.

Okay, I'll I'll give you an example. Imagine you are asked to prove whether you have a particular disease or not, right? And publicly revealing that is not a good situation, right? If you were to as a zero-knowledge protocol, you can actually prove to somebody who knows whether you have it or not what what it is, whether you have it or not. And the end of the protocol, uh the person who asked the question would know for certain that, you know, whatever your status would be known to them. But everybody else who's actually watching this would know nothing. And this is interesting. Imagine that you have a bunch of VMs. They want to transfer information between themselves as in validate information between themselves.

They can do this. Cuz they have no way to know whether other VMs are malicious or not. And similarly, if you have your brain and you have an entity that sat outside, you have the ability to prove to your brain that the the entity that sat outside is a real one or not. Am I making sense? Did I lose you guys? Okay. I'll try once more. Okay. So, imagine I have something sat outside my brain, right? So, my iPhone has this interface that actually talks to my brain directly, right? And I want to make sure that uh my iPhone is my iPhone and my brain is my brain. And I don't want anybody else to recognize what the hell I'm talking to my iPhone about.

Okay? So, I'm doing this in the public and I'm just doing this protocol where I run this question, right? Like this. At the end of the protocol, I know my phone is my phone. And a phone knows my brain is my brain. Okay? Okay. So, let me explain Let me ask you a question given you asked a question. So, now, okay. Now, you know how symmetric encryption works, right? So, imagine I have a box and a box has a lock. And the lock has keys, right? Unless you have the keys in your hand, there's no point in giving you a box. Right? Yeah, that's the problem. So, here the thing is like you are allowing everybody to actually see what you're transacting.

But what are you trying to do is literally the people that really know what they are interested in, they get the information. I mean, they get validation. They don't want to open anything. They just get the validation. If it's a symmetric key operation, it's like you actually need to do a a key management and a key setup. And then everybody has to be enrolled on the system. And once everybody's enrolled on the system and in a symmetric key scenario, anybody who actually has the keys to it, whatever the box. Everybody has the same key, so the box opens. And everything that's inside it is known to them, right? Sorry. Yeah, so this is one of the things.

And one of the other things that I found very interesting was the construct of ethics. I know what time is it? This one went down. Do you know what time is it? This one down. Okay, sorry guys. So, then this is the question of ethics. We were talking about difference between humans and machines and Oh, yes, I have 15 minutes. Yes. Right, this is a bit of work that I've been doing on the side. This is my spare time activity. And I've been thinking about Okay, so I shall admit I've been thinking about this very hard and I've been thinking about you know, what could I do in my limited ability to move this a bit forward.

So, I convinced a couple of my friends to join me in my crazy search for stuff. So, these are like a high high level uh high level scenarios that we use to look at what humans normally do and how humans make decisions and how could you actually translate that into either machine to machine or between agents. Okay, let me try describing that slightly better. So, imagine uh that there's an interaction between two humans, right? And we know what is ethical and what is non-ethical, right? So, what we could know is like if somebody does something that's non-ethical, both the parties involved know this is non-ethical. But imagine it's if the case of EMs, right? So, how do they know this is ethical or non-ethical?

Can you actually have a a third entity in the ecosystem which actually tells them whether this is ethical or non-ethical? Or can you actually have cohorts of the system that could actually tell them whether this is ethical or non-ethical? And that to me seems interesting. Because it's like if you actually have a copy of your brain given, what is the assurance that you have that that that entity is actually going to be used in an ethical manner? That to me sounded like an interesting start of a question. Interesting start of a question. We don't know about what the next level is, but this is like the basic level that I thought would be interesting. If you actually had your brains, uh, how could I say, processes exposed, you want a level of assurance about privacy.

Then there's a level of assurance about authenticity, and a level of assurance about the ethical use of it. So, you know, at least you know some semblance of fairness occurs at the end of it. So, this was an attempt at doing it, but let me see. Okay, we'll skip that one. So, this is like a classic one that you know. Can anybody see the key challenges on this slide? It's a very simple thing. You guys drive. I don't drive, okay? So, for me it might be more difficult. You guys drive, so you should be able to find it easily. So, the way we tried to do this was, you know, look look at all the key challenges and see what the action should be and how could you actually make a machine understand what you do?

So, this is what it is, okay? Cases, object ahead in the lane, moves out of lane. Okay, this is classic case and I'm sure you recognize it. Uh then there's a question of erratic driver. Uh you might have alternate solutions to this problem, but you know, these are the obvious ones we found. So, you know, you could actually, you know, use this as a basis to build other systems. This is like a very dumbed down version to actually understand, you know, how humans generally look at things and translate that into that takes and how how that could be understood by machines and how you could actually make that uh how could I say uh in a comparable state and make two machines agree on something.

All right. Okay, this is the stuff if you guys get really interested and you want to hack it, go hack it. And if you have questions, drop me a note. Right. Uh I'm going to skip that. And uh what I would say is like given I've been saying so much about all these things, I think it is possible that they might be, you know, near EMs that exist. When I say near EMs, uh I mean is that you know, for smaller uh organisms, we have the ability to actually scan them to a degree. Uh I think open worm was an example that you actually can actually look at a worm and like look at nematodes which is like a lower lower organism with a lower number of neurons and lower number connections, you could copy them and then you could actually run it and then look at what the outcome is.

So, you know, if you could do that at some point in time, maybe. So, the distribution is varied and uh that's all, folks. Now it's your time to shoot me. Yeah, go for it. You want to So, in one of the last slides, you had this trolley problem, I think, kind of outlined with a self-driving car. So, my question to you is, would you would you buy a self-driving car Okay. So, would you would you buy a self-driving car that in certain situation would value another life over yours? So, so would save another life So, the question that I was asked was something along the lines of, you know, do I really value my life?

Summary answer, uh even though I consider myself to be an effective altruist, to a large degree, I'm selfish. I'm unlikely to do that. But But that kind of solves the trolley problem. Because in the end, self-driving cars would only be successful if people um if people buy them, and people would only buy them if they could be sure that their life is the highest priority, right? That's already not the case. Yeah, that's that's what I was saying. Well, people people buy Teslas all the time, and they certainly don't make sure that you stay alive. Well, that's it's it's It's a very long debate. It's Some of it is actually up to politicians to decide uh how things should be handled, okay?

And legal There will be a bunch of test cases. So, I'll give you a couple of examples from health care where things have gone south. Uh so, a long while back, uh there was like a radiotherapy device that was created. They gave her a a fatal dose to somebody, and somebody passed away, and they actually sued the company. So, some precedence of things like that that will actually make people think about things. So, what that would mean is like if you actually bought a car and if the car killed somebody and if you are going to be sued for manslaughter as in you the occupant of the thing, then you start thinking about it much more carefully rather than thinking about my what my value of life is, right?

If I go to jail for until the end of my life, man, I will really consider it, right? Yeah, that's a different question. I am not a politician yet. It's an open question. It's an open question. And open questions don't come with easy answers. I mean as humans have to really decide which way we need to go and lots of points in life. I mean and evolution as such and this is one of those things that we need to really recognize. You know, should we have machines as equals? Should we have other humans as equals? What should be the what should be the guidance to machines given in terms of questions like this. And if you can come up with an answer that's effective over a large set of times, I think that will solve a lot of these problems.

Because we'll be forced to do it. That's what one of the person in the audience said. It's like Tesla already has crossed the path. So, it's a matter of time. You know, when when we had cars and deals on the roads, we had the same question. We actually had like people running in front of it with the I can't remember what it was, a flag or something like that saying a car is becoming behind, right? We have come to a point where we cars can't drive on the road. So, I hope something will evolve. You had a question? You want to shout? So, so my understanding is that you thinking about how do you evaluate Yeah.

Things if your brain was used by Yeah. Yes, yes. Not just to you know, the the the concept is if you have a there are multiple things that you can think of. You can think of your brain brain talking to something else. So Yeah, yeah, yes, yes, yes. Yeah. Mhm. Okay, so so the first thing that you need to actually recognize is you know, how do you capture all the so-called ethics in the machine readable form? That's what we were trying to do the very minimum, right? Then there's a question of valuation, meaning you should actually have some quorum mechanism or validation mechanism where you know for sure that you know, at least a whole bunch of entities get the same results.

So if you go to a court, you know exactly this decision was made using this ethics applied in this manner rather than randomly. That's the best we can do right now. So you can actually have like a verified system. So the trusted computing base would allow you to have a verified system. So the verified system will allow to have you you can actually attest to a set of properties to a third party without revealing really what it was done. So that's what DAA really helps you. And then once it's done, if you have anything that's formally verifiable, that should give you the same results. So uh How do you tune or Okay. Okay, that's an open problem.

I haven't really fixed it yet. Oh, yes, go for it. Adapting, yes. Yes. So, I fully agree with you on that problem. So, part of the part of the way we are looking at it was in a slightly what I would describe as stupid way. That's my classic way of looking at problems. So, sorry. So, I was thinking of it as a graph, okay? And I looked at the graph and I looked at the I assume there's a tree in the graph which has a high probability of traversing. And I I mean we you are able to apply a whole bunch of logic at that point in time to validate that. You know, will this actually go at this decision point or not?

If it is, then it's you know, we have an idea. Where state space is really, really large and we have no way of going forwards and backwards, all these mechanisms totally fail. Put your hand up and tell you this truth. I mean, we can only try you know, very basic stuff. So, an attempt Somebody needs to start somewhere. Guys, thank you all. Thank you.

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.