What if we're on the wrong path to AGI? Ft. Alexander Naumenko

Is today’s AI built on a flawed foundation? In this episode of AI Rebels, Alexander Naumenko, software developer turned AI theorist, challenges the dominant paradigms of logic-driven and statistical AI. Drawing from his concept of "semantic binary search," Naumenko proposes a new model of intelligence inspired by the game 20 Questions—where intelligence is about selecting the best option under constraints, not predicting outcomes. He argues that true intelligence hinges on recognizing relevant differences and defining features, not statistical similarity. His vision? A radical rethinking of AI—one that may finally bring us closer to real general intelligence.

Transcript

0:18

hey everybody welcome to another episode of the AI Rebels podcast I'm Jacob I am Spencer and he we are here with Alexander uh Jake say his last name you had it way better than I did Naumenko thank you for inviting me thank you for joining us uh just to start Alex um what kind of just tell us about yourself what lead you into AI to start with uh what were you doing before AI if there was anything um we'd love to hear about it yeah uh so I am a software developer uh and I have uh like uh more than 10 years of experience now but I wasn't really uh into a I even though I started to like be being interested in the a I er from 2,018 so it's about seven years now uh but uh I really didn't have any practical experience in that area so uh huh uh I was pretty much least likely person to contribute to the field but uh uh when the war started I I needed some distraction from those stressful news and because of that I started just thinking about AI and I needed to focus on something and I selected the NLP Natural Language Processing and it was pretty lucky choice as I see now because starting from like from the top uh because it it works with symbols and and and concepts and the stuff allowed me to discover some key principles behind intelligence and behind natural languages and now I'm trying to like uh share my ideas in those areas initially I hoped that I will try to contribute at least something in some tiny area of AI and now I I I hope uh to like redirect the the whole a I research agenda that's big that's ambitious yeah and it's so that's fantastic Alex can you in layman's terms in simple terms tell us what what it is now that you're focusing on what are you researching what are you thinking about right now so uh mainly I was uh interested in uh what concepts are and how language relies on them on their usage but now I am more into the algorithm behind intelligence and I am not talking about artificial intelligence I'm talking about natural intelligence and from what I discovered I see that all life is intelligent and in that I agree with Professor Michael Levin and he he does a great job in his area and I really enjoy uh that my ideas uh correspond to what he does uh and I I find some support from his work to my ideas and that encourages me to continue in my research but basically it's intelligence it's components and how those components work together okay can you give can you give me an example of what it is that you're with this intelligence algorithm like what is it that what does that look like okay so uh I uh I I like to mention frequently in my posts and in my tweets that uh intelligence is best demonstrated by the game 20 questions and uh that game uh demonstrates the the the core algorithm and this algorithm is applies not only to intelligence it applies to to all cognition all cognitive functions and how I see it is well I defined it as the selection from available options of the most fitting option given some relevant constraints so let's consider the game 20 questions and for instance the task of object recognition we have some object at hand and we want to recognize it and we ask questions based on properties of that object so properties of that object play the role of constraints in the algorithm and all the categories known to us are available options and then when we apply those properties that we check and compare to and we filter like select the the best option from available ones so applying each consecutive property of the object we filter out non fitting ones and when we are down enough we eh to the bottom of the tree we are left with only those uh categories that apply to that object because we we took into account all the properties of of that object but really that game teaches uh much more than just the core algorithm because I see for instance if we uh going down we introduce differences between different categories but when we start ignoring them we demonstrate what generalization is the the central topic of a G I because a I is is easy but general intelligence is difficult for for machines now and it it's that simple just ignore differences and and we we are generalizing but as I claim we we require equally both specialization and generalization so going down the tree and going up the tree okay and it depends on the task which we pursue at the moment because at different levels of the specialization tree different rules apply and it may be relevant to the task or task dependent so what is the biggest what's the biggest difference right now between what you're proposing this this neural pathway this way of structuring intelligence versus where a I is at currently with its reasoning capabilities and all of that uh so uh basically so far we had uh two major approaches to AI uh the first one was uh good old fashioned AI and it was uh relying on uh rigorous logic and I claim that rigorous logic is is not applicable to intelligence first of all because it is impossible to reach like ultimate precision and is it a good way to maybe is this like a if then statement this rigorous logic you're thinking of yes is that it maybe a good example okay yes but often we have those like vague moments we when we we are not sure is it day yet or is it already night so that's twilight Twilight Zone er in between so and the those areas are quite common in real life and in real life scenarios and we we shouldn't forget that we are real time animals we need to make decisions fast immediately uh like within one second we do not have time like er selecting different er developing new approaches and we only have what what is available to us at the moment and we we may either apply those or go go some some somewhere else and find something else so and the because of that rigorous logic is not applicable as I see and the also intelligence is definitely not relying on statistics on guessing so we do not predict our future actions we select actions and I like this example of Roger Federer who is contemplating the shot and if he is predicting that the that he will miss the shot he wouldn't go with it right it's unreasonable but he selects the the manner in which he will perform the shot that's it and then a cesserara cesserara Italian phrase which means be it what it may so we selected the best way to proceed and then be it what it may so yeah uh we we we we we cannot reliably predict future and uh but we can select uh given our current knowledge what will work best for us given our current knowledge to know what we know so far if we have some gaps in our knowledge it's a completely different story we we need to fill fill in those gaps and because of that we we may decide to go and experiment and the collect necessary evidence to to know how to behave in in the future and the that also is one of the critical pieces that I consider novel in my approach because for instance in reinforcement learning we rely on known reward function but in case of unknown recipe we we cannot calculate that reward we cannot we do not have any reliable information to to to use and I claim that collecting this information maybe even more important than the success as such and because of that filling those gaps in our knowledge is is kind of rewarding itself the same applies to failures because if we fail at something it just means that what we did didn't lead to uh those results that we wanted but it definitely LED to some uh result and we should analyze how those parameters initial LED to that result and we then will have that information to take into account when we select how to behave in similar situation there are a few studies out there that have shown um that the the impulse to act occurs in the brain before the conscious realization of of that impulse to act and so a lot of our rationalizations and and reasons for doing things are post hoc rationalizations right you know we make the decision subconsciously and then and then we rationalize it with what we know I'd be curious um how do you think that your kind of theory of intelligence interacts with with uh that phenomenal okay how can I explain that right yeah yeah so I am perfectly fine with the uh with intuition and subconscious uh so when we uh collect when we interact with the world we subconsciously we collect information about how different parameters lead to uh different results uh huh like uh in uh in the along the idea of Einstein's definition of insanity uh huh right doing the same thing uh over and over expecting different results but in fact we observe how different parameters lead to different results and eh because of that eh our intelligence eh works eh again eh by eh chopping so if we observe that some perimeter always lead to some result then we know that okay this perimeter is responsible for and is relevant to that result and again if uh some result is always different no matter how we change that particular parameter then that parameter is irrelevant and this relevant versus irrelevant is pretty much important for intelligence and it is it is pretty good at spotting what's relevant and what's what's not and it keeps in mind keeps in memory what is relevant for each particular action so I consider that actions are like building blocks so they they stand the side in our brain so we consider each action uh separately right and and we track how different parameters uh are relevant or or not to to that action and because of that for for each action we have some set of algorithms behaviors recipes to select from so what what and we or or we can say that it um it is about the result that we want to achieve so that result can be achieved by using those uh different actions and we we know that they lead to that result so we may select from them so that's the the idea and what what you were talking about is that some of those realizations internally they they do not need to be rationalized or communicated to other people because because they are hmm too tiny or too subjective so if I know how to play snooker for instance and how to kick uh balls still you will need to develop this skill on your own and we can only discuss like high level stuff so just uh try to uh kick it sideways or uh push it forward uh the the the cable or whatever so uh but uh the tiny details are connected to to each individual that's why by the way it is necessary probably to really develop some skills with with connection to to particular instance of the to particular agent so to say yeah be it robot or human or whatever so and what it brings us that internally in our head we have tiny details and large details we uh we can talk about large details like generalized uh versions but not small details because we do not uh consciously aware of them yeah so uh and uh and that makes but intelligence is is aware of them he he knows about them and he takes them into account and because the the whole mechanism relies on just a quick comparisons of quite limited number of features then it it works like instantaneously and we intuitively know what to do in each situation in real time but sometimes those decision may be not so perfect or rational or whatever but it it's perfectly can be explained by our past experience because interesting uh we we we think that uh that information that we acquired in the past is basically what we know what we can select from so yeah eh we we cannot apply what we do not have at our hands yeah so okay I think this kind of I think it's starting to make sense to me I think I think I'm getting there ha ha but I'm curious with artificial intelligence in its current state obviously its intelligence is completely based on the data we have fed it right it's it has no solid at least at this point solid interaction with the the external world it's fed data and it responds accordingly but what you're saying do you think a I at is is it possible for a I to get to this level of intelligence without having these human experiences it is possible so uh what I'm saying is that uh for instance when we uh categorize objects we follow this procedure described in the game 20 questions so we introduce those differences like is it tangible is it alive er does it er is it animal or and and does it feed its its cups with milk and er and and so on so and er you see we only have few comparisons few uh features and we already down to mammals uh huh but imagine what uh what uh neural networks do they are fed with whatever information they are fed for instance number of limbs uh the type of skin and uh size uh weight and so on and probably uh average life life expectancy is it enough to indicate that that animal is mammal no the defining feature is not there and it doesn't know about it it has to go with what it it was fed with so because of that they uh they they do not rely on the proper uh defining Features mhuh and that 20 questions game it teaches the importance of defining features mhuh because they are relevant all other features are not uh huh so if if I am talking about that some features define some animal as a cat do do I need to consider weight size colour and so on uh name as a defining feature no and it brings us to one crucial misconception about current AI approaches they rely in the in their categorization and it really applies to anything because everywhere we work in AI or in cognition like uh natural intelligence we need to to match patterns right and it's it's basically categorisation and the current approaches rely on similarities so they consider cats and try to figure out what makes cats similar so what what brings them together right but in my approach cat is defined not by the similarities with cats it is defined by differences from anything else mm hmm hmm so what we have internal set of cats and we have external set of anything else yeah mm hmm defining features they differentiate cats from all other animals and not only animals from because uh for instance if we consider one interesting theory uh exemplar theory so it explains that we recognize uh objects by comparing them to exemplar of of that category mm hmm but how many categories do we know like millions right and do we compare a given object to to to a million categories no we go by eliminating roughly half of the possible categories by following the 20 questions procedure uh huh is it tangible and immediately love democracy justice are given away is it alive yes no so mountain is no longer considered and immediately we narrow down the the possible uh and because of that when we are down to cats we probably checked only 10 comparisons uh huh that's it but if we are talking about comparing 1 million categories we have even more problems with that approach because what are the defining features to compare yeah this is a colour weight size this is reminding me of the uh the old apocryphal story of I think it's Socrates on the agora and they're they're trying to define what is a man and Socrates comes up with the definition of a of a featherless biped and then Diogenes shows up the next day with a chicken that he's plucked all the feathers from and says behold a man hahaha reminds me of that hahaha yeah that's exactly why trying to uh come up with features based on the class instances yeah it's it's it's useless right so we need and one more key aspect of this key observation uh we cannot come up or suggest one concept uh huh uh like uh not connected to anything else right right yeah so there is a Russian saying uh which is roughly translated as uh everything is recognised in comparison we need at least two concerts right mm and and we need to to know what differentiates them so yeah this is fascinating so you're wanting to essentially 1 I'm gonna like say some things back at you and just correct me if any of it's wrong you're you're trying to 1 decode the way the human brain reasons through recognition and decision making yeah that sound about right yes and and with that it's taking that reasoning of yeah the 20 questions so with a I as it stands what kind of an effect would this have what benefits are there to to adopting this uh okay so uh first of all uh there was a funny example in Twitter this problem of a farmer wolf cabbage and a goat but there there was a a bridge mentioned and the alelem in question er started to like enumerating all those steps that take the goat to the other side go back and so on so it didn't consider that bridge was in anyway relevant to the task but we we all know that it is right but what if it is something else was mentioned like a tree it is irrelevant okay but LLM doesn't know the difference between relevant and irrelevant preamble to the task and if it was trained that in such and such task when goat farmer cabbage and wolf are mentioned list those steps and you will be fine but change those parameters to like ogre chocolate and some small girl and you will be making new game all together to Dale Lamb and it will not know what to do with it uh huh but the the task didn't change at all so that's what I am talking about relevance and not and this er relevance is is one key that er will change a lot in what we have today but what I propose additionally is this er concept of defining features uh because really we will uh faster recognize objects we will perform much fewer computations and by the way this allegory of computer computation eh inside our head we need to specialize it better it is comparisons so what what I'm talking about that eh when we er perform some er reasoning inside our heads we are relying on our internal representation of some object uh uh and we operate with all the knowledge that we have about that object at that moment but when we talk about it we mention only only its differentiating features so like category yeah but in fact each action relies on a on the subset of the properties of that whole object and this is a very important concept because when we when we see how large language models produce their answers they most likely rely on mentioning those algorithms applied by humans and described in different sources that were included in the training set but this is not obvious and it's not directly follows the the general process and the general like workflow because when we talk about stuff we do not rely even on the required properties of those objects let me demonstrate on a very simple example so if if I say uh take a seat uh I I pronounced only like uh two words right but in those words is it possible to encode the location of a chair or some other in instance of furniture which is mm hmm sitable right yeah right so how can you uh like perform this action what what I did I only invited your cognitive abilities like perception reasoning memory imagination even to perform all the necessary heavy lifting to perform that that uh comment that request right but the the cognitive heavy lifting was performed by by your other cognitive abilities and that's what's what I consider is important and it it turns out that language doesn't rely on those in effect language only invites attention of of the listener of the reader and that's it and then other cognitive abilities kick in and and but in language those details are not encoded and that's very important and what we have in elephants is they they lack those additional cognitive abilities to compensate for the lack of that information and that's crucial because of that we will require uh multimodality and so on and in fact we will need even to uh like uh fine tune those knowledge to each agent individually so there's there's a lot of work to do with this approach and unfortunately we are too too invested in the old ones and the this like Japanese proverb saying that when you are on the wrong train you you you'd better to go off quickly and turn back unfortunately we are not going off the wrong train so far hopefully eh I will at some point succeed in this mission and the uh some uh company hopefully in big company will will get interested and they will at least try to check if this approach will will be in some respect better and promise uh better results uh be more efficient uh yeah in terms of resources more meaningful in terms of answers and the value delivered reliable as well so there are many promises in this approach and I hope it will see the day soon it makes me a lot of stuff that Yankees says about and he's backed off slightly from this I believe but stuff that he says about LMS not being as smart as a house cat cause you know on one level that's very wrong but on another level it it is very correct cause you think about what a house cat needs to survive in the the internal models that it needs to be able to create to navigate the world right like you know my cat jumping up onto the the counter like he he's created a pretty sophisticated internal model of the physics involved in making that jump he just doesn't have the you know the other aspect is the language to describe those those models externally um yes but yeah right yeah it's very fascinating imagine a cat injured leg yeah all those internal parameters no longer work it will need to adjust internal and return to the original one when the leg is healed so and imagine writing from scratch all those models it's it's insane and it is yeah and it is not necessary because all all we have is take into account additional constraints right selection from available options respecting relevant constraints so if one leg is injured so I will move in different manner I will not jump I will not rely on the injured leg and so on so very simple and it it it corresponds to the favourite definition of Michael Levin achieving the goal er no matter what er constraints are in front of you so you you apply different er algorithms and you you find your way so to say yeah and so obviously it's being talked about so much right now is this idea of you know deep seek all these different things that the training the models fine tuning the models um with your approach what does training a model really look like it it cause it's not as much here's the data here's how to work with that data interpret the data it's much more here's a it's like much more multi modal almost like here's the here's the cat right we we love cats yes um and here are all the different parts of it I like I'm trying to envision what this looks like to train a model this way so basically uh what we will have to do uh we already have a lot of information encoded in texts so it is okay to go with text so far because I I don't do perception so I I cannot like uh provide accurate algorithm on how to uh do vision audio smell taste and so on so uh so we will need to work on that and I believe that these principles will apply there as well just okay we need to take into account er relevant constraints again er so er but er what I'm saying that we need to construct those er specialization trees so if we are trying to affect uh for instance uh location of of an object so we can kick it we can uh push it we can uh move it in in in any way with air with some pressure or we can ask someone else so many many approaches and we know we we need to know how different actions and their intensities or other parameters apply and lead to which results so and that what I mean is by by knowledge how so recipes so we we need to work out those recipes it is quite possible to collect a lot of in of such information from texts of course we will need to work out the details as soon as we have the need for for for this additional information and additional details like in different kinds of robots and machines and so on and and it is possible to then when we processed all those texts and filtered out irrelevant ones like hallucinations or um hypotheses that went wrong and were refuted and so on so but then we will still have some gaps in knowledge and that's when we in fact it will be lalaams not lalaams this novel AIS they will instruct us on how to experiment in order to fill the gaps in in this refined knowledge hmm interesting so and it will be possible for this new type of AI to like lead us in generating new knowledge and bringing us closer to to new discoveries really so and that's definitely different from what L L amps can do for us at the moment and as I say it will be done in a pretty efficient way without so much resources wasted yeah yeah that's what yeah I wonder and it will be really one time training because er we will not need to repeat the same joke twice even once that tree is built yeah the only thing uh that may be interesting uh is uh for instance if uh different authors mentioned the same joke then we will probably need to know just references that that author and that author mentioned it okay so for for references only or that that joke was mentioned 20 times and that joke was mentioned 100 times so okay that that one is more popular okay just to know the basic statistics and that's it or it it applies not only to jokes I hope you understand so yeah yeah that makes sense wow fascinating but again it's one time training and sometimes it's not even training in just updating a a parameter for instance if you had have some bunch of friends and one of them betrays you you just update this property unreliable and that's it next time you need to ask someone you do not retain all your historic historical data right you right you just select uh from the set of reliable friends and that's it that friend is no longer reliable so he's out he's filtered out so the same process applies so selection okay selection with respect to properties yeah maybe you should build your own uh your own 20 questions game Alex ha ha well there there are many many really uh versions of of it online even and by the way er that game was er interesting to famous American philosopher Charles Charles Thunderspears but he he failed to understand the full potential of that game again because Pearce was more interested in rigorous logic yeah and and and that game it is not so about rigorous logic right so it's yeah it's about and uh by the way I I start to call my approach semantic binary search after that game because essentially it follows the idea of binary search it divides the the remaining uh like the the currently remaining options roughly into half but it relies on different properties each time that that makes it semantic yeah and each at each level we have like uh different kinds of uh properties to continue with based on the previous uh question uh answers really so that teaches us a lot about er what properties to expect next yeah and and it's it's yeah this is this is I mean what I'm loving as I'm hearing this obviously we're called the a I Rebels podcast right because we we try to feature and highlight I I loved how my approach is uh yeah fine tuned for for your for your format well you're you're rebelling against the rebels ha ha ha ha which is just I I am rebelling so poetic I love it against current approaches so yeah right and I'm trying to like redirect the the whole the whole field the whole discipline into new direction because really we we don't want to uh to to stay on the wrong train yeah I mean there's something there cause we're already seeing it with the currently it seems a thought process right like the answer is more data oh okay we're running out of data let's use synthetic data okay now Deep Sea is doing it at a fraction of the cost and now it's gonna be a race to the bottom it's how far can we go with the current model I think it might take something like this a rethink of the logic and the system that we're using I I I think you're onto something I think something like this is is what it's gonna take to really get a I to that that true a G I level that a lot of us think of when we think of a G I yes and you already mentioned er young Liquan but the the problem is that a lot of experts call for a new paradigm right mm but they do not propose one mm hmm and uh for instance uh young Leon proposes his own uh version of uh what what can lead to a G I but it it is relying on on the same deep learning yeah methods and so basically we need a new approach so let's add one more layer to the neural network it doesn't work that way if you need radically new approach then probably you need to get rid of those neural networks altogether so right basically what what I propose of course you you should not remember the alpha go impact yeah right yeah but uh do do you remember that it relies on the research and effectively that's what I propose the only difference between Monticello research and my approach is that I propose a quick comparison of existing properties to check with which option to go next and not rely on this Monticello simulations really which which take long and the it makes the the whole process but again if we if we will talk about the approach to beat go then I will consider different kinds of alpha go algorithms so basically we need to consider options for the task and if we don't know anything better than that's the way to go so yeah yeah amazing well Alex thank you for coming on as we wrap up I've I've been wanting to ask you as we've been talking about this what is your definition of a G I uh so again I do not define AGI because it's a model of intelligence right so I defined intelligence so basically because a lot of folks try to avoid that topic because it's too difficult for them and for me it's not difficult because I know that intelligence relies on differences so my definition of intelligence is the ability to handle in all the forms possible to handle differences recognize them process them and produce them so that's it do do you remember what Heraclitus said about the river that no man can enter the same river twice yeah right because it's not the same river and he is not the same man so everything is unique but I continue that for some purposes some objects are interchangeable and we we should know the difference between relevant irrelevant interchangeable or different if we know that we are fine to go and reliant on simple comparisons fantastic well Alex yeah people wanna follow this this rebellious journey of yours what's the what's the best way for them to do that well I I have a Twitter account and I have substack page where I post my ideas and I am open for all sorts of collaboration or discussions and by the way I am even ready for pivot in my career so from software development to AI research so and you you can say that I'm pretty interested in in that so I can tell it will be definitely exciting period for me yeah yeah that's awesome we'll we'll drop links and we'll spread the word yeah thank thank you for that appreciate that really and so much for coming on and it was great discussion thank you for your questions guys yeah thank you