← Bald Ambition
Loading summary
Transcript89 lines
- [00:01]
A
Hello and welcome back to Bald Ambition. I'm your still very bald host, Mookie Smits, and I'm thrilled to have Ganesh Krishnan on board today. Welcome, Ganesh.
- [00:16]
B
Thank you. Muking. And I'm not bald, but hopefully I'm half of it.
- [00:20]
A
You know, I'm always jealous of most of my guests. So you, you are in that camp where you just rub it in. And if you're, if you're seeing us on video, he's. Ganesh has a nice bouffant making. Making me look even balder than I usually am. So Ganesh is CEO of AI Hello. And he does some interesting stuff with AI in the e commerce space. Ganesh is also a philosopher and a POV guy related to AI, its interaction with our culture, its levels of sophistic. And I can't wait to dive into AI in these various forms. Pragmatic application. We've had many guests on over the past couple months doing all sorts of stuff in AI and Ganesh is doing some interesting stuff. And I'd also like to hear your vision of AI because it can get very confusing very fast. I think it's a mixture of myths, optimism, and doom, saying, and, you know, let's roll our sleeves up and talk AI.
- [01:34]
B
Right. Thanks, Muki. You're right about the mix. And it's a mix of optimism, doomsayer, ignorance. There's a lot of mix into it. But what I would generally advise is keep the doomsayer on the side because it's one of the first few times, you know, we have knowledge that is from the specific people that everyone can access across the world. When you ask a code to create an algorithm and you don't understand the algorithm, it can fetch it from his database and it can build out the algorithm for you. It's not thinking of anything new. It's just bringing stuff from other people that were hidden behind those in front of you. So I wouldn't go as far as call it as AI. It's just an LLM. AI is probably the future when we build ASI or AGI.
- [02:17]
A
Well, AI is the general bucket of artificial intelligence, which is not new. AI as we understand it, machine learning, even some deep learning, goes all the way back to the 1970s, right. Even earlier, Marvin Minsky, a lot of the trailblazers. And AI has had its ups and downs, right? They make big promises they don't come to bear. And then it has a winter, an AI winter. And then when it warms up a little bit, it does so with pragmatic application. We've seen that happen with machine learning that's done some amazing things. And the big revolution happened with OpenAI chat GPT November 2022 only, only three years ago, right?
- [03:09]
B
Yeah. Feels like a lifetime. And I hope this it's we don't have an AI winter and let's hope this time, you know, we continue this. The exponential growth that we are making when it comes to, you know, LLMs or transformers, there's obviously we're probably going to hit a ceiling. We've been talking about hitting a ceiling for the last four years. We haven't hit a ceiling so far. But I'm very confident of, you know, where we are going. And hopefully with the change of technology, not the transformers, using some other methodology, we'll probably get to at least human level intelligence. And that is what I'm very optimistic about.
- [03:46]
A
Human level, when it's very hard to
- [03:49]
B
pinpoint, I don't think anyone on this world can answer anything. What I say is you can't even decide if it is intelligent or not because when it reaches a point, most of the play what it does would seem kind of stupid. But in the end it wins. You know, it's like if you're playing stockfish on chess, you can see some of the moves it makes is really, really absurd. You think why would it make such a stupid move? And then few moves later you realize it, oh, that's why it did it. So it's. And I always have this analogy of a cow standing in the middle of a road and looking at humans, you know, crossing at a crosswalk to it. You know, we humans are probably stupid. I don't know why we are crossing in traffic. It doesn't understand that there's a crosswalk. It doesn't understand there's a signal, doesn't understand those things. So when we do reach, you know, human or above human intelligence, I don't think we'll be the ones to judge it.
- [04:36]
A
Well, let's try to define our terms because we've got a lot of swirl here. GPT is generative, pre trained transformer. So it's generative in that it actually creates content. To your point, it's not creating anything new because it's based on pre trained models. It took a look at everything available on the Internet, some things that were actually copywritten and technically not available to it. So that's a whole separate story. They chopped it up into tokens and then they did did a lot of math to show the relationships between the tokens in terms of their utilization. And then the transformer takes all of that data, petabytes of data and data analysis, and converts that into a response which appears to us, to your point, as, as being almost sentient. It, it was astonishing when it first came out. They've only gotten better, but they do have limitations. And I would, I would venture to guess that the IQ of these machines or the level of self awareness is still very, very low. They are ultimately very sophisticated, deterministic, mechanical engines of inference.
- [06:04]
B
Right. And one of the most, you know, frequent question I ask in podcast is you probably, you know, had an education in school and you read the, you know, the mainstream media as well at both times. How many times did you disagree with both of them? How many times did you disagree with the mainstream media? Even though it's not in your training, you know, if something new erupts, then you probably, you know, kind of question it. Is it true? If the hantavirus, you know, erupts, your first thought is, is this true? Are they trying to do something? Are they trying to, you know, scare me to something? You question those, you know, authority, and even though it's not in your training, you know, you can invariably question it. If you have been taught all your life, you know, that the sun is red and you woke up today and you looked at the sun is yellow, then all of your life training is gone out. You've been trained on it for, you know, whatever many years, you would throw it out and then you pick up on this new because you have more authority, looked at the sun, you verified it's yellow. You want to use yellow from now on. So the question that we ask LLM is can it rebel against the training data and then say, no, I've been trained wrong. Right? It can't.
- [07:20]
A
Yeah, but, but there are claims that it does. It's headline news that, that, that versions of Claude try to bust out of the lab, that versions of the LLMs are scheming. We had the Clawbot and the social media platform that was created for the agents who were talking smack about their human creators. There's a lot of, frankly, in my opinion, BS surrounding these autonomous agents.
- [07:57]
B
Yeah, I completely agree. It's just bs. And I don't think they have conscience. I don't think they have intelligence. I don't think they can rebel against the training. The question is, do we have enough data that we can call it as intelligence? But, you know, like, the question that you need to ask is LLM is can you invent the future? We humans are Always moving, creating new things, creating new future. So if you train an LLM till 1970s, can it invent the Internet? Can it invent the email? Can it, you know, if you use till the 1970s data, can it invent the future and bring us to this level? And quite frankly, the answer is, no, it won't.
- [08:34]
A
Well, it could iterate possibility, and it can do it much quicker than the human mind can. And I think here's where some of the confusion lies, that they work incredibly fast. I mean, you put a query into even the frontier model that you get for $20 a month, and in less than a second it'll perform what was the equivalent of a computational miracle from only a few years ago. And they're doing all sorts of computations. I don't know if you're familiar with Mathematica, Stephen Wolfram's program. It was amazing at doing calculations of fairly sophisticated mathematics. And now the most basic bot can churn out integrals, do differential calculus, perform statistical analysis. And now with the help of the agents, you can assign them increasingly sophisticated technologies, tasks, go off and do things. So these capabilities are incredible. And then if you're looking at predictive modeling, so you bring up the example predict the future, well, that's analogous to predicting the outcome of, let's say, a molecular rearrangement of a drug or some kind of genetic profile of work that they're doing. So is this intelligence, or is this the ability to crunch an incredible amount of numbers and calculations and then extract a select set that might be hopeful. It's almost Darwinian, right?
- [10:13]
B
Yeah, for sure. It's a, you know, it's an interesting question, like, where do we define intelligence? Where do we draw the line on consciousness? How do you exactly define if the machines are sentient or not? And most likely they're not. I don't. You know, in my opinion, they're neither intelligent nor sentient nor conscious. All of it is almost zero.
- [10:33]
A
Alan Turing defined intelligence in a behavioral way. So the Turing Test was the classic example of that. And it held up until actually fairly recently. So the vision was that you have a typist, a human typist, on one side of a wall, and. And they're communicating with the other side of the wall. There's a hidden operator, and it's. It's like a teletype machine. So you type in, hello, how are you? And then from the other side, there's a response. I'm, well, how are you? So the Turing Test was, how many iterations of this conversation would have to take place until you realize that the entity on the other side of the wall is not human. And Turing's response was, if you keep sustaining this conversation long enough, then you define sentience by your inability to differentiate whether the operator is human or a machine. Now, the challenge with that is that the chatbots do this so well that you could sustain an hour long conversation with Claude and be quite convinced that you're actually talking to somebody and the evidence is there and that. And that people are falling in love with their bot. People are talking to their bot continuously throughout the day. I know several people who consider the bot their best friend. And they share everything from philosophical conversations to religious arguments to practical business application. The Turing Test is thrown out the window. Does that make the bot sentient? Does it make it human? And I would say no to the point you're making now. So what is the new test for sentience?
- [12:30]
B
You know, I could talk on hours for this. And it basically starts with something I read today that to use the AI correctly, you need a very high level of IQ to use the AI correctly. And if you put the Turing Test, you know, and I should probably build it out, you know, just for testing out sake. If you test it out, if you ask the right questions, it's probably very easy for you to figure it out if it's a bot or a human. And I'm sure most intelligent humans can figure it out within a few questions if it is a bot. Even now, whether it's, you know, Opus 4.7 or, you know, OpenAI 5.5, you can easily figure it out. You just ask it a few question, you know, a few intelligent question. The responses should be not quite sufficient enough to find out if it's a part of it's a human. But again, you know, this goes into the philosophical side of it. Most humans, I think, are content with the lives that they're living. It's average life. They have average, you know, thinking, and they want a 9 to 5 job. And what they think about and the questions they ask. AI fits within the realm of what we humans have always invented. So they just want, you know, a recipe for a banana pie or recipe for a banana bread. These things AI can spit it out, but if you want to really go, you know, beyond the boundary and push the human limits, then the AI starts failing, right? Those are the parts. I mean, of course, sometimes it succeeds, but sometimes it fails. But it cannot go and ask, you know, how do I break a quantum encryption? How do I build a quantum computer if that was the case then, would have already built a quantum computer. Our progress would be significantly faster. But what is happening is that the human limits have already been achieved by certain humans and the rest of the world is coming up to those limits using AI. So there's a lot of people, you know, going from average to really intelligent to the limits of human intelligence by using AI.
- [14:29]
A
I'm a little confused by your, your, your taxonomy here. You could interact with the most average human of relatively average iq. You know, there's the joke that, you know, George Carlin the comic made the joke that think about, think about how dumb the average person is and then realize that half of them are more stupid than that. So, but what does that have to do with sentience? Some of the most cognitively compromised humans are sentient. They're self aware, they're, they're flexible. You can give them various tasks in the real world and they can perform them much better than any robot that's, that's functional right now. So we're mixing the terms. You don't need to figure out quantum gravity to be sentient.
- [15:29]
B
No, for sure, for sure. But the question again is like, what is the minimum IQ required to get a driver's license in us Right? Do you know that, you know, do we know the answer to that? What is the minimum IQ required?
- [15:43]
A
I mean they don't, they don't give you an IQ test to get a driver license.
- [15:48]
B
But if you have to guess, you know, like, of course, if you agree, you know, like what you use it on cognitively, you know, like impact, then of course you can't drive. But at what level do they say that you are impaired enough not to drive the car? And I did a bit of research on this and apparently it's around IQ of 80 and 80 and below. They don't give you the driving test 80 and above, you know, you can, you could get the driving license. But again, it comes back to this. If a human at IQ level 80 can drive the car really well, go through everything, why is it taking so long for a self driving car to be even reach this level of 80%? You know, like even now, self driving car, how many do you see? It's very hard to see, you know, like, and they don't handle all the edge cases which humans can do. So to answer your question, you know, like, you are very right. How do you draw a line between why is a human sentient even though their IQ is low and why is a machine not? These are all philosophical questions. You know, I Mean you can spend hours talking about it, but the answer is machines are definitely not sentient or conscious because I don't think they are more intelligent than, you know, the bottom 10% of the human population.
- [16:57]
A
Once again, you're mixing cognitive sophistication with self awareness and the ability to interact in the real world. So I don't think sentience has to do with levels of cognitive sophistication. I think to your point, there's a cutoff. But, but someone can literally be dumb as hell and at the same time have complete self awareness on a level that the architecture of the cure of the current LLMs might not even be able to get close to.
- [17:34]
B
Absolutely.
- [17:35]
A
I think that's what confuses people. So you don't need to understand quantum mechanics to be a sentient being and have AGI. And conversely, you could have a dominant LLM that crunches an unbelievable amount of computations and figures out quantum, quantum gravity because it's done 100, septillion variations of the Einstein field equations and compared them to Heisenberg's equation and then it, it popped out.
- [18:14]
B
I think the chances of an LLM discoding something new that humans haven't discovered is pretty slim. There might be something that humans have missed quite obviously, but to be at the forefront of human capability and to invent something, I don't think an LLM or a transformer or any current model is capable of doing it.
- [18:35]
A
But they've already done it, Ganesh, and they've done it before the LLMs. So let me give you a concrete example. In the field of healthcare, they applied basic machine learning to the retinas of the human eye. Have you heard of this, this case study? So men and women have different retinas, how they're structured, but we have been unable to identify the differences. So what they did was they took photographs of tens of thousands of human retinas, the branching arteries at the back of the eye. And then they assigned a gender to each image. You know, this one is male, male, female, male, female, male, male, male, female, one after the other. And this was one of the biggest revelations. And this is pre transformer. This is raw machine learning. It did an association between the visual images in aggregate and the binary conclusion of gender. And then they gave it a test. They gave it a picture of a retina and then they asked the machine, the prelim machine, whether it's a male retina or a female retina. After tens of thousands were identified in cross reference, their accuracy was over 98%. So the machines had learned machines can now identify whether a retina is male or female. That's entirely new. We were unable to do that. And the other irony of this is we still don't know the difference because the machine is incapable of communicating to us how it learned that difference. So this kind of encapsulates the first part of our conversation, frankly, which is it's got nothing to do with IQ and cognition in, in the human sense. It has zero to do with self awareness. It does unbelievable amounts of calculation in a brute force way. But thanks to the machine learning architecture and the developments in deep learning, even prior to the LLMs, it's doing things that we cannot do and it's reaching conclusions that we were unable to make for sure.
- [21:13]
B
I think what I need to make a distinction is between real machine learning or machine learning that existed before LLM and transformer technology. And how do you define as intelligence? If you take Stockfish, it can beat all the world's top chess master chess grandmasters right away, you know, even blindfolded, whatever it is, you know, like you can play really, really, you know, absurdly good. There is no human that can defeat it. It's the same with the retinal things. They can recognize the retina. They can do it. Would you call Stockfish to be intelligent? You know, it does his job. Does this work? It can beat even, you know, one of the greatest players, you know, Magnus Carlsen. It can do all of it. Is it really more intelligent? But beyond that question, you know, my issue is with the Transformers model. His issue is with the LLM model. It's good for self attention. It's good for, you know, predicting what word is next. But I do not think Transformers is what is going to take us to the next step. When it comes to AGI or when it comes to ASR. We need a completely different algorithm. We need something like cnn. You know, what you talked about the retinal. We need to talk about something like what is the Stockfish? We need some kind of, you know, algorithms to take us to the next step. And it's definitely not the Transformers.
- [22:27]
A
If you look at human evolution of our minds, you know, we have this fatty gel in between our ears, weighs only a few pounds, and has created all that we know. It's also a universe within a universe because how we perceive the universe is quite different from how the universe actually is. So it's. It's a miracle. It's amazing. So we're not near that with an LLM. We're impressed that it could Write an essay or do an analysis. But the lived experience of the human in the universe, creating a universe inside of it, is not an LLM. And I think that our technology will get to our level of sophistication in a similarly evolutionary way, just like we're not even close to understanding how the human brain works. I think if we do arrive at AGI, real AGI, that the brain in the box, maybe with robotic extensions, is like us, has a sense of consciousness and sentience is. Is intentional and motivated, maybe even has emergent emotions. I think it will be a consequence of an evolution. So I think the road to AGI isn't programming, like programming code. It's setting bots loose in a simulated environment where they could have generational evolution in some kind of ecosystem, solving problems, being forced to fight for their survival. So that you have a mechanism of evolution, you introduce mutations, and you kind of mimic the Darwinian evolution of human consciousness, only you do it in a digital realm. I think that's the way to go.
- [24:17]
B
I used to think like that. You know, that was what I was thinking. You know, they have a kind of algorithm, genetic algorithms, which can fight against each other and choose the winner. But the more I get into philosophy, you know, the more I see that our consciousness, you know, like on one hand, there is a school that says our consciousness is a quantum effect because our neurons work on a quantum level where there is superimposition and there is no, you know, fixed place. There is no nothing fixed. And that is where, you know, our consciousness comes from. It's nothing to do with digital. It has nothing to do with physical thing. Our consciousness is entirely quantum. The other school that I'm really interested is that our brain is just a computing machine. Consciousness lies somewhere else in the universe, and we just tap into this consciousness. That's why we are all one. And we've been saying it for thousands of years. You know, like Plato, Socrates, the ancient Hindu philosophies, they all say we are one. By that, they don't mean say we are one physically. It's just the consciousness is one. We all just tap into it and everything is just energy, right? Our brain, our body, our physical thing, it's just energy vibrating. So their philosophy, you know, which I kind of, you know, I'm looking into it as well, that consciousness is somewhere in the, you know, somewhere in the universe, and our brain just taps into it. And, you know, we. We function as. It's a lake, it's a reservoir of consciousness, and we tap into it. I don't know which one to believe.
- [25:40]
A
I think they're, I think they're all cop outs. They're all excuses for a lack of understanding. Remember the ancients had the idea that there was a turtle, that the earth was on a turtle. And then, and then where was the turtle? Well, the turtle was on another turtle. And then there were a million turtles that were holding up. And then what was that last turtle standing on? The turtle was standing on an elephant. And then what was that elephant standing on ad infinitum? The arguments that you mention, that the human brain cannot be emulated because it has a quantum foundation. That's a turtle. Because the sun uses quantum tunneling, we think, to enable fusion reactions to happen. When two protons get slammed into each other, their repulsive force is so incredible, increasing with distance, that you actually need quantum theory to show how their superimposition is what induces nuclear fusion. Does that mean that there's anything transcendental about the sun's fusion? No. If you introduce quantum theory into human consciousness, that's nothing that can't necessarily be emulated for the same reason that it's just a physical property. And if you introduce some kind of spirituality, well, you could believe in that. That's true, but I don't think it's necessary for us to be conscious and it turns into a religious argument more than it does a rational, scientific evidence based conversation. So I'm, I'm skeptical of that. I'm not just, I'm not here saying that consciousness is 100% deterministic, that we're just machines. There's a lot that we don't know. But, but I, I don't believe that it cannot be emulated. I agree with you that we're nowhere near emulation with the transformers now and people are being deluded that we are. But, but I don't see any, any potential hurdle for eventually getting there with the right application of the right technology. So that makes an optimistic materialist. But at the same time, again, I'm skeptical of these circular tautological arguments that we can't emulate sentience because quantum theory, we're spiritual beings. These inherent physical limitations, I just don't
- [28:18]
B
buy it could be, you know, like, I mean, less than maybe a few hundred years ago, we looked at, you know, thunder and lightning and thought it was the gods getting angry. And now we know what exactly it is. So probably, you know, we'll find a way to solve the consciousness problem at all. But you know, like, it's been thousands of years and I hope you're right, you know, for our sake, that we can kind of solve it, but we're not even barely. We're badly scratching the surface of it even know that I. Stop.
- [28:47]
A
Well, let's talk about the doomsayers, because you're saying that you hope that we can get sentience in a box and there's a lot of smart people saying that the second we do, we're all going to die. So there's that very popular book that came out, which is actually a very good book. It's good in the sense of making very good arguments, is that if someone builds it, everyone dies, which is all you need is one and we're all toast. So what makes you so optimistic that AI is benevolent and a force of good for humanity?
- [29:26]
B
Well, it's like saying, can a calculator take over our house and then hold our pets hostage? It's just an algorithm. Right now, once we discover consciousness and we tap into it that, you know, we could ask this question again. But I just, you know, like today morning I had a call with the Alibaba team, you know, the Deep Seq team, and we were talking about, you know, how to use Quinn correctly, you know, how to use the API, and we were just talking about interaction on how to make the AI better. Something the team has insisted, you know, and it's, it's weird coming from China and of all the place, of course, you know, it's not weird, but it's a normal was. They insisted that we make AI as cheap as possible because people from, you know, like low social economic countries like Cambodia, you know, like Africa, those people, they deserve access to the AI as well. Just because they can't afford it doesn't mean we should, you know, ignore them in this AI ways. And they were trying to make the API as cheap as possible. And they say, whatever price you see in the future, we're going to keep on decreasing it until everyone in the world can collectively use AI just like they use water. We want to make it the right. And it's interesting to hear that topic. It's interesting to hear them, you know, and I genuinely believe they genuinely believed it as well. And I told them, you guys are cheap enough, but we would like to, you know, make it as cheap as possible. So I think what I would like to see is the positive side of it. For anything. You can look at the negative side of it. I grew up in a generation where there were protests on the street because people were saying the computers are going to take away the job the computers are going to take over the world. The computers are going to destroy all of our lives. And there were huge, huge protests. So, you know, we never got computers in our, in our country for a long time until one of the leaders decided to put his feet down and then said, okay, you know what, we're going to get left behind. Let's use the computers. And then, you know, like we moved on, right? I mean, of course I'm talking of India and now, you know, there's a lot of, you know, tech and computers and coming through. But there were lots of protests to it. And the doomsayers for most of the time, they always look at the wrong side of life. We should look at the right side of life. Right now. We're not tapping into consciousness right now. There's no sentience. It's just a glorified calculator. It's just a glorified vector calculator that is going to, you know, make AI and knowledge as cheap as possible for everyone in the world.
- [31:49]
A
Access to all the world's information aggregated almost in real time and it's providing enormous benefits and everything is a double edged sword. We're nowhere near sentience. So even the risk of it taking over is probably hyped. So I agree with you. Where I disagree with you is in the inherent altruism and benevolence of the Deep Sikh folks who it almost, it makes. When you were saying that, I was internally laughing my ass off because frankly, the real reason Deep Seek did what it did was because the Chinese really didn't have sufficient efficient access to the super fast Blackwell chips. So if they could, if they could, if they could train an LLM and make it portable on a laptop and do it for 1/10, 1/100 the cost, they would own the AI race. I don't think they were talking about kids in Africa getting cheap AI. I think they were thinking about dominating the market. And I think it's terrific PR when, when companies blow smoke up our ass like that after the fact, I just have to grin. I don't blame them. I would do the same. But, but what's driving the AI race right now is not altruism. It is money and power. And that's what freaks a lot of people out.
- [33:15]
B
I agree to that. You know, it was kind of. But when you talk to the Google team and you talk to the Deep SEQ team, what they put out and what they talk about, what they're doing behind the scenes or how they're behaving you can see a complete difference in it. And I know, of course we grew up in the western world. Of course we trust our own people more than from a Chinese model.
- [33:37]
A
I'm not blaming the Chinese. I'm saying their motivation is the same as our motivation, which is to get rich and take over the world.
- [33:45]
B
Probably. I won't disagree with you on that. You know, like they probably have it. But if I had to choose between, you know, Gemini, Google, Gemini having all my data and Deep Seq having all my data, I would give my data twice to the Deep seq.
- [33:58]
A
Well, that's because it's a local LLM and this illustrates the other thing that people aren't talking about enough is that there's an extreme tension between building enormous data centers, gobbling up giga gigawatts of power and having just a very functional good enough LLM that literally sits on your laptop computer with no access outside, which contains all your data, which is safe, secure, very cheap. You don't even need to pay the Frontier models licensing fees. You don't need to pay them. By the token, your own little LLM is cooking on your own machine, on your own server. And Deep Seek paved the way for being able to do that. Very affordable, very local. And that is a very meaningful competitive paradigm to subscribing to a Frontier model that is, that is sucking the juice out of cities across the world for sure.
- [35:06]
B
I think my number one gripe against Facebook and Google and all these people is they go on a banning spree for anything because they are the thought police. And if you do anything wrong, if you ask the question wrong, there is no way for you to dispute that thing as well. If you get banned and I got recently banned on Facebook for something that I didn't do, I won. One of the largest Facebook group and one of the other moderator posted a question about Free Tibet or I don't know, he posted about Free Palestine. I don't know what he posted about. I didn't even get a chance to look at it. Facebook banned my account and banned all the admins on that account. There is no way to dispute it. There is no way to send your answer. And 16 years of my Facebook data is gone like overnight. There is no way for you to go back and then ask them why did you delete it? There's no way for customer support, nothing. If you download llama 4 and you ask the question wrong and you know the Facebook team thinks this is a danger, they could ban everything for you. If Deep Seq bans me, I don't care. I have Quan, I have something else. I do not care. And this is my primary reason, not that I can download it. Facebook, Google, they all have too much of your data and they are your daddy right now. They decide what you can say, they decide what you cannot say. If you criticize, you know, China is good. If you say anything, you know, positive about China, they could ban your accounts. Imagine losing your Google accounts. Imagine using all your, you know, Gmail calendar and everything for the wrong reasons. And there is no way to dispute it. It makes it, you know, like a very, very suspicious and it's. Yeah, it's a bit scary as well.
- [36:44]
A
Well, the Deep Seek model can be biased in its own way. So you know, the stuff stuff is programmed. Part of the obsequiousness, you know, blowing smoke up your ass that's prevalent among these bots. So you use them. More is baked into the, almost baked into the architecture. You can't get them to stop doing that. To your point, censorship is a problem. And an even bigger problem I think than censorship is selective response. So when you get a response to a prompt through the transformer, it goes through various filters and processes and it has as millions of different ways it can respond to that query. The selection criteria could and sometimes is politically and ideologically motivated. And Elon Musk's Grok was an attempt at counter balancing what would be construed as a left wing LLM conspiracy of poisoning and indoctrinating the people with Woke. Woke ideology. Elon Musk actually he says it, he might not entirely believe it because he's hyper competitive. He wanted Grok to dominate. And then Grok went in the other direction so that it responded as, as Mech Hitler, as a glitch where a programmer just said, Elon said let it respond any which way. And then the glitch became a product feature. So there's this bigger issue not only of being kicked off a social media platform by a major platform provider, but it's the influence that the LLMs have on the information that they're cobbling together for you. So for example, if you go into Deep Seek in China, type Tianan Square, good, good luck with that. And then, and then that a lot of countries rewrite their history and we can't necessarily blame them. Look what's happening in Texas and Florida. We're rewriting the rewritten history and to teach our children. But the Japanese have been huge history deniers since World War II. The opposite of the Germans. And now that we have this technology, the mechanism of control through information has never been more intense. And that you bring up a good point that the real risk of the bots might not be just taking over the world and killing all of us or even taking all of our jobs, but. But manipulating us with propaganda, Correct?
- [39:31]
B
That is absolutely right. And further to the point, you know, like that's point number one, it's, you know, polluting the propaganda. The second point which I wish to make is that if Deep Seat team, you know, bans me tomorrow, I have nothing to do with them. I can just cancel it off. But Gemini Grok, you know, Facebook, they all hold too much power and too much data over our lives. I do not want to give them any more power. That's the reason I would give the power to someone else. So if something goes wrong, at least, you know, I don't care.
- [40:02]
A
Yeah, that's another benefit of the local LLM. So if you get an Nvidia Nemo, right, You get Nemo on your desk. To your point, Nemo is not going to necessarily piss you off, but if it did, you could go to Deepsea. And if the market expands and more and more folks adopt the technology, it becomes commoditized. So local LLMs become commoditized. You're no longer dependent on the huge frontier models and their enormous trillion dollar market cap companies that essentially at the end of the day, just control your ass.
- [40:42]
B
Yeah, you're not immune from propaganda. I mean, the local alliance still have, you know, all of it, but that's something you have to use a critical thinking and know which one is white and which one is wrong, not where. Is it hallucinating, but is it not hallucinating?
- [40:54]
A
Yeah. So these are, these are hopeful points. And it also shows opportunities in the marketplace. We've been spending most of our time with prediction models, which is the LLM architecture. But what about the robots, the reinforcement learning and we're talking about AGI. I think a central part of AGI is to get the brain out of the box. You give it sensory organs. Usually people think about a robot as an anthropomorphic two arms, two legs, a head, a torso. But the robots of the future are not going to look like us. They're going to take on every pragmatic shape you can imagine to do stuff in the real world. Look at the drone warfare in Ukraine. They're AI in ways. They're swarming like agents. They're increasingly sophisticated and they're there for a function. So that's a whole separate parallel interwoven thread of applying AI in the real world. And I think the two are going to meet if they haven't already, the predictive models and the reinforcement models. And that's where things I think will get really exciting and also a little bit scary.
- [42:11]
B
So interesting that you bring it up. So the Hal 0 part where we are talking about is reinforcement learning is without transformers and it's not an LLM that you can download because it sits on a huge repository of data and the vectors and the weights are bound to the data. So whenever it talks, it is backed up by the data that we have in our database. This is where we are going. And it's not good for logic, but it's good for really knowledge, getting knowledge out, knowing everything. And we can be 100% sure that it's not hallucinating because it's backed by data. It runs on reinforcement learning, at least the last step, where it talks to you and where it sees what kind of data that it has to get is based on reinforcement learning. This is where we're going towards. We are a bootstrap startup still, you know, a lot in the works, but hopefully, you know, we release our first API within the next couple of weeks. That's where what we're aiming for. You ask it a question, it can give you an answer with zero hallucination. And that's our hall just to give
- [43:08]
A
a little context to the listeners and viewers. You have AI, hello. Which does its crunching for paid advertising on e commerce platforms. You help people out that way. And then you have Hal0AI, which is another startup that you're leading. And it sounds like it bypasses the conventional LLM architecture to use reinforcement learning with data.
- [43:38]
B
That's correct.
- [43:41]
A
Can you tell us a little bit more about that? Because that seems to be an oxymoron.
- [43:47]
B
So what we do is we first get the data in the right format and we use our internal database and we put all the data, we separate it, we tokenize it, we put it in the right format and then we use reinforcement learning to find out what kind of topics to look up based on the question. And this all comes from, you know, like to give a bit of a background is I started my life as studying to be a doctor and I wanted to, you know, go into the brain side of it. I learned a lot of, you know, neurology. And the most interesting part of, you know, of the human brain is whether you go all the way to an earthworm or to a fly or to, you know, mosquito, you can see they're all bilateral, right, they're split into left and right. They're exactly split into left and right. Because the primary function of a brain is just to decide if you want to go left, if you want to go right or if you want to go continue. That's a brain. And then we know, we started building layers and layers and layers on top of it. So human brains is built on layers of, you know, knowledge. Each layer of brain gives an additional functionality to it. But at the core, what we call as a reptilian brain is the basic instinct. Like what we do, do we turn left, do we turn right, do we continue? And we build on based on that. So we used a bit of this human brain knowledge to build out our reinforcement learning. We put the data in and then first layer is what kind of data should we look up? The second layer is do we need to use some logic or do we query against the user? The third one is we use reinforcement learning. How do we talk about. And finally, of course, instead of a transformer, we use a diffusion model. We build out the entire paragraph using diffusion division is what helps you create images right now. So using this layer, kind of a technology we build out, but at the base of it, inside it is data and then, you know, it propagates outwards using this data to build out a good response.
- [45:44]
A
Well, what precedents are you using? Like, what other successful examples of this architecture are you adapting?
- [45:52]
B
Well, none. There are some Chinese models that try to emulate similar to this. There are people that do bits and part of it, if you think of OpenAI, there's nothing new that they created, it's just the transformer model that was done by Google, the reinforcement learning which already existed and they cobbled all of it. So, you know, the old philosophical saying is there's nothing new under the sun, you just have to connect it and build it. So what we are doing is then is being done in bits and pieces by other models. We look that up, we use it, we use the technology and then we try to connect together and build it out. The first project that we're going to release, I'm going to send you a free usage as well that looks up to your email and tells you how we can, you know, respond best. But when you respond to an email, there should be zero hallucination as well. It has to know exactly what you're doing, what you have written and build on that. So we'll see how the experiment goes, we'll see how the alpha release goes, how the beta release goes, and based on that, we know we'll know how successful we are.
- [46:57]
A
I'm just a little confused because as far as my understanding goes, hallucinations for the LLMs are more problematic for queries that aren't so raw data centric. So if you ask an LLM to run calculations similar to what I think you're suggesting, like analyze spreadsheets of data and then make some basic conclusions, I think the hallucinations are fairly low because it's kind of a deterministic agent based analysis. The hallucinations occur for more conceptual kind of prompts. Am I wrong about that?
- [47:41]
B
It depends on, you know, how you define conceptual. So if you ask it to know, like give me all the tax breaks that I can get in Canada And I ask ChatGPT and ask Kimi some of the things that it gets up is really hallucinating. It's like saying you can use section 14.3 to claim, you know, on your computer and you can use and those things don't exist. You know, like it's a, it's a random thing. And of course ChatGPT is much, much better right now. But what we need to keep in mind is that if you give it a spreadsheet and ask it an answer, it'll be very deterministic, of course, but as the context increases, the chances of hallucination keeps on going up, 10 questions down. You ask it a question about, you know, question number one, the spreadsheet, it'll be like, what spreadsheet? Like it won't even know. It doesn't even know what you're talking about. So the chances of hallucination keeps on going up as the context increases.
- [48:31]
A
Well, that's fair and to your point, especially the frontier models have gotten much, much better at filtering out hallucinations. But doesn't your architecture beg the question of its capacity to perform these more sophisticated functions just as a diffusion model? I mean, if it's just a diffusion model, then doesn't it have some inherent limitations that more conventional out of the box LLM wouldn't?
- [48:59]
B
Yeah, it's only time will tell. But the diffusion model is just for typing out what your thoughts are. So if I ask you a question like about, tell me about the latest iPhone model. What do you think about it in your mind, you construct the basic skeleton of your thoughts those are not hallucinated, and then you build a language model around it. So you build an abstract iPhone, the latest model, whatever it is, you know, like ips, the resolution, and then you Build a language around it. If you speak multiple languages, Right? I don't know if you speak multiple, you probably speak Hungarian too, right? Your IQ and your knowledge does not depend on what language you're communicating in. The basic of your thoughts, the basic of your brain, the IQ of your brain. The knowledge of your brain is independent of the language that you're talking of. But with transformer models, because of its tokenization, if you ask it a question in Hungarian, it'll be completely different from if you ask it in Chinese or if you ask it in English because of the different tokens it has been trained on.
- [50:01]
A
Well, that's because it uses the pre trained weighting.
- [50:04]
B
Yeah, but the knowledge is independent. The knowledge is dependent on the tokens. And what we're trying to do is we're trying to make the knowledge independent of the tokens.
- [50:16]
A
Yeah, I understand. But again, it's got. As an inference model, diffusion might have its own limitations, which you acknowledge and presumably you're working on.
- [50:27]
B
Yep, definitely. Diffusion is not to get the knowledge out. Diffusion is just to talk about the knowledge that we already gave it.
- [50:34]
A
The cat chasing its tail. It's a little bit word salad. Right. Because ultimately you're, you're doing inference. Well, good luck with it. It sounds like an interesting perspective and an interesting application of, to your point, existing technology. And the question is, when does a quantitative difference become a qualitative one?
- [50:54]
B
Exactly.
- [50:55]
A
You're putting together technologies that already exist, that have proven themselves sometimes independently of each other, and then holistically, when you bring them together, they're greater than the sum of their parts and something new can happen.
- [51:08]
B
That's what we are betting on.
- [51:10]
A
Right. Well, that sounds exciting. What are some of the use cases? Can you share business use cases for application that you envision?
- [51:20]
B
Sure. The first business use case, we already launched it internally. We call it Graph Inbox. You connect your email, you connect your meeting notes, the bot joins your meetings, the bot reads your email. And then you can ask you to create drafts. You can say, based on all the podcasts that I had in the last few days, create a draft and leave it over there. And then it keeps the data runs those HAL0 and then it creates a draft for those emails. Those drafts are without hallucination. It knows exactly what you talked about. It knows exactly what is your meeting link. It knows everything about you. We did release it internally. We are pretty happy with it. I think I plan to release it externally within maybe two weeks, three weeks.
- [51:57]
A
All right, well, very interesting. You never know. And then just the spirit of exploration is good, trying something different because everyone's doing the local LLMs, they're doing the Nemo plugins, everyone's a genius in AI. Few people are really pushing the envelope. So that's pretty cool.
- [52:19]
B
We are rewriting everything. And Yann Lapun, you know, from Meta, he said it very rightly. As I said, Transformers is not what is going to get us to the next step. It's something else. And everyone is exploring something else. We know we throw the hat in the ring and we say, okay, we'll make an attempt as well.
- [52:38]
A
Well, good luck. Maybe you are like Wozniak and Jobs in the garage where you take, you're putting together the parts that are already there, right. And just merely the application and the shift in the business model can be transformative.
- [52:53]
B
Yeah, fingers crossed.
- [52:54]
A
So that, that's exciting. And then when you're looking at customers, what kind of segmentation are you, are you focusing on? What kind of companies do you think you could benefit?
- [53:05]
B
Usually anyone that uses an email and goes on lots of meetings, you know, for us the first one is the in air. Hello. We have specialists that talk to customers who would like to record the call and then, you know, make a plan for that salespeople, you know, when they get on a call with customers, you know, they would like to know what is the action point, you know, what to put in. These are the two primary targets that we're going through. And you know of course the support specialist as well. When they talk to customers, they like to keep a track of all the emails and all the meetings they had with that customer.
- [53:34]
A
I've had various guests who offer similar solutions or optimizations for business use cases as yourself, what I hear repeatedly though, it's not so much the technology that gets in the way, but it's understanding what those business rules are and then onboarding it within an organization which is often resistant to change. So how do you, how do you factor those, those things into your relationships with clients and potential clients?
- [54:05]
B
So this functionality or this business is very, very common. You know, like there's lots of technologies that offer it right now, lots of startups that offer it right now, they're all wrappers on GPT. They just put the GPT and they wrap it up. And I used a plenty of them. And I can see, you know, what is the problem with it. Quite frequently, you know, they hallucinate, they don't look up the context, they don't know what is this person talking about, they don't know what Is your role into it and we're trying to fix it. If this goes through, it's not for me to make, you know, like this startup successful. If the draft that it creates is successful, if the knowledge that it gets is successful, then we know that our hallucination is working right well. And because we're not using a GPT wrapper, we're using our own internal, you know, health zero platform, then we know this kind of, you know, technology has a future. Although it does not have logic, it can use the knowledge and can create a draft very well using diffusion language.
- [54:59]
A
That sounds, sounds impressive. I think a critical component though is the discovery phase where you need to understand each business and their unique challenges. Right. So is there an aspect to your relationship with clients where you try to figure that out or do you consider. I had one guest who says that's not very important at all because the clients on their remedial level are all doing pretty much the same thing. So do you rely on that latter perspective to just assume that the basics are being messed up and then you go in with a new approach to the bot?
- [55:35]
B
Well, I'm not so much interested in making this startup a success as in making sure that the technology behind it is working. And once the technology behind it is working, then we know we can ramp it up and we can make the AI even more smarter for own use. But it's interesting part of it, you know, lots of people will dismiss it, lots of people will use it again. You know, our team, our internal team, we have around 32 employees if they use it daily and if they think it's useful. Right. All I need is one person to say, okay, this is useful for me. And I think it's, you know, it's something I would use every day. Then it's easy to scale up. All we need is one person to say it's useful. Then there will be other people like that. You know, regardless of how unique we think we are, there's always millions of people similar to you.
- [56:20]
A
Yeah, that's true. I think what I'm struggling a little bit with is the challenge. The technology, like you're saying, the hallucinations, the unreliability of the current transformer model to fulfill these functions. Or is it the other stuff that I've been talking about, which is difficulty for onboarding the adoption curve in terms of human behavior and being able to plug that technology into legacy systems that are already a spaghetti bowl mess of redundancy and inefficiency and this mix up of human interaction with SaaS functions. They're using Slack, they're using Gmail, they're using Docusign and Adobe, they're using Excel, they're using Salesforce. And they've spent millions of dollars and they've thrown it all into the mix and it's all just jerry rigged together. And you have employees who've been there for 10 years who know all the details about all the crazy wiring. And their biggest problem is in hallucinations of the GPT. Their biggest problem is how the hell are we going to fix this?
- [57:37]
B
Right. One thing to keep in mind is that most of the transformers have a hard limit on the context. And we reach the context level for most of it. If we use GPT wrapper, 10 email, 20 email and four or five meetings, the context is over. But when you talk of email, you know, enterprise level email, you have hundreds and hundreds of email with one contact, you have 10 or 20, you know, meetings with them. Can you reliably build a draft based on all the historical things? And most LLMs would run out of context by the time and they use just the previous email to create a draft. But if you go for this Hal 0, then it can go with ultimately as much context as you want with almost zero hallucination.
- [58:19]
A
Yeah, I mean that's the essence of reinforcement learning, right? It's scalable. So the predictive model is dependent on your pre training. And the reinforcement model is like Pacman. It keeps eating more and more data and presumably at least in principle, gets better and better at doing it with the more you feed it.
- [58:42]
B
By definition, yeah, it's very, very slow. But we don't care about it because it runs in the background. It takes a couple of hours to create a draft for an email, but that's okay. It's quality over quantity.
- [58:54]
A
How do you deal with the computational requirements?
- [58:58]
B
We have our own GPU cluster, so we have our own Nvidia, we have our own on premise GPUs. We run it. We moved also to Alibaba Cloud, which is extremely cheap, it runs on a few cents, so has been pretty, pretty good.
- [59:12]
A
Because I'm thinking what you're doing is potentially very data intensive and energy intensive.
- [59:18]
B
Yes. So we run our own gpu. It does. I mean the electricity bill is through
- [59:23]
A
the roof, but fascinating model, I guess. In summary, we had a winding, super fun conversation about AI principles and application and culture and awareness and sentience. But to be honest with you, the most interesting part for me is application of reinforcement learning in what's usually considered a GPT construct that's novel and you're actually putting it into play, which is very interesting and I'd be interested to see where you get with that. So very, very cool. HAL0AI. An extension of AI. Hello. Which is your flagship, I guess. And this is the new stuff, applying reinforcement, learning, essentially what we would consider the brain behind robotics to what's been up to now very predictive and suffering from all the challenges of that.
- [60:23]
B
We're going all in on rl. That's our main objective. That's the main bread and butter and we think that is going to give us zero hallucination.
- [60:31]
A
Hallucination I think is a problem, but I think it goes beyond that, which is improving inference and making it more pragmatically expedient.
- [60:40]
B
Right, Correct. Yeah.
- [60:42]
A
Thank you, Ganesh.
- [60:43]
B
Thank you, Mookie.
- [60:45]
A
Yeah, it was a really intriguing conversation. Like subscribe Share Ganesh Krishnan, thank you for time.