Sometimes, getting spammed leads you somewhere unexpectedly brilliant.
Agree?
If not, read along and see how my spam story played out well :)
One fine night, some random LinkedIn connection in the US added me (along with 50+ others) to a random LinkedIn group.
He created this group because he was super delighted to share the groundbreaking thing that Mr. Srini Pagidyala and Mr. Peter Voss (one of the men who coined the term “AGI”) were building.
The next morning, I woke up to 200+ LinkedIn messages in that group, and I had an early morning flight to catch.
Ironically, this random connection was not-so-random. I had met him at an in-person event in San Francisco in 2023, when I was jobless and almost homeless in the Bay Area. Back then, we connected on LinkedIn and shared contacts (of course, in the hopes that he would help me with my job hunt), and he really did try his best to help me out.
Some people you scroll past. Some people, you remember who showed up for you. So while most of the group saw him as a spammer, I saw his name and thought: let me hear him out.
That morning, after I boarded my flight, I went through all the messages in the LinkedIn group. Most of the group members were bashing him for adding them without their permission, calling it spam, and leaving the group. Meanwhile, I ignored them and instead went through every link he had shared in the group to understand his enthusiasm.
By the time the plane was ready to take off and it was time to put my phone on flight mode, I had already reached out to both Srini Pagidyala and Peter Voss to learn more about their revolutionary work with aigo.ai.
During my flight, instead of sleeping or listening to music, I quickly prepared offline notes with questions I would ask either of them in my first conversation.
I had a chat with Srini Pagidyala almost two days later. After my chat, I knew this had to be a full-fledged podcast episode for my audience.
I’m choosing not to name this not-so-random connection. He doesn’t deserve to be remembered as a spammer. He deserves to be remembered as my “Best Podcast Matchmaker”: thanks to him, I got to ask Srini so many questions on my list 🤗
Both Srini and his business partner, Peter, have been saying from the start that “LLMs could never reach AGI.” Srini published an article on VentureBeat about this, and Peter wrote about it right after LLMs were released. Something that the AI industry is only now starting to admit.
So I asked Srini:
📺 What did they know about LLMs from the start?
📺 What is the missing piece of AGI? How did they crack it?
📺 What is so special about AIGO that they’re confident it will lead to AGI?
📺 Where exactly is the AI bubble?
And many more questions…
He also spoke about a super interesting Chess Flip Test: LLMs vs AIGO (much like UNO Flip):
They changed the rules of chess. Flipped the king to a queen, queen to a king. Changed the rook to a bishop, and the bishop to a rook. Now let’s play!
Guess who fell apart within seconds, and who adapted to the new rules?
Or, well, you can watch our episode to find out:
Watch on YouTube; listen on Apple Podcasts or Spotify.
Episode Summary:
Srini shares his insights on the limitations of current LLMs, the architecture of true AGI, and how AIGO is pioneering a cognition-first approach to artificial intelligence that can learn, adapt, and serve humanity at an unprecedented scale. Srini shares his vision for AIGO, a revolutionary AI architecture aimed at human flourishing, and discusses the future of AI, safety, and societal impact. This episode explores how advanced AI can transform humanity's future positively.
Timestamps:
00:00 AIR Bites (Precap)
02:05 Zero LLMs, Zero Hallucinations: Is This AGI?
03:50 The AI Bubble: Utility Is Real, Intelligence Isn’t
05:26 3 Non-Negotiables of Real Intelligence
09:55 LLMs Are a Microwave, Not a Chef
10:46 Concepts Is All You Need: Inside AIGO’s Brain
15:35 Clone 9,300 PhD Cancer Researchers Instantly
19:50 The Chess Test LLMs Fail in Seconds
24:35 Language Isn’t Intelligence: Dynamic World Models
29:53 Is INSA the Fourth Type of Neurosymbolic AI?
31:35 AGI on a 3-Year-Old Laptop: 1,000,000x Less Compute
33:05 Stop Paying AI Companies Twice: Full Data Sovereignty
35:35 The $100M Proof Before Autonomy
37:42 The “We Just Had a Baby” AGI Test
39:49 Why IBM Walked Away from Neurosymbolic AI
41:17 Why Classic Neurosymbolic AI Fails
44:10 15% to 93%: How AIGO Transformed 1-800-Flowers
47:04 Train an AI Financial Analyst in 3 Days
50:51 How Much Will AIGO Cost?
52:23 4 Ways AIGO Will Change Work
53:35 Why Japan’s Robotics Giant Wants to Fund AIGO
55:47 AI for Every Human: Ending Unnecessary Suffering
57:44 Flourishing Is 36x More Important Than Survival
59:05 Why Srini Quit Seven-Figure Jobs for AIGO
01:02:27 Scarcity to Abundance to Flourishing
01:07:35 Why Silicon Valley Won’t Fund Non-LLM AGI
01:12:48 What If a Bad Actor Uses AGI?
01:13:59 “Truth Is Optional”: Why AI Guardrails Are Marketing
01:14:35 Who Secures Your Private AGI?
01:15:29 Continual Learning ≠ Continual Training
01:17:26 Cognition on Tap: The End of LLM Models?
01:19:49 Universal High Income: Jobs, Timing and the Transition
01:22:22 Frontier Labs Have a Coding Certificate, Not AGI
01:24:10 OpenAI’s Price Cuts: The AI Race to Zero
01:27:04 “Resistance Is a Slow Surrender”: The LLM Alternative
01:28:05 Light Bulb vs Candle: Will AIGO Replace ChatGPT?
01:28:49 How AI Can Know What It Doesn’t Know
01:31:51 AIGO’s Roadmap: Toddler, Teen, Graduate
01:33:05 Compute Race vs Architecture: What If Both Win?
01:35:55 “ASI Is Marketing”: Why AGI Already Equals AGI
01:39:14 The King Midas Problem: Can Abundance Break Us?
01:46:05 The Podcaster’s Ask: Free AIGO
01:26:44 How AI Can Know What It Doesn’t Know
01:29:46 AIGO’s Roadmap: Toddler, Teen, Graduate
01:31:00 Compute Race vs Architecture: What If Both Win?
01:33:50 “ASI Is Marketing”: Why AGI Already Equals ASI
01:37:09 The King Midas Problem: Can Abundance Break Us?
01:44:00 The Podcaster’s Ask: Free AIGO
01:45:58 AIR BTS
01:46:43 End Credits
Transcript:
Aashka (00:49)
hello and welcome to On AIR with Aashka Thank you so much for joining us finally, Srini and thank you so much for joining us on a Sunday. It really means a lot. Thank you so much.
Srini (01:00)
Thanks for having me. Yeah.
Aashka (01:02)
Yeah, let’s dive right into the questions. zero LLMs, no hallucinations.
No massive data, compute or energy and yet you built AIGO While OpenAI, Google and Anthropic spent hundreds of billions betting on exact opposite. You have been saying LLMs can’t reach AGI back when that was deeply unpopular. Now even Ilya Satskuver is agreeing we are back to the age of research. So what does it feel like to watch the Scaling Emperor is naked narrative go mainstream now?
Srini (01:37)
No, it’s it’s it’s something, you know, I I’ve I’ve been saying this for since LLM came out. but my business partner and our CEO Peter Voss he’s he’s been saying this for almost two decades. also Gary Marcus, Professor Gary Marcus, he has been saying this for more than two decades, right? So these are the two people I really look up to. Gary is fantastic in terms of the AI critic. he critiques it.
Correctly. And Peter, on the other hand, he’s not n much of a critique. he points out all the things that LLMs cannot do. but he’s more of a a solution guy. You know, he’s been building this integrated neurosymbolic architecture. and and I became an a part of it as a co founder in two thousand eighteen when we started AIGO
no, it’s absolutely thrilling. This is the best time to build AGI because the real race of intelligence hasn’t even begun yet. You know, they’re brute forcing their way towards intelligence and and and then they you know they’re selling a probabilistic generation tool as real intelligence. and it is not, it can never be. It is still useful, right? That’s why when people ask me, hey Srini what do you think of the bubble, right? My answer is the utility is real, the intelligence is not.
No. The valuation gap between the two is where the bubble lives. They didn’t sell the utility as utility. They sold it as intelligence That’s where the problem is. That’s where the whatever ten trillion dollars are going in. And that valuation gap is where the bubble lives. So that that basically we are thrilled to to be in this space because we do not see any other way to achieve real intelligence without solving cognition. So you really need a cognition first architecture that’s ground up.
It’s not something you can kind of bolt on at a later stage, right? So therefore, all paths leads to cognition first. All cognition are what we call integrated neurosymbolic AI, which is INSA. you got to start with that first and then continue to evolve from there.
Aashka (03:43)
Yeah, yeah, yeah, makes sense. So, you mentioned that your partner Peter has been talking about LLMs won’t reach AGI. So, what early evidence did Peter and the team have that made you say no to all of this LLMs and the massive infrastructure game even before it was cool? what evidence did you have?
Srini (04:06)
Yeah. So it’s actually, you know, three things, right? These are what what Peter refers to as non-negotiables when it comes to intelligence or real intelligence or AGI, whatever you want to call it, right? And these three things are number one is real-time incremental learning. So right now, as we are speaking right now, you know, with each other, we are able to learn new information and as that information makes sense.
We you know some of our assumptions, beliefs, behaviors, actions, some or all of them change about this particular topic, right? Yeah, that’s what learning means. So real time incremental learning is critical. LLMs could never do that. They could not do it on day one. They can never they they they cannot do it now. Architecturally, they are frozen. Architecturally they’re
Aashka (04:52)
frozen in time.
Srini (04:54)
basically that’s a ceiling that they can never learn you know information in real time.
And you know, make sense of it and and and and kind of propagate the changes accordingly. That’s number one, right? Number two is concept formation or abstraction. Humans are really, really good at concept formation and abstractions. LLMs can’t do that, right? With completely a brand new concept, you know, you’re trying to understand whatever it is, photosynthesis that are are you know, a any number of the even kid, you show a giraffe and then you show a dog and a cat, immediately they know.
They belong to a different category, maybe animals or something, but they don’t belong to human category. Right. So these kind of abstractions, we are very, very good at at kind of in making that understanding, which are intelligence, even at the I’m talking about three year old, four-year-old, five-year-old, LLMs can’t do that. Unless it it was not already in the pre-trained data, math evolution of data, all of that stuff. That’s number two, right? The third one, which is the highest form of intelligence, which is metacognition.
Right. Metacognition is the highest form of intelligence. It is continuously validating every piece of information coming in. Does this make sense? Does it not make sense? What do I want from it? What does this person want from me? Or what does this author, you know, whether you’re watching a book or you know, moving in the nature or watching a movie, reading a book, any of the are interacting with other person? it is basically doing that constantly. Right now, you are doing it.
Every everything I’m saying, you’re valid. Does it make a demon that not makes sense? What question should I ask next? You know, all of these things, right? That’s pretty, pretty normal. What do I want? And and what does this person want? so this is a very common thing that that happens kind of you know without our knowledge. That’s metacognition. There is no such thing as metacognition, you know, within LLM. And that’s right. They’re pure statistical, you know, generation tools based on the training data.
and you know this is the first time humans ha have the access to for a tool that can generate so much content. You give 10 words, it can give you a thousand words. Right? That is beautiful. Some of it is relevant not all of it is irrelevant but the you know humans are super excited about it and they say, okay, this is what intelligence is and do that. No, the the model is frozen upon training. It can once training stops, learning stops. That’s that’s a key you know key thing. And therefore, because the training stops, industry compensates it.
With by throwing more data, more compute, more energy, you know, more data centers, more power plants and more capital, right? let me kind of step back and ask you this, right? You know, we you and I, are we born with Wikipedia of information? No. We we learn how to learn, therefore we continue to nobody brute force fed us everything about everything and then and then we start you know giving answers for everything. No. We learn how to learn, we learn some foundational knowledge.
And then based on that, we start learning anything new any point of time we want, whether it’s a new language, new topic, you know, by talking, by reading, by watching, by observation, you know, all of this stuff, right? Now that is the key when it comes to intelligence. Ability to learn continuously and autonomously. that is not something LLMs could do. So we knew that from day one. in fact, I published an article on Venture Beat and Peter wrote about it within a week after LLMs are released. Why?
they cannot the answer. In fact we are, you know, quite honestly we are surprised how long it’s taking for people to realize that that there is utility in LLM I’m not undermining the utility of LLM by any means, right? but that is not real intelligence. It’s a tool. You know, it’s like it’s like they are selling a micro oven at a as a most creative chef on the planet. No, it’s it’s a microwave. You know it’s a tool that
Aashka (08:37)
Nice analogy
Srini (08:38)
that’s what it’s supposed to bo to do and it’s doing that, which is fine, right? But you know that’s what I say. Utility is real, intelligence is not, the valuation gap between the two is where the bubble lives.
Aashka (08:49)
Yeah, yeah, makes sense, makes sense.
you talked about metacognition, incremental learning and stuff. In one of your recent LinkedIn post, you mentioned that self-learning AI requires a full cognitive machinery of intelligence. Persistent memory, concept formation, reasoning, contextual understanding, belief revision, metacognition.
goal management, dynamic world model and continuous learning. And you say that intelligence cannot be brute-forced. So, how did you build each piece of this intelligence machinery in AIGO without brute-forcing? Can you explain on each?
Srini (09:26)
Absolutely. So AIGO is centered around concepts, right? In fact, Peter Voss my business partner, AGI pioneer, he published a paper called Concepts Is All You Need. kind of on the lines of attention is all you need, is what LLM?
Aashka (09:41)
Yeah.
Srini (09:42)
Yeah. So so we published Concepts Is All You Need. definitely a worthy paper to check it out. so in order for AIGO to understand the concept, we ground each one of these concepts. In other words, AIGO is not saying any word.
that it doesn’t know. it knows every single word, every single concept is come fully grounded visually, it it understands it, right? now when we say full cognitive machinery, that’s absolutely true. So what we did is, and here is the core difference between LLMs and and and AIGO’s cognitive architecture, right? So instead of going after the outputs of intelligence, we look at language, music, coding, you know, images, movies, you know, all of these things are outputs of intelligence.
Somebody created it. Somebody created it. you know, a group of people or one person, two person, you know, they created mathematics, they created physics, they created coding, a number of these subjects are created, right? So we look at them as outputs of intelligence. So the question that we ask, this is the Peter’s question, almost two decades ago, right? What mechanisms that humans have that created these outputs?
And these are cognitive mechanisms. And that’s the machinery I’m talking about, right? And that comes down to perception, understanding, learning, adaptation, metacognition, self-directed learning, goal-directed behavior, dynamic world model, concept formation. so they they’re like 12 to 15 of these belief revision. so they’re 12 to 15 of these mechanisms that we have that we use to produce these outputs, right? So the moment you go after the outputs of intelligence, that’s the outcome, that’s the dead end.
But if you go to the source that produced these outputs, now you can you can think and reason and learn like a human. This is exactly what Peter wrote about wrote about in 2002. And Peter’s chapter is called The Direct Path to AGI, right? Ray Ray Cardwell posted it. I’m I’m happy to share the link. so so in that he talks about.
Why cognitive mechanisms are essential. So instead of going after the outputs of intelligence, let’s go to the source of intelligence. And when you deeply integrate them, because y none of these parts, including persistent memory, can be you know good by themselves. They have to work in conjunction with the entire you know cognitive machinery to make sense of what somebody is saying, understand the intent, understand the context, and be able to respond. And and we don’t
talk in in 10, 20 sentences. We talk one, two, three, four sentences. That’s exactly how AIGO goes about it. When people ask me, so how how how is I know how is AIGO going to do this? Think of it as the smartest individual that you have ever worked with, working with you as a coworker, as your assistant, that remembers you, reminds you of things and all that, and that is there to serve you, your agenda, your goals and your you know choices and all that stuff. That’s how AIGO will work. And that
You could be an organization, you could be an individual. AIGO with you there ready to serve you. So that’s what we mean by the cognitive machinery. And what we did is, yeah, what we did is once we integrated all of this, we put this on top of a vector graph database. so it becomes a learning machine, right? It’s not you know like machine learning people use the the the the subject reverse that it has to be a learning machine. So the learning
Aashka (13:01)
learning machine.
Srini (13:02)
machine, right? So the contrarian insight I’m sharing is
The AI era, the winner in the AI era is not going to be a company with the largest model or largest data center or largest power plant. It will be the company who built learning architecture. That means it can learn continuously, adapt autonomously, and compound intelligence through use. That means the more you use it, the more intelligent it becomes. So you don’t need to brute force train it and all that stuff. Once it is at a at a
you know, teenager level, and then you know we make it a graduate. Once it’s a graduate level, it’s essentially game over. It can learn any complex subject on its own, just like you and I. Right? If I give you Aashka you’re gonna go to Spain and the next six months you’re gonna be in Spain, you’re gonna do you know hundred podcasts in Spain with all the people, you’re gonna learn Spanish, you’re gonna that’s exactly how AIGO, you know, would would go about it. so it can learn any subject. In fact, if I give you a graduate, AIGO.
And if you happen to be a cancer research institute director, and and you you always said, you know, I I only have 700 cancer researchers, PhD level cancer researchers, if only I had another 9,300 more, we would be 10,000 people and and we would be solving, you know, cancer, curing cancer, right? Well, I give you one AIGO, and and then AIGO will ask questions, what would you want me to do? And you say, I need you to become a PhD level cancer researcher. AIGO like, okay.
Can you help me understand? basically, whatever curriculum you went through to become a PhD level cancer researcher, you give that curriculum, access to that curriculum to AIGO AIGO will read it in the same sequence, by the way. You don’t give what you studied in in a final year in the first year. But there’s no brute force. You you feed the first year first, second year, third year. You give access to it. You’re not brute force feeding anything. You give access to the book. AIGO will read it.
And AIGO will understand it. Whatever it doesn’t understand, it cannot understand, it’ll come back and ask you additional questions. So it will go through the same process like human does, except because AIGO doesn’t have any biological limitations, it’s working 24 by 7. There’s no physical needs, financial needs, emotional needs, AIGO needs, none of that stuff. So it’s on it 24 by 7. and and it it wants you to test it. You know, did I understand this correctly Can you put me through the test kind of thing? So it’ll go through that. What took you 10 years to go to kind of become that PhD level cancer researcher?
AIGO will get there in a few weeks or a few, you know, two to three months, right? Once one AIGO gets there and you certify, you’re good to go. You are a PhD level cancer researcher right now. AIGO will ask you, how many more of me would do you need, Aashka Right? And you say, I need 9,300 more. That way I become 10,000 people, 10,000 PhD level cancer researchers. Literally immediately you have 9300 AIGOs right? In addition to the 700 you PhD level. Now you have 10,000 PhD level cancer researchers.
Out of which 9300 are digital researchers and they are on 24 by 7. Imagine unleashing that kind of brain power because it learns like us, it thinks like us, they have no AIGO needs they’re freely sharing information. now they’re able to solve cancer, right? Imagine unleashing that kind of brain power for every single problem that we as humanity are struggling to solve. Right? That is where we are talking about in a real explosion in frontier technology.
you know, medicine, climate change, healthcare, so it’s so many areas that are ripe for that kind of innovation. that that really makes abundance, what we call radical abundance, a reality. A reality. Yeah. Right? So medicine, healthcare, energy, education, safety, you know, shelter, all these things become abundantly available at almost near zero cost for everyone on the planet, not just for so we move humanity.
from from where it is right now, we refer to that as artificial scarcity, to
Aashka (17:01)
BOOM
Srini (17:02)
abundance, and then move move them towards human flourishing. So that’s basically the mission of AIGO
Aashka (17:08)
Yeah, yeah, that’s like a lot you said. Let me unpack everything step by step. about the copies that you mentioned, like the 9300 AIGos cancer researchers. like, do we need to wait for like that same week or monthly period for other copies to get trained or one is okay, immediately it can be copied. Okay, okay. Makes sense.
Srini (17:36)
And
AIGO will do it by itself. You don’t have to do anything. As long as the infrastructure needs are taken care of, boom, immediately. Okay, here is what I need. Do I don’t have enough capacity, enough storage. It’ll ask you what it needs. And once you set that up, boom.
Aashka (17:51)
Got it, got it. Interesting, interesting. So you mentioned about Peter’s research like two decades before. So he lays out eight characteristics that any real AGI must have for continual learning. Incremental, real-time, one-shot, proactive, integrative, autonomous, multimodal, limited resources.
He says GenAI based systems simply cannot meet these requirements, impossible that we all know now. So walk us through what continual learning actually looks like INSA Aigo and explain Aigo’s brain basically.
Srini (18:30)
Right, right, right. So in fact we we did a demo. I’ll explain what we did in the demo. It is kind of very self-explanatory, right? So recently we did a demo of LLM playing chess and AIGO playing. You know, AIGO right now is more of a a kid, a five to eight-year-old kid, right? Last year AIGO with a toddler, kind of what we call guided learning. and this year AIGO is going through assisted learning, five to eight-year-old child Next year AIGO will be a teenager.
assuming you we secure funding we need next year we’re talking about self-directed learning as a teenager. No. And and shortly thereafter AIGO becomes fully autonomous intelligence. That’s a graduate, right? And that’s when that’s when it’s essentially game over. It can learn anything it wants. Right. Now what does continuous learning look like? So as a we we just created a demo and in the demo we showcase we have LLMs that have been trained on billions of games of chess.
Let them play chess. And then we have AIGO doesn’t know anything about chess at all. It learns the game on the fly and and then it applies them, right? It applies the rules and all that stuff. But he did the kicker. So what we said is instead of playing regular chess, let’s change the rules of the chess. Let’s switch king to a queen, queen to a king. wow. Change rook to a bishop, bishop to a rook, right? Now you say, let’s play, right? So LLMs and pretty much every AI model.
They they fall apart in seconds. Yes. Because
Aashka (20:01)
Yeah.
Srini (20:02)
they never understood the game the same way humans do.
Then they have a problem in understanding the context, meaning and consequences of what what they are saying, right? but e with with chess, absolutely the same thing because they don’t have any patterns to draw from based on the new rules. There was no game that that that has been fed into the system that with
Aashka (20:28)
played like this.
Srini (20:30)
yeah, with like this, with the new rules. King king like a queen, queen like a king, rook like a bishop, bishop like a they fall apart in the second.
Aashka (20:35)
Interesting test here.
Srini (20:37)
So the real test of intelligence is not about how well they play chess.
It’s about how well they can play chess with new rules. When you change the rule, you know, can it does it still play, right? they fall apart, which when they make illegal moves, they do all kinds of crazy stuff, they fall apart, right? In seconds, right? Now we do the same thing with AIGO, right? So AIGO is like, I don’t know what chess game is. It it just has basic board ontology. That means that that means there are two players and there is a board, there are some rules, and to that extent it knows, it doesn’t know anything about chess. So it learns the chess on the fly.
both both verbally, multimodally, you know, through the interactions, all that stuff right there in real time, right? Now once it learns it, it learns like a like a kid, five to eight year old kid, right? Now once a kid learns chess for even for the first time, now you tell you tell that kid, hey, right now we are changing the rules, move king like a queen, queen like a king, right? They’ll start moving king like a queen and queen is now moving one step at a time, right? AIGO that exact same thing. There is no retraining, there is no pre-training, none of that stuff.
Because they understand the game exactly the same way, you know, you and I do, because that that’s humans. AIGO to the same thing. Now we do, right? Now now we say we start moving you know, a bishop like a rook, right? and we actually show it on the system and then we say, you know, are you are you watching it? AIGO I say, yes, I’m watching it. Now move move the bishop like a rook. And AIGO, now it’s not moving. So no retraining, no pre-training, it’s learning in real-time adaptation.
So that’s what continuous learning is. It understands what the right rules are. So at any given point of time, you say, you know, switch it back to the standard rules. It goes back to the standard rules. It knows what the standard rules are, and then it also knows that, you know, for this, you know, whatever demonstration, all that, we are doing this as a you know a variation of the existing rules. It and and then it remembers that too. Every single piece of information that comes into AIGO brain has a timestamp, date and timestamp, number one, right? and number two, it is fully auditable.
That means unlike LLMs that are black box, you can clearly see where the information came from, when the information came from, and what transformation this particular information went through for it to give you the kind of answer it’s giving right now. So fully explainable and auditable, you know, brain, access to that brain is what you will have, not something that that LLMs can do. It just cannot do that, right?
and you know that’s what continuous learning is. So to do this, AIGO is using all of the cognitive machinery that we talked about, right? let me talk about world models also, right? We talked about that that being a part of the cognitive machinery, right? Yeah. Now, right now, you know, let’s let’s make that assumption what LLM made, right? Which is they’ll they look at language that’s a sign of intelligence, right? And then that’s what humans believe. Somebody can can speak really well, it’s a sign of intelligence and all that stuff. Language is kind of a
Aashka (23:31)
Yeah.
Srini (23:32)
it’s a sign of great intelligence and all that stuff, which is, you know, obviously you know human expression is is there. Now, you you look at people who are born deaf and dumb, they don’t have language, you know, they’re completely deaf. They never had out there. Yet they’re able to navigate through life just fine. They’re able to hold a
Aashka (23:50)
Fine, yeah.
Srini (23:51)
they’re able to hold a job, no problem at all. So language is an output of intelligence.
it’s not the source of intelligence. Source, yeah. So so right now, I’m using language to express the thought that I have. Right? Now, where does the thought come from? Thought comes from an inner model of the world, how I see the world. You know, how I see that that’s the world model, by the way. Yeah, you know, how I see the world. So and this is constantly changing with every interaction, you know, every piece of information. you know, my you know, di that’s why we we don’t call it just world model, we call it dynamic world model.
Aashka (24:27)
The is not frozen.
Srini (24:27)
Yeah, because it’s constant.
And
the LLMs are frozen. So so it is dynamically changing. And every piece of information I get, you know, you know, it’s going to be super hot today, or it’s going to be super cold today, somebody, you know, something happened, you know, elections happen, you know, leadership changed, any number of the things, my job is transferred, any number of the things immediately change our world model. And that is constantly happening, right? so that that’s that’s what it is. Now, that you know, how do you form that world model?
cognitive mechanisms. So it really comes back to again cognitive
Aashka (25:05)
cognitive machine.
Srini (25:07)
adaptation, understanding, contextual understanding. That is how you form that world model. So that is the source or the, you know, that’s the mechanisms that actually produce the output. That’s why we ended up building the integrated neurosymbolic architecture with all of these cognitive mechanisms that are deeply integrated and then put that on top of a vector graph database. So that’s how we have evolved right.
Aashka (25:30)
Very, very interesting. So since you mentioned about the world models thing recently on your LinkedIn post again, you proposed a great north star metric that I would say for continual learning. Tell your AI, we just had our first child and a continual learning AI shouldn’t just log that. It should understand your world has changed like the dynamic world model that you mentioned. Your priorities may change, your schedule may change, et cetera, et cetera.
So it should understand what follows from what it just learned, recognize what it doesn’t know, ask relevant questions. have you tried this experiment with Aigo and how did it respond?
Srini (26:08)
Yeah. No, great question. We we didn’t try this with AIGO at this stage because AIGO is still at a at a kid level. Five to eight year old child, but as AIGO you know evolves through its stages, the teenager and the graduate. Absolutely. Graduate I have absolutely no doubt about it. even as a teenager it should be able to address that, right? So that’s the case. So what we are saying is does it understand the consequences? Does it under understand the meaning, the context and the consequences of
Of of this particular one piece of information,
Aashka (26:40)
situation.
Srini (26:42)
it completely changes your travel schedules and any number of those things that are there. You know, that is where I’m saying, you know, your understanding, your beliefs, behaviors, actions, assumptions, all these things, some of these things are all these things change as a result of one piece of information coming in, right? No, what what LLMs do is they brute force, you know, put it in the memory and they retrieve that information. That’s not understanding
Aashka (27:04)
time to time.
Srini (27:06)
what that actually implies.
You know, the consequences, what you know, what are the implications of this piece of information? You know, what action should I take? What what does it change? How do my priorities change as a result of this, right? So there are any number of those things. None of that stuff happened because the model is frozen. And more importantly, the LLM model is incapable of understanding and implementing those those implications because it is pure you know statistical system to be able to actually understand it, understand the consequences and actually propagate those.
Changes into the model itself in real time. All of this happens in real time. And that’s where you go back to the three non negotiable for for intelligence. Number one, real time incremental learning. Also
Aashka (27:50)
to demand.
Srini (27:51)
the learning automatically implies adaptation. And you know, the moment you say learning, what does that mean? It changes something in you. Either your understanding, your assumptions, your actions, your beliefs, behavior, something or all of these things change. That’s what learning means, right? So that’s why it’s number one and then comes concept formation and
And and abstractions and then metacognition. So all these things are essential for it to do what it’s supposed to do. LLMs cannot do it architecturally, they’re limited. And while they’re useful, you know, for other things, but LLMs cannot ever deliver AGI. It’s I think Yan LeCunn put it the strongest possible way. LLMs are a dead end and a distraction for for AGI. Right? And I we totally agree with you.
Aashka (28:32)
Yeah,
so you have been talking about the INSA architecture. Let’s dive deeper into that. as per my knowledge, Neurosymbolic AI has four main types based on Kahneman’s fast thinking and slow thinking, system one and system two. Two them compress symbolic knowledge.
Srini (28:52)
System two is the metacognition, by the way.
Aashka (28:55)
Okay.
Srini (28:55)
System two is metacognition. Yeah, keep going. Yeah.
Aashka (28:57)
Yeah, so two of them compress symbolic knowledge either as knowledge graph embeddings or as formal logic into neural networks and the other two do the reverse. Neural side extracts symbolic structures either as decoupled federated pipelines like LangChain routing to a calculator or a search tool or as tightly intertwined systems where neural and symbolic run on the same shared representation and continuously inform each other.
The last approach scores highest on reasoning depth, explainability as you mentioned and continual adaptation but it’s also the hardest to build. So does INSA fall under any of these four types? Is it entirely different?
Srini (29:41)
It’s it’s a fourth one. It it’s a closer to the fourth one, I would say. It’s not exactly the fourth one. Yeah, I absolutely right. So once we built these mechanisms ground up and deeply integrated them, so so what it does is it actually, you know, takes away you know resource requirements. So we we we you can run AIGO with at least a million times less data and less compute. At least a million times. That’s huge in the internet world, right?
And and that that’s one of the benefits. I’ll talk about the benefits and then I talk about how we are doing it.
Aashka (30:12)
Yeah
Srini (30:13)
And then secondly, what AIGO allows is let’s say once we give you graduate AIGO, think of three concentric circles, and we are giving you the the core, which is the graduate AIGO. The second layer is given to enterprises or even you know consumers, right? And that second layer is owned by you. What that means is you have full privacy, full sovereignty.
Right. I think Satya recently said you’re paying these model companies twice. Not only you’re paying for the tokens, but because they have access to your data,
Aashka (30:43)
Data.
Srini (30:44)
they’re actually starting their own, you know, companies competing with their customers. You know, like Instacart becoming a manufacturing company, and and OpenAI becoming I don’t know, insurance company, you know, all this stuff, right? So you’re paying paying paying them twice, right? So what AIGO offers is full sovereignty.
That means you control your data, your intelligence, your memory and your business. You are fully in charge of it. So you never have to send that data back to AIGO. AIGO is there. I give you you know, AIGO, and you can put this all of your data on your cloud or on your servers. So it never has to leave premise and then it uses AIGO to make sense of what everything you know is saying and coming in and essentially does that. It reads your policies, your your principles, all of it, your company stuff, or it could be your individual stuff, right?
Aashka (31:31)
The vector databases that you mentioned that is stored on like personal servers or enterprise servers.
Srini (31:38)
On enterprise cloud or personal cloud. Absolutely, yeah. you know,
Aashka (31:42)
Right now.
Srini (31:43)
kid AIGO, if you have to find the internet, we are running it on a three year old laptop. That’s how little computer, that’s how little, you know, energy and and how little data we
Aashka (31:53)
Compute.
Srini (31:54)
need. Exactly. Exactly. It it fundamentally that’s why I say it’s it’s a paradigm shift. It’s fundamentally shift the focus
Aashka (32:01)
Yes.
Srini (32:01)
to to to the architecture, away from all of the craziness going on right now.
Right. so and then the outer layer, I talked about three layers, right? We call the core AIGO and then the expert layer is is what we talked about. That’s companies and and even individual creators, you can use that. the outer layer is the hyperpersonalized layer. So that is when you deliver whatever it is that you have to your customers, right? And your customer could be an employee or citizens or end consumers, whoever it is, they have the outer layer, right? And even in that layer, every individual gets their own layer.
Where AIGO with hyper personalized to them. That means it gets to know your preferences, your choices, your you know, everything to do about you and it can and it is private to you and it’s personal to you. It continues to enhance you, that kind of thing. But if you are getting that through an organization, a bank or an e-commerce company, you will have only that amount of whatever interaction that you have with them. You only have that. But if you can have it in your personal life, now that is completely owned by you. No one can ever see it. So our commitment to all our customers.
Whether they’re consumers or enterprise. They see we will never ever see your data, sell your data, share your data, or mine your data. Never. It’s contractual for us. Because AIGO knows what it needs to do with your data. We don’t have to see it, you know, sell it, share it, mine it. AIGO will base it. And it is there. The purpose of its entire existence is to serve you. And you could be an individual consumer, you could be an enterprise. And AIGO is by the way.
It’s a multi-trillion dollar market because it is language agnostic worldwide, domain agnostic, industry agnostic, and device agnostic. It can sit on any device. So it’s always there. That’s why we say AIGO where you go, because it’s always right.
Aashka (33:47)
Nice one. Nice one.
Srini (33:48)
It’s always there with you. Now, now let’s talk about how did we get here, right? Again, everything I’m saying is already proven. It’s proven. So so we commercialized it, prior generation of AIGO, through a company called Smart Action.
That produced $100 million in enterprise revenue, but it was pre-autonomy. Now we are going for fully autonomous intelligence, right? So we we were basically, you know, we would always dedicate at least two people to six people for every single project. We were defining as an example for an e-commerce company that AIGO, you know, ended up working for and ended up replacing thousands of agents in the call center space. we were defining what shipping addresses, billing addresses, order is order items. AIGO didn’t know any other stuff, right?
Now, once AIGO is fully autonomous, we don’t be we are not going to be involved. We give you AIGO. AIGO will ask all the right questions. AIGO will self configure itself and ask you to test it. And once you test it and you certify it, AIGO will ask you how many more of me do you need? Immediately you have 50,000 customer service agents, or you can have a million of them, right? That’s essentially how we go about it. so that that’s the power of it. So everything I’m I’m I’m sharing right now in this podcast is proven pre autonomy, except now we are going for
Fully autonomous the
Aashka (35:01)
fully autonomous.
Srini (35:02)
path that we are on, right? now, how did we end up building the INSA architecture?
Aashka (35:06)
Insa,
Srini (35:08)
So think of it as these are all learning mechanisms, right? You use memory, short-term, long-term memory, a perception, you know, based on what you are seeing and all that stuff, right? and and and the grounded concepts that we have and be able to form abstract abstractions based on the concept, right? So all of these cognitive mechanisms that we talked about, they have a purpose.
I and and at the end of the day, think of it as one more different type of learning mechanism. Because perception is getting it from what it is seeing, right? Understanding is getting it from what it understands from the context, right? Memory is saying it, you know, what do I know about this person, this context from before, from the long-term memory, and what do I know from the short term, from the recent past or from the you know past? Do I know a year ago, six months back, whatever? And then in the recent past, in the last week, last day.
You know, last hour, is there anything I know? So each one of these cognitive components have a purpose to make sense of, to disseminate, you know, what it is that it is looking at, whatever information it’s looking at, right? And it is, by the way, super fast, even though it’s going through all of these mechanisms. All of them. It’s very, very fast because it is proprietary built you know knowledge graph that we have. So once we integrated them deeply and put it on vector graph database, now it is constantly learning, adapting.
And evolving. And that makes all the difference. Yeah. So that’s essentially how how we are going about it. So you have all of these mechanisms. And so the moment a piece of information comes in, you know, let’s talk about the same example that you just said. You know, we just had a child. We just had a child, you know, we just had a newborn, right? And and immediately, okay, now the memory and the context that, yes, you know, somebody was pregnant and now they delivered a child and you know, boy child, girl child. Okay, they’re going to be happy, but it also changes. my god, they had a
vacation plan two weeks from now now they need to make sure that they can accommodate but any number of these things whatever future plans they have and and you know all the things that they need to now accommodate AIGO immediately get to know how does AIGO know because it already is learning and evolving it has common sense knowledge also at that level. So when we say we are moving from a toddler to a kid to teenager to a graduate, right? So it’s also understanding the metacognition.
And common sense knowledge at that level. So as a teenager, it will have that level of common sense knowledge. And so so will be are they are they so it will be we are the graduate also, right? So it knows all the different things that it needs to, and whatever it doesn’t know, it doesn’t make any assumptions, it’s not going to just guess what you need, it will ask you. I know, you know, I went and I made sure that you know all these things are changing. Is there anything else I’m missing I need to know that’s critical? Or are it will actually help you.
think things through. What are all the implications? Now that I’m having a child, what will what are all the implications? Do I need to save for education? You know, do I need to move to a different location where there are better schools? Any number of these things, right? I do will help you think through these things and then they actually help you plan for these things also. Right. And it will also ask you, it’s a very collaborative, super smart
intelligent the it will also ask you, is there anything else that’s important for you that that we did not discuss about, just to make sure that you didn’t miss anything, right? That’s the type of you know collaborative, super smart, adaptive assistant that’s private to you will be available for every single human being on the planet.
Aashka (38:29)
Interesting. I think this child example can be your Turing test when AIGO reaches graduate level.
Srini (38:38)
Yeah it’s
any number of those things.
Aashka (38:42)
Yeah,
that’s one example, but it’s the best one.
Srini (38:45)
number of examples that it understand the implications of what I just made
Aashka (38:49)
Yeah, yes exactly.
Srini (38:51)
the changes throughout the model selectively based on what that made that shows you you know what it actually understood about what I just said and and first you’d have to validate the information also and then catch
Aashka (39:04)
Yeah.
Srini (39:05)
it all the way through. Yeah.
Aashka (39:07)
Absolutely. Makes sense. So regarding this neuro symbolic or INSA architecture. So back in 2024, I connected with IBM’s NSAI program director. And IBM has a lot of research done on NSAI. So he said that NSAI is more of an academic research area and IBM had until recently stopped working on
NSAI and producing more content and research on that, even after spending years and serious money on neuro symbolic AI. So they seem to have discontinued this. What did you see? what was opportunity that you saw with NSAI that you acted and they stopped acting on that?
Srini (39:58)
So so we we we never cared what IBM was doing, anything like that. I think DARPA came close to what we are doing right now. DARPA published a piece on three waves of AI. First wave was about good old fashioned AI, all the rules handled Second
Aashka (40:12)
Yes, Go FAI
Srini (40:14)
Wave, GoFAI exactly. the second wave is what we are going through right now, explosion of data and neural networks, all that stuff. Now you have generation tools and and simulation tools like world models and all that.
Which is great, right? They talked about third wave of AI, and that is all about contextual understanding and adapting, right? And they call that cognitive AI, right? So they are the ones who came closer to what we have. What we have right now in terms of the you know integrated neuro symbolic architecture. Now the other thing we did is we knew neuro symbolic, the classic neuro symbolic itself had number of limitations. I think Peter wrote wrote an article about it. Because
It it’s not in t integrated. So what what that means is everything has to be built ground up. You cannot say what’s the best memory memory engine. let’s take the you know, a Stanford Professors are and and memory is great great from this company. You cannot do that. You have to build everything ground up. You have to deeply integrate them because every piece of information coming in, it has to make sense to the to the system. what what do I know about let’s say you are saying something, right? Let’s say hey city get like can we get on a podcast, right? Let’s say this is your
You fifth attempt, we already had two meetings about the podcast. We never did the podcast. And then we send a request. If AIGO receiving it, AIGO like, okay, what do I know about Aashka yeah, she already reached out, and we had five, you know, five exchanges, two of them are actually, you know, or or you know, on Zoom, we talked about it and all the stuff. So AIGO basically and need to understand so entire cognitive machinery is in action. and and okay, we have seen some samples of Aashka we like that, you know, A, B, C, D, check, check, check, check, check, and that makes sense and all that.
Okay, now the right action is to let Aashka know when we are going to be available, we are going to be ready for the podcast, right? So it gets to that conclusion and and that requires the entire cognitive machinery. and and you know the LLMs just don’t have that kind of machinery. AIGO does. so we ended up building the integrated neuro symbolic architecture ground up. We are not using any standard out-of-the-box tools, anything like that. So we had to build our own proprietary way of going about doing things, right? So that’s where some of our IP.
also is some of our trade secrets out there. so that is exactly how you know Peter basically came to that conclusion that we cannot use any kind of standard out of the box kind of modules. We have to build this ground up. And we’ve proven it. You know, when we delivered this as an example for a company called 1800 Flowers, they have 20 million customers, even though initially we went to improve the productivity of the call center agent, very quickly they realized the the power of AIGO.
we literally went from for one business unit, we w literally went from fifteen percent containment. Containment is you know without transferring it to human. is your digital assistant, you know, is it capable of handling the request all by itself, which means it understands the request, it fully addresses the request, no transfer to human, right? They were struggling at 15%. and with with AIGO, we were able to move that to 93% in in six months. It’s always huge. And then they said
Aashka (43:18)
What language was it? English?
Srini (43:20)
English. English. Only English. Okay. Right now AIGO with English only, by the way, right now. Okay. But once it’s a graduate, it will learn Spanish the same way. You would learn any language it will learn on its own. Right? Once AIGO with a graduate, that’s the beauty, beauty of I.
Aashka (43:32)
Yes.
Srini (43:33)
So so what then what they did is hey Srini and and and the executives basically said, in instead of just addressing any request coming in, what if AIGO becomes a a a hyper personalized gifting assistant?
For each one of our 20 million customers. So immediately we had 20 million Aigos working one-to-one with each customer that that they have, right? Let’s say if you happen to be their customer, Aashka AIGO will get to know who is Aashka what kind of things does Aashka like, who are most important people in Ash Aashka’s life, what kind of things they like, what events in their life does Aashka care about, and what kind of budgets Aashka has to spend on these events, you know, for these people, right? Let’s say you sent something for your mom last year, right, for her birthday.
now AIGO with would remember that and then basically remind you a week or two weeks before the event, hey Aashka, your mom’s birthday is coming up. You want to do same, you know, something that you did last year, she really loved it, or you want to do something special because she’s turning whatever, 60 now and 70 now, whatever it is, and it’ll actually make some recommendation kind of thing. So it becomes your hyper personalized gifting assistant. so the the the principle there is we want to make sure whether you have 10 million customers or 200 million customers, each customer needs to feel.
as if your entire company exists to serve them. Right? So that customer feels now, you know, one eight hundred flowers is whatever, three or four billion dollar company. Now they say this entire company exists to serve me through this digitalization that really cares about you and knows your needs, knows you know what kind of things you like and you don’t like, know the important people, all other stuff. Now it basically gives you those options and make it frictionless to get it done for you. Right? So imagine a bank doing that, imagine an insurance company doing that.
That’s the type of service now companies can provide. Once AIGO with a graduate, we can still do that right now, but that is still pre-autonomy. We have to be in modern. Once AIGO with a graduate, that’s when it’s literally game over. It’s not just customer service. It can do that same thing for finance, for cancer research, for HR, every single domain within a company, every single industry and every single language, and it can sit on any device. Everything is up for that. That’s why I say the real race to AGI, real race to intelligence.
hasn’t even begun yet.
Aashka (45:46)
Yeah, yeah, yeah, makes sense, makes sense. So you mentioned about this cancer researcher or financial analyst or whatever it is. So the learning times for each of those AIGOs would be different, right? The timelines.
Srini (46:00)
The way yeah, yeah, it it’s different. Exactly same as human, right? So the way you want to think about it is imagine you’re you’re just hiring a brand new graduate off of top university, Stanford, Harvard, right? They just finished their undergrad and grad, right? Four years of undergrad, two years of grad. Now you hired them, right? And however much time that a human takes to get good at their job, right? You you do you do only those things to AIGO as you would do to a human.
In order to bring a financial analyst, let’s say you hired them, they they graduated Harvard in finance, you hired them as a financial analyst, right? What do you do with them? You you tell them here are the things, here are the day to day jobs and here are some of the analysis and all that stuff. Can you read this and let us know if you have any questions and all that? They go through the process. In about three minimum three weeks to six weeks, you you validate and then you say you are ready to go get get the job done. And if you have any questions along the way, come back and ask us and anything we knew.
That’s exactly what you do with AIGO, except two things. AIGO will get there a lot faster than a human, consuming the same amount of resources.
Aashka (47:08)
Yeah, of course. That’s why we need Aigo.
Srini (47:10)
Yeah, exactly. What takes you three to six weeks? AIGO will get there in a couple of days, two to three days. And then once you certify you are good to go. AIGO, AIGO will ask you, how many more of me do you need? Let’s say you have 20,000 financial analysts working in your organization. Immediately you can tell me I I need 20,000 more of you. Immediately you have 20,000 financial analysts, AIGOs working for you that are now good at your business, at your company financials, at your analysis, you know, financial analysis of your company, right? That’s the power, right? and now.
Yeah. Imagine the same thing on the cancer research, right? If you bring a a fresh graduate off of a university and you say, I need you to become cancer researcher, and that’s it they did biology and and all that stuff. So they’re they’re prime, but they are not PhD. And then how long it probably takes them whatever, six years, you know, seven and AIGO will do that in in few weeks to few months, like I said. So but it’s the same process, the process is the same because it is learning like a human without any biological limitations.
Right? It is thinking. And it is c it is coming to you asking questions. It’ll bother you with number of questions, just like humans. A smart, super active, you know, a person would do the same thing. I read this and this, and that made sense. This then this did not make sense. Did you help understand what it is? Where can I get that information? Where can I learn it? So now because I give you graduate AIGO AIGO can learn from any of these means. AIGO can learn from reading a paper, AIGO can learn from watching a video.
AIGO can learn from a lecture, from watching a lecture, from you know, any number of that’s with multimodal learning. When Peter says continual learning is multimodal, it can learn from any of those mechanisms by watching, by speaking to somebody, all of these things are within the realm of possibility. That’s how AIGO will work The speed at which it becomes really depends on the subject. It will be a lot faster than human. And the most important thing here is I’m not brute force feeding any other stuff. It will use less resources than a human. And that’s powerful. Compare that with LLMs.
You know, you’re talking about whatever billions of dollars, trillions of dollars, the data center, the power plant. None of that stuff. None of that stuff. It uses less resources than a human. Now I’m not saying less resources than LLM. Less resources than a human. but it it’ll get there faster and then you have it scales immediately. You need twenty thousand of them, you need ten thousand of them, boom, instant scaling, right? That’s the power.
Aashka (49:31)
Yeah, yeah, makes sense. So the cost economics for like the pricing of AIGO will also depend on if I want a financial analyst versus a PhD level cancer researcher, right? Yeah. Okay.
Srini (49:45)
Yeah, exactly. So it could you know, just come up with what are you paying for a financial analyst today? Maybe we charge some small percentage of that. You know, maybe fifteen percent of it, twenty percent of it, ten percent of it. Well, we we we’ll yeah, that’s basically the the business model. For that there are for others we have different b business models. And if we are looking at, you know, attending you as employee, however much you pay for that employee in real life for a human, you pay a tiny fraction of that for AIGO
But we there’s also volume discount, there’s gonna be tiered models and all that stuff. and if you want AIGO, because AIGO can work 24 hours, it doesn’t need to be 9 to 5 anymore, right? If you want that, then then there’s an extra fee for because of the all the extra hours, right? Or maybe you know we’ll just sell it by the hour and you decide how many hours you want, then we can give you that, right? So any number of these things are are are possible once you have the graduate AIGO. And by the way, graduate AIGO, one of the, you know, that the we look at it four different applications, right?
What we have been talking about is what we call adaptive coworker in the enterprise. Whether
Aashka (50:45)
Yeah.
Srini (50:46)
you are a two-people company or two million people company, AIGO becomes an adaptive coworker. AIGO can also build adaptive tools for your enterprise. The moment it is working within your organization, and you say, you know, Aashka, you’re using four different tools to accomplish this thing. There’s actually a better way of going about it. Why are you paying four different vendors?
Here is one vendor or I can actually create a tool that can you know skip all these four vendors. Here is the best way to go about it because you’re not using many of the features that they are offering anyway, right? So that’s adaptive tools for the enterprise. So that’s the first bucket of applications, right? The second bucket of application, this is very, very interesting. In fact, a Japanese company, robotics company right now is actually looking to fund AIGO completely, full round. because they basically concluded after playing with LLM for last year and app yet,
Two things LLMs cannot deliver when it comes to robotics, physical AI, right? Number one, reliability. Number two, deterministic safety. LLMs cannot be reliable. LLMs cannot be deterministically safe in every scenario, every situation. So they basically stop that, right? Now they’re looking for alternative, they found us, and we are in in a deep discussion. but bottom line is AIGO become the mind for the body. They’re they’re bringing the body physically.
Aashka (52:00)
Body.
Srini (52:01)
You
become the mind for the body. And here is what happened. The moment AIGO, AIGO’s brain is is there in that body, you don’t have to train gazillion times. You know, today when just to fold clothes or just to do dishes, they train it gazillion times, right? AIGO watch you do it once and it does it. Let’s say if it makes some mistake, you say, Hey, AIGO, there is still a stain on that on that particular jet, make sure there are no stain AIGO, Okay, understand it. And no stain ever. That’s it. One instruction, it understands because that’s continual real time.
Aashka (52:29)
Continual learning.
Srini (52:30)
It
learns it. Okay, I understand it. And then it cleans it up and you never have to say that again for the rest of its life. And it’s there in the in the core brain. So therefore, any robot that comes after that we have would have already come with this kind of instruction, right? That’s what we are talking about, right? So you know, training costs, all of this stuff, that’s where no need for massive data compute, energy and no retraining, no hallucinations. And and some of the smartest people they talk about, how come it doesn’t hallucinate? Well
Hallucination is happening because your tool is statistical. It is just probably a
Aashka (53:05)
Yeah.
Srini (53:05)
tool. Right? Our time know none of the statistical mechanism. It’s cognitive. It’s the evidence-based approach. It gathered the evidence before it answered. That’s why, you know, when I talk about you know in integrating neurosymbolic
Aashka (53:16)
Yeah
Srini (53:17)
collecting evidence from each one. What do I know about this person? What do I know short-term memory in the long-term memory? Did I meet Aashka six months back? Did I meet Aashka in the last two weeks, three weeks?
What did Aashka talk about? Is she consistent with what she talked about? These are all the things that the human mind goes through, right? That’s exactly what AIGO is AIGO is collecting evidence from all of these cognitive mechanisms and then concluding this is what I need to do, this is my next step, this is my right action, this is what I understand. Sometimes it comes back and say, I still have a couple of questions, Aashka. Here is where I’m coming from, right? That’s exactly how AIGO is is going about getting and that is super important and powerful when it comes to physical AI. That’s the second bucket, right? Third
Aashka (53:56)
Yeah.
Srini (53:56)
bucket is
frontier tech. This is the cancer research, climate change, any of the you know, all the research going on on the energy, you know, education. Hard
Aashka (54:05)
Solving the Hard Problems
Srini (54:07)
problem. Absolutely. That’s the third bucket. The last and and and the most important bucket is it becomes a personal, adaptive, private assistant for every single human being on the planet. It will speak to you in your local language, right? So so let me give you an example, right? For some countries, we may give it away for free for a period of time, as an example.
You know, I I I use this example quite a bit because I really deeply feel for it, which is an eight eight-year-old girl child, you know, she goes through in Sierra Leone, she goes through almost a million I’m sorry, mile deep tunnel to go and handpick cobalt to pull it out and to kind of and then come back, right? now and then she gets paid whatever, a dollar a week or something like that.
Right. What if she had access to AIGO that can speak Sierra Leone’s local language, right? And empowers her for better working conditions, better wage, and and any number of these things. So there’s a lot of unnecessary human suffering is going on for
Aashka (55:10)
See you.
Srini (55:10)
whatever is the reason. You know, there are other humans that are taking advantage of it, any number of reasons, right? So we want to empower humanity. and and also, you know, when you when you take care of the
the food, safety, shelter, education, energy, healthcare, medicines, abundantly available for every single one. You know, how how else now you are going to you know flourish because some people are using it in in a negative way. You know, but I mean I AIGO will help you create machines to, you know, to do that. But unnecessary suffering is something that that I believe AIGO that my personal opinion, AIGO will reduce significantly from from the people who are being taken advantage of.
Aashka (55:50)
Yeah, that’s
Srini (55:51)
That’s the particular categories.
Aashka (55:56)
very powerful.
Srini (55:59)
Yeah, the in in fact I now that you said so one of the reasons I get up every day do what I do, it’s it’s more than a company, it’s more than a mission.
Aashka (56:06)
Yeah, it’s the
Srini (56:08)
Devoted to the mission of AIGO and the mission is to produce maximally beneficial AI for human flourishing. For human flourishing. Right. So that’s basically why I do what I do. That’s what gets me going every single day and
And know, all the you know, it it’s it’s not easy. I’m going against every single big company, every trillion dollar company
Aashka (56:30)
Yeah
Srini (56:31)
out there, every single media company. They’re saying one thing I’m not I’m not like I don’t I don’t like dislike them or anything like that. But the fact is, what I’m I’m I’m talking about it goes in in in the opposite direction of what these guys have been talking about for ages. And they have the cloud, they have the money, they have the power, they have all of the stuff. but but they will fail in in achieving AGI.
they will fail. but if they continue going down this path. Now if their power, you know, INSA architecture, that’s a different story. right. but going on the side that they’re gone, they will not dead
Aashka (57:05)
It’s a dead end for them.
Srini (57:07)
end. Exactly.
Aashka (57:08)
Yeah, yeah, makes sense.
Srini (57:10)
And that includes all of the MANGO companies, right? They the the Microsoft, you know, Apple, Amazon, Anthropic, Nvidia, Google, SpaceX,
Aashka (57:19)
when they
Srini (57:20)
you know, all of them, right? Yeah. Yeah. I want them to succeed in whatever they’re doing, but when it comes to AI, they’re hundred percent in the wrong direction.
Aashka (57:30)
Yeah, yeah, makes sense. So, since you mentioned about the flourishing part maximally beneficial for human flourishing. here’s a take from Will MacAskill, the philosopher and the originators of the effective altruism movement. So, he has this framework, the expected value of humanity’s future is survival into flourishing, like survival and flourishing.
Yeah, the probability that humanity survives the century and the quality of life given we survive. So, most of the research is going on towards survival right now, like saving ourselves from AGI or the bad things of AI. So, he argues that flourishing could be roughly 36 times more important than survival. Most AI safety work is, as I mentioned, is focused on survival.
And Aigo’s mission is maximally beneficial AI for human flourishing. So can you expand on how human flourishing became a mission for Aigo and how you are advancing towards that mission?
Srini (58:40)
Yeah. So for me, you know, Peter Peter actually shared that. This is why he’s doing what he’s doing, right? I’m I’m you know, extraordinarily inspired by that mission. Right. I walked away from whatever seven figure gigs, everything that I had, I put you know
Aashka (58:56)
Yeah.
Srini (58:56)
my life on hold to to jump into AIGO and continuing going down there. That’s right, the T absolutely it it it it it’s it’s not easy, but I love it. I I wouldn’t want you know, otherwise, I wouldn’t want any any other thing, right? That that’s how deeply I’m
I’m devoted to the to the case. Right? now, so so I did not arrive. It made sense to me. You know, it made sense to me how Peter arrived at that and the possibilities with the technology. Right. When when was the last time you said the technology that I’m working on will change humanity for better? Positively. 100%
Aashka (59:29)
Yeah, not many.
Srini (59:34)
possibility, right? We are we are made we want to make that reality. We want the world to say.
In in my mind it happened already. I’m living in that world, right? So that’s why I’m everything I’m doing is to make that reality come closer and closer and closer so the rest of the world can can see it and and live in it. Right. so absolutely. That’s why I said, you know, when I mentioned this, I said we are living in artificial scarcity and and that that’s a survival mode, right? So we move that to abundance, right? And then we move to human flourishing So those are the steps, right? So we cannot go from here to flourishing straight up.
And and how do you go from you know right now s survival is not just for saving from AGI, it’s more about you’re running behind the dollar every single day. You know? what if all your basic needs are taken care of? You know, that is food and safety, shelter, you know, energy, education, transportation, medicine, you know, all if all these things take in care, who are you? Is the question for humanity. Yeah. Who are you and all the basic so that’s a deep.
You don’t need checking account because your kid now there’s always going to be limited space on the ocean in in order for you to build your home and all that stuff. Money will have a completely different meaning at the time. So it won’t be, you know, you’re not you know fighting for survival anymore. So survival goes away The moment you move humanity from scarcity to abundance, survival.
Aashka (1:00:59)
abundance.
Srini (1:01:00)
Now from abundance to flourishing is all these breakthroughs, which means some of the chronic diseases need to become history.
Aashka (1:01:07)
Yes.
Srini (1:01:08)
Poverty need to become history.
So and then longevity becomes standard. So people living for whatever hundred and fifty years, two hundred years need to become standard. So these are any number of breakthroughs. You know, we don’t have to mindlessly burn the planet for a prompt that they are doing right now, or for any number of things that they are doing that, because once AIGO is available, how could we create a massive energy? Maybe it’s baking a grain of a sand. I’m just throwing this out. Energy
Aashka (1:01:33)
Right? Yeah.
Srini (1:01:35)
hidden in it, if you can use that.
unlimited sand grains available and AIGO will figure out a way. I’m just throwing that at an idea. What I’m saying is how can we produce what we need for humanity, food by the energy wave and all that, without having to destroy the planet. That becomes one of the major frontier the third category of application that I talked about. And let them go do research and let them come up with all the options because AIGO will do all the stuff. You know, you can have ten human beings, you can have ten thousand or a million AIGO to work around the clock to solve these these kind of problems, right? So that is where
the the power is to move them from abundance to flourishing, right? So survival, you know, we get through the survival from scarcity to abundance and then from abundance we go to flourishing. So and then the next question is is AGI going to destroy humanity, right? Well this flourishing is the only thing that is actually built into AIGO. In other words, yeah let’s say you have AIGO now you want to build a biobomb or you want to destroy people, cities or whatever it is, AIGO by itself will stop you.
So that’s the only built-in mechanism that AIGO has at any point of time, anyone doing, you know, it is there to help you, right? So what does human flourishing mean for us, right? Two things. At a micro level, at an individual level, it will enhance your agency. It will enhance your well-being, right? It will enhance the quality of your relationship. It will enhance the quality of your life overall, right? That is what it is.
You have a competence. You want you want to learn a guitar, how to play guitar. AIGO like, let me teach you. And then and then it it runs. Okay, here is the beginner, intermediate, advanced, and it will take you through any subject matter, anything. You want to become good in photography, whatever it is, AIGO in there. Enhance your competence, enhance, and it’ll help you think through things. So you don’t have to suffer unnecessarily. If you Aashka, you have three options right now. If you take option one, here are the consequences you have to face. And let’s say if you face that one, here are the second degree, third degree of consequences.
Do you still want to go through that path or you want me to talk about option two and option three? It’ll help you think think through these things, right? So this kind of thing. So it’s there. And all of this is happening in private. So no one will ever know because it’s it’s your AIGO your agenda, your preference. You it is your personal, private, adaptive assistant, right? So that that is there. So that’s at the individual level, right? Now at a at a macro level, the way we define human flourishing is planetary abundance, social abundance, that is education, healthcare, all other stuff.
personal abundance and then frontier technology abundance. That’s how we break at a macro level. All of this, by the way, is worldwide. Not just one country, not just one region, not just one. Everyone has to you know receive the benefits of AGI. that’s right. So when Peter shared this with me when I first met him, I scheduled a 15-minute meeting that went for three and a half hours. I was completely blown away. I referred to that as one of my transcendental movements.
And and I clearly saw the possibility, even though I didn’t understand everything about AIGO at that time. Yeah. Because it took me a few months to really because number one challenge I had, and I still see that the challenge for number of people, unlearning what you know. Unlearning what you know, right? Because people think they know AI. What they know is statistical, machine learning, deep learning. but real intelligence i that’s a very small portion of it. Real intelligence is more cognition. So you that’s why.
You need to solve cognition first before you go to computation. And you know, their starting point is how much data, how much volume, how much compute, and how much energy. Our starting point is what is intelligence and how does human intelligence you know work and and what makes human intelligence so special, so different, that a child that’s born today that can’t do anything, right? goes and earns a PhD in 25 years. How is that possible? You know, you know, because some people argue and and it’s very dominant thought process in the Bay Area.
From Elon Musk to everyone else. Intelligence is what intelligence does. They say it casually. Intelligence is what intelligence does. Well, not necessarily, because you you you look at the drill that punches a hole in the wall. It punches a perfect hole in the wall. You and I can’t do that. A tool does it. That’s what it is supposed to do, right? But then you look at a child that’s born today, can’t do anything. You have to do everything for the child. And that same child goes and earns a PhD in 25 years.
So intelligence is not just what you know, it’s about your ability to interact with the real world. That means you’re learning in real time, right? Adapt and improve continuously and autonomously. That is why a child becomes a PhD in 25 years. That is what is so special about human intelligence. That is the type of intelligence that we are building through integrated neurosymbolic architecture.
Aashka (1:06:15)
Yeah, interesting, very interesting. So I want to unpack two things from what you said. So you mentioned about unlearning.
you recently had this investor meetup. So what would you tell the person or the investor who is walking out of that demo, the AIGO demo, still seeing AIGO as just another chatbot? What are they missing?
Srini (1:06:38)
Most of them seem to get wrong, right? Number one, they think of us still a a research company. We have zero researchers in the company, right? We are a development company, not a research company. We were already proven commercially, right? Now we are developing it for fully autonomous intelligence, right? That’s where we are going. And we are not like two, two years away from from getting their up on securing funding, right? So that’s one thing that they get wrong, number one, right? Number two, they’re still looking at AIGO.
Through the LLM’s lens. Through the lens of it. It’s like it’s like I’m introducing you know a light bulb and electricity. They’re still talking about how much wax do you need? you know.
Aashka (1:07:20)
For the candle
Srini (1:07:20)
Exactly. Exactly. So they’re still looking at at the at the at the car. I’m introducing car, they’re still looking at it from lens of a horse, right? so any number of these examples, right? So you cannot look at it from it, it’s paradigm shift.
That’s requires a new learning, new new thing. No, they cannot get out of their way. And some of it is also coming from the fact that they have put way too much money. Right? We’re talking about trillions. There’s not a single thing you know, in the that I’ve known in the last whatever hundred years that has taken this amount of money and level of misunderstanding. This level of misunderstanding that they’re dealing with. It’s absolutely mind boggling. Right? but you know in what it it’s what they call sunk cost falla you know fallacy.
So they because they sunk all all this cost, you know, now they don’t want to hear anything, they don’t want to look at anything else, you know, they bury their head in the sand and basically say, This is all I see. You know, show me where it is. Well, get your head out, you’ll see it.
Yeah. Exactly. But then things will change the moment some of these the when they get their money back. Some of these companies go IPO and then they cash out and they you know dump their you know these things on on and retail investors. Now they can buy, right? I think some of them are starting to realize most of them but most of them don’t want to acknowledge it because it’s risk for them because it’s their own money. It is and I have at least some of them being honest enough to put that in writing.
That I see I really love what you guys are doing. but it it’s
Aashka (1:08:54)
you all the best.
Srini (1:08:56)
no that they say towards the end anyway. But they actually give me b be honest reasons, right? Number one, it goes it goes against our thesis how we raise funds. You know, they raise whatever a hundred million dollars, 150 million dollars. They’re they’re raising it from other people, right? The thesis that they submitted to them was that LLM is gonna be the next big thing and based on this we are raising it. Now they have to go back and say.
Sorry, LLMs are wrong and all that stuff. But that doesn’t make them look good, right? So it goes against our thesis, number one. The second point that they make is it goes against our port coast, they call it. The port portfolio companies. They’ve already invested from this fund. They invested 10 companies, 20 companies, 50 companies, all of them are going in that direction. The moment you know AIGO succeeds, the value and the valuation of these companies will not be the same. So the you know, it goes literally against their DNA.
Aashka (1:09:45)
Yeah.
Srini (1:09:46)
How they raise funds and how they have invested the these funds, right? And that applies to pretty much every tier one and tier two I know in the you know in the space. Yeah.
Aashka (1:09:56)
Soon.
Srini (1:09:57)
So so only some of them, you know, they’re really good at saying no in five hundred different ways. Yeah, exactly. Right? so you really have to dissect it on all the if they’re not writing me a check, giving me a thumbshit, it’s a no for me. So I know I know that I’m not gonna, you know, waste my time. Of course, you know, I’ll I’ll spend time if they’re open to learning and all that.
but bottom line is it goes against the against the thesis that they’ve been making better on in the last three, four years, right? so you know, unless it becomes very clear out in the open where there’s bubble bursting valuations going down and all that, that’s the only sound that they will hear until then they they bury themselves in a
Aashka (1:10:37)
Yeah, in the sand. Yeah, yeah, makes sense. So, you mentioned about Aigo learning just like humans and you touched a bit on AI safety and the values are ingrained in Aigo. what stops a bad actor from giving Aigo? Like you mentioned the example of okay, it will
help you evaluate different options and if you go this way then this would happen and stuff like that. But if I am a misaligned human and I am telling Aigo I want to go with the worst or the bad option anyways. So in that case like how will Aigo help me not pursue a bad goal or maybe Aigo is itself not a bad actor. So how do you ingrain?
Srini (1:11:28)
So, AIGO will reflect your values and all that stuff, right? if you’re doing something bad knowingly, if it’s doing only bad to you, it’s it’s your call. It’s your call. AIGO will let you know, educate you and all that stuff. Ultimately, okay. You know, it’s your call. It’s your call. It’ll tell you about the consequences and all that stuff and and to that extent, right? that you know you’re gonna damage your relationships and all that stuff. AIGO will say that’s your call if you wanna do that. That’s if somebody is drunk and and they’re calling their their
ex boyfriend, ex girlfriend, whatever it is and all that stuff we can educate you. But then if you still want to make a call, go ahead and do that. That’s fine, right? Now if you are doing anything that causes damage to
Aashka (1:12:09)
Bombing or suicide or like...
Srini (1:12:11)
So I AIGO will stop you essentially. AIGO we’ll we’ll basically warn you first and then stop you from taking any action, that kind of thing. So to the extent that if you need to call authorities, it’ll it’ll do that also, right? to that extent. yeah. So AIGO so that’s where the human flourishing is built in mechanism into AIGO. So we look at all the external safety guardrails and all that stuff. we there is no such thing as safe AI right now.
Aashka (1:12:39)
Yes, exactly.
Srini (1:12:41)
You know, the people are like safe and True.
Aashka (1:12:43)
seeking and yeah it’s not
Srini (1:12:46)
People words to say your guardrails are marketing
Aashka (1:12:49)
Marketing words
Srini (1:12:50)
word, your guardrails are a joke. and truth is always optional. so when truth is optional within LLM, trust is impossible. How can you trust a system when truth is optional? And truth is always optional because architecturally it’s a probabilistic content generation tool. therefore, you know if truth cannot be verified, trust cannot be promised, trust cannot be delivered.
So now you have to have babysitter sitting on every single output, is it right, is it not right, and all that stuff, right? so no thing. But with AIGO, that’s not the case because we are not a probabilistic generation system. It’s an evidence-based system. It’ll back it up and it’ll actually show you exactly where it came from, all of that stuff, right? So if somebody wants to kind of do some bad thing and get out, we have a full track track record of every piece of information that came in, when it came in, who it came in and who came who it came from, the whole thing is is there, right? So it’s fully auditable.
Aashka (1:13:45)
Hmm. Hmm.
Srini (1:13:45)
Fully scrutable
Aashka (1:13:46)
God.
Srini (1:13:47)
fully scrutable It’s not a black box, right? So those are the mechanisms that that we have. yeah.
Aashka (1:13:53)
Yeah, and also you
Srini (1:13:55)
Government asks the and I think governments have the power to do that. They want some more, obviously we we’ll we’ll work with the governments to, you know, okay to that extent, yes. but AIGO itself will know and it’ll stop.
Aashka (1:14:09)
Yeah. You mentioned about this, like my data or the processing is done on my personal cloud and for enterprises, it’s their own cloud. Correct. So how do you basically ensure the cybersecurity of those things? is that the responsibility of AIGO as a company or is it mine?
Srini (1:14:32)
We’ll do our part. They have to do their part because now it’s sitting on your cloud. I cannot take one of these. You’re already protecting all your data anyway in the cloud. Okay. Because the rest of your company data is also there. Where you know you don’t put AIGO data somewhere else. You put them exactly where the rest of your data is. So you already have the security, you already have the firewalls, all that stuff. That’s where you know AIGO AIGO data will sit. Right? It will never leave your your premise, whether it’s on the cloud or your service, all that stuff. And then we will have our own.
You know, cybersecurity tools and all other stuff available on our side. okay. so that’s how we are gonna go about it. That’s exactly what we did with the enterprise customers that we served already. So it was on trend on their cloud. they had all the security stuff and we had our own security. So yeah. Anytime AIGO core, AIGO is updated, they’ll get on the air updates. So they don’t have to whatever. So we give them an executable copy that is updated you know, by the core. We will never touch the you know, layers, layer layer, the expert layer or the person.
Aashka (1:15:31)
Yeah, interesting. So, Dwarkesh Patel, the famous podcaster argues that once continual learning is real, switching AI providers stops being easy, just like switching cloud providers is today. So, you are not just changing tools, you are firing an employee who has accumulated months of context and replacing them with an inexperienced new intern. So, that’s how the big labs eventually build pricing power.
Do you agree with that and how does that work with AIGO? Like how do see it in the future?
Srini (1:16:06)
So once cognitive AI, integrated neuro symbolic architecture comes into foray continual learning is definitely important. Continual learning is not continual training. That’s the direction that they’re heading. Yes. Right? It’s not continual training. you know, they they think they can change the weight forcefully And even if they do that, they are still looking at the statistical correlations, the causation is missing, and a number of other things of the continual learning, the multimodal learning, integrative learning.
Everything that Peter talked about, the eight things that you mentioned about continual learning Yes. Many of those things are missing, right? So I don’t see the models the way they are today, LLM models, being relevant in that world. Because that world will have not models, learning architectures like INSA. And and the moment INSA comes into play, you completely change the landscape anyway. The infrastructure requirements change, the limitations change, the resource requirement change. Everything is fundamentally different now.
where you you’re now having cognition on tap. Anytime anyone wants more cognition, you know, you get it. And and to your point of employment. So and and somebody asked this question to Peter during during our dinner. And Peter said, yeah, this location is is kind of built-in mechanism, but we want to make sure that everyone can benefit from it. So what that means is, you know, in order you know for people to have you know whatever level of money you need, all that stuff.
whether universal high income governments are providing it, but also your food and safety and shelter and education, energy, transportation, all of these things are available at near zero cost. And and so it they have to come together, right? It’s not going to happen on day one. We you know we are not going to replace whatever, you know, thousands or millions of employees and by the time as we escalate, there are some of these policies that need to come into play, where governments are offering this universal high income on one hand.
Also the cost of goods and services are going down significantly, where you know living healthily and happily is becoming a norm for everyone worldwide. Right?
Aashka (1:18:12)
Yes.
Srini (1:18:13)
So they say it’s a timing issue, how we manage that timing of it. And then whatever people need money for is something that needs to be provided. and and how they go buy it and all that stuff. So it’s not a a simple answer at the time. We really need to get together with economists, you know, lawmakers.
to come up with a with a model that works. once one country adapts to that and and proves it, you know, the rest of the world will will follow. so that’s a great challenge to have because bottom line is that is moving towards the flourishing. You’re moving the new towards f flourishing. So the current challenges that we have will no longer be the challenges. We will not be destroying the planet the same way we are doing. We will not be suffering from some of the diseases that we are suffering today.
Because all of those things are being researched, developed, and basically kind of beaten down. Basically, no no more chronic diseases kind of thing. That’s a wonderful place to be at that’s that’s where we are going, right? The sooner we you know we can show one or two of these breakthroughs, I think the resistance will be lot less at that time. But right now, resistance is high because these guys have not delivered it. They’re hollowing out you know
Aashka (1:19:20)
Promises of curing cancer.
Srini (1:19:22)
They
cannot ever deliver these things. and they’re hollowing out humanity. You know, people are basically dependent on their tools for every single thing that you do on one hand. And on the other hand, they’re burning the planet, they’re burning, you know, resources like there’s no tomorrow, and not able to deliver the kind of value that is actually even for the enterprises that they’re not able to deliver. Forget about humanity and forget about so
Aashka (1:19:43)
Yeah
Srini (1:19:46)
so and they cannot deliver bottom line. And for all those reasons, the resistance is very high right now.
and we believe that resistance will would continue to get higher until unless we have a technology that starts making difference in their real lives. You know, make things easier, you know, cure some diseases, whatever it is, we got to go down that path. The sooner we can have one or two of those kind of major breakthroughs, either in a diseases wise or energy-wise or or some kind of a you know a tech based breakthrough.
that makes everything that much easier, cheaper, faster and all that. I think now now you build that in a human capital, if you will.
Aashka (1:20:26)
Yeah, yeah, yeah, makes sense. So like this Dwarkesh Patel reference came from Dario’s interview in which he mentioned basically, how will and when will these Frontier AI labs start making money?
So that’s when Dario mentioned that it is very similar to what cloud providers are today. Initially, they were also questioned like how will you make money and stuff like that. And that’s the bet they are taking towards making money in the end. make the cost.
Srini (1:21:02)
It’s
a wrong answer. They have the wrong architecture. They cannot deliver their promises. Right. other than coding, I don’t see many, many applications working at all.
Aashka (1:21:11)
Yeah, and they are like specifically trained for that purpose. Just like you are talking about like it can become a financial analyst or a cancer researcher or any of that stuff.
Srini (1:21:23)
Once AIGO comes out, it can do coding, it can do financial analysis They don’t have AGI. What they have is coding certificate, but they’re selling it at AGI That’s where they’re it’s a very convenient answer. Absolutely. they cannot you know kind of show that much of revenue, trillions in revenue, and they need trillions, to be
Aashka (1:21:43)
Yeah, yes
Srini (1:21:45)
able to a new architecture will completely overtake their if they if they truly believe that then
Then they’re wrong.
Aashka (1:21:54)
Yes, and continual learning is mostly the missing piece that the industry is now agreeing to. But from the different components. Yes, yes.
Srini (1:22:02)
Yeah. They have been agreeing that do that with LLM.
Dwarkesh has been talking about it for at least two years and even before that Sam Altman also talked about it. But no. You cannot do that, right? and therefore yeah. That requires a different architecture that LLMs are incapable of. Yeah.
Aashka (1:22:22)
So you recently wrote that OpenAI cut model prices by 80 % and then by 20 % just weeks after their launch. And that when customers can’t distinguish between models on intelligence, they inevitably compare on prices and the infrastructure race has now become a pricing race. So can you unpack what’s actually happening right now in the AI race among Frontier AI Labs?
Srini (1:22:50)
Right. So I believe both the Chinese companies and even Meta are offering what the Frontier Labs are offering at a tiny fraction of the cost, right? What that tells you is that the models and the capabilities are being quickly commoditized. Yes. Quickly commoditized. They’re only like few weeks, maybe a couple of months away from kind of doing that, right? So
Aashka (1:23:19)
I
think six months gap yeah
Srini (1:23:21)
Yeah, that amount of time to charge a premium and then offer it, they say, right? So there’s only so long you can go without re you know, changing the architecture. and that gap is closing. It used to be six months. I don’t think it is six months right now. I believe it is less than three months right now. and some capabilities are only a few weeks, in a maximum few weeks to three months. So that that’s exactly where we are, right? so commoditize capabilities, basically create this this competition.
that becomes you know a pricing race to the bottom. You know, hey, I offer it for ten bucks, I offer it for you know a six bucks, and somebody says I offer it for sixty cents, come on over, right? So this is exactly what is happening. And I believe that will continue to happen. And that’s exactly why these guys will have a hard time just to survive, because they still need billions to continue to survive on one hand. and on the other hand, revenues are kind of the the revenue window is closing.
The opportunity to make a lot of money is is closing. It’s a very short window. How much are you going to make? What in three months people are going to pay you millions? You know, token maxim is seemed to be kind of going down the drain right now. they don’t want to companies don’t want to spend. If the enterprises don’t want to spend, who else is going to spend? How else are you going to recover you know hundreds of billions of dollars? So no, I don’t see them. I don’t see a path for them. And everyone that’s saying that this is going to continue, is somebody who has invested heavily into them.
Right. Because they have a lot to lose, to give up on on them. Right. They continue to support it, push it and all that. but I don’t see quite honestly, I don’t see a path for them to kind of get the you know, hundreds of billions and trillions in revenue for the kind of money that they are spending right now. Right. and especially once you have a viable alternative, right? So what I always say is resistance is a slow surrender.
There are people who have surrendered themselves to LLMs, right? And there are people who are resisting. To me, that’s a slow surrender surrender, right? what’s more important is come up with a viable alternative that is far superior and far cheaper. Far cheaper and far superior. Right? And now you have you know apple to apple comparison. Let’s go. Now
Aashka (1:25:37)
Yeah
Srini (1:25:39)
you go, right? That’s when you see the beginning of the end. Yeah. For some of these models. So these questions would.
Would become a lot more obvious for everyone. Okay. This thing cannot survive. You know, they whatever 10 billion, 20 billion, 30 billion per month, to survive and and per quarter to survive. And and here we are, we are still struggling for revenue. And this other model that delivers far superior value and you know at far better price. that you just can’t make it beat it because it’s a it’s a different architecture altogether. You can just brute force it, you can just throw money at it. Yeah. Yeah. So that’s why.
Presenting a viable alternative for LLM is the most important thing. That’s what we’re working on. AIGO is that alternative. You can also look at it as that that was a wave one and this is wave two. and as wave one becomes very clear that it cannot deliver. Now, having said that, AIGO can use them as a tool. If you say, can you give me you know Trump speech in Shakespearean style, AIGO says, you know, Chat GPT is really good at that.
do you want to do Chat GPT? And you say yes, it can take your credentials and it’ll write the prompt. It’ll take get the output, it’ll it’ll give you the output. Here is what Chat GPT said kind of thing. So it it can use the it’s a it’s a tool user. It can use any tool, including Chat GPT or Claude or anything like that, right? Yeah. So I’m not saying just because we introduced electricity and a light bulb, that the candles are gonna go away. Candles will have their own You know, you should use candles for birthdays and and you know, any number of
different occasions, but they for meditations exactly.
Aashka (1:27:10)
Meditation.
Srini (1:27:13)
So so they will find their place, right? it’s just not going to be the promise of it that it’s gonna light the entire city, entire world is not going to be there. You just cannot do that with the candles. Yeah. So that’s becomes
Aashka (1:27:30)
So, we touched upon metacognition, the ability to reflect on one’s own thinking that is explicitly in AIGO’s training curriculum. and this is something LLMs have been struggling with, with hallucinations and stuff. So, how does AIGO actually recognize the edges of its own knowledge? What does that look like in practice?
Srini (1:27:52)
Right. So because AIGO is fully transparent and you have all these modules, metacognition is looking at every module. Memory, concept formation. Is it right? Is it not right? Are we doing this right? Right. And it’s looking at the memory. What do we know about this person? And then it also has its own belief revision and goal oriented. Right. What’s what’s my goal from this call? Right? and and you know, it only values what is important to it. You know, if you share a lot of information.
ten pieces of information, eight of them are not relevant, either for AIGO or for this conversation. It’ll listen to it just like humans. You know, we we ignore what we don’t need, all that stuff. But whatever is thrown
Aashka (1:28:29)
It will zone out.
Srini (1:28:31)
out, exactly. It will listen to it, right? It’ll it’ll remember that. But it doesn’t take action. It doesn’t make any other none of that stuff.
Aashka (1:28:38)
Put it in its context.
Srini (1:28:40)
Put
it the context of all that stuff. But the the two pieces of inventory you give is very valuable. Now that changes my goal, that changes my next action I need to take and all this stuff. Now it’s valid. So that in metacognition is at you know, right now it is at still at a kid level. Definitely one of the hardest problems. But we know how to solve it. We are solving it incrementally with that as AIGO evolves it. So by the time it’s a teenager, that’s when you will see you know, basically you give access to a book to AIGO, AIGO will read it.
And and it’ll tell you what it understood. More importantly, it’ll tell you what it did not understand. That’s metacognition. If you’re asking questions about things that you did not understand, that’s a sign of intelligence. Metacognition does that. That says, you know, this is important for me because I need to understand this, but I don’t understand it. So I need you to tell me, you know, answer for this. because it but that means it’s a goal-oriented system.
And it also un understood what it did not understand in the asking the question, right? That’s what we want to be able to show other as a teenager, just to show we believe once we achieve teenager self-directed learning, that will give us the financial escape velocity In other words, we don’t have to struggle for raising again and all that stuff. It d not that we will not, right? We will raise one more time, but now you have people who’s who kind of people banging the doors want to get in, get get into get into the action kind of.
Aashka (1:30:00)
Hmm.
Srini (1:30:02)
Because everyone wants it. Because now I think it was Bill Gates who said it once. if you have a system that can learn by itself, it is self-directed, it’s goal oriented and all that. It’ll be what, like 10 Microsoft or something like that. So essentially that’s that’s the direction. It will become, once we get to autonomous intelligence graduate, AIGO will be, mark my words, AIGO will be the most valuable AI company on the planet. and and it’ll continue to be to be that way. Yeah.
Aashka (1:30:31)
Yeah, Manifesting that. Yes, and yeah, it’s possible. The harder the problem, the more the value and more the rewards. So that’s what history has proven to us. So it makes sense. So you mentioned about
AIGO using Chat GPT or like acting as a tool user as well.
Srini (1:31:01)
Telco.com SAP any of those tools.
Aashka (1:31:03)
Yeah, yeah. So basically, I had this weird idea. So let’s say AIGO or a system like it actually solves continual learning, the graduate level AIGO, the thing that OpenAI and Anthropic haven’t cracked yet. And say the Frontier Labs separately keep winning on raw infrastructure and compute as they are right now. So after AGI is reached by either path,
most likely is AIGO reaching
Srini (1:31:33)
But either path. It you it has to be cognition.
Aashka (1:31:36)
Yeah, it’s through AIGO so
Srini (1:31:40)
Are anyone else like AIGO? It does it doesn’t have to be AIGO only, right? Anyone else that’s building on integrated neuro symbolic architecture that’s solving cognition ground up, absolutely it’s a possibility. Just not through LLM, all I’m saying. It doesn’t have to be AIGO only, right? It could be
Aashka (1:31:52)
and not throw it. Yeah, yeah, yes.
Srini (1:31:56)
that’s using similar architectures like AIGO.
Aashka (1:31:59)
Similar, yes. So, the continual learning, architecture and the infrastructure scale combine into something that accelerates humanity towards superintelligence or intelligence explosion faster than either path alone. No, How do you see that world?
Srini (1:32:16)
Absolutely. So infrastructure for us is only needed on as needed basis for some of the applications. If you want to do DNA analysis, of course I need massive infrastructure, it’s a compute intense and all Not every application needs that. Right? so I and we believe by the time we get there, infrastructure is already available. It’s there. You know, I don’t need to burn the planet and I don’t need to build a, you know, whatever number of data centers that these guys are building, you know, all the power plant that they’re building.
I don’t I don’t need to do all that up. There are already existing enough infrastructure that that that we could use. And obviously depending on the application and the complexity of the application, computational intensity of the application. You know, we we would need more, but by then these guys are already building all that stuff. if anything, there will be excess capacity of that infrastructure. So infrastructure is is necessary, not sufficient for us. but what is what is necessary is the architecture.
It’s the architecture that’s the answer. Infrastructure is kind of backburner. It’s a back burner. We will we’ll we’ll need it wh as we need it, when we need it for certain types of applications. but other than that, it’s not I I don’t need massive infrastructure. I mean we need infrastructure. We need the data, we need the compute, we need the energy. But at a very, very minimal level. Like I said, the kid kid go we are running it on three year old laptop, right? so even for the commercial engagement that we had with 20 million customers.
We were paying under 10 grand for the so for the storage, for the compute, for all of the infrastructure combined per month for twenty million c consumers, right? So that’s very, very little, right? So no, we don’t see infrastructure is not our bottleneck, put it that way. It’s not our bottleneck. what’s bottleneck is be you know, getting AIGO to that graduate level, that it can learn anything on its own, like human, consuming same or even less resources than a human. That’s the key.
That’s it. The moment it gets there and after that, it’ll come up with the most efficient way to get to from point A to point B for many, many things. Now it can invent things.
Aashka (1:34:21)
Hmm. Yeah. So just to rephrase this question, let’s say AIGO or any similar architecture has achieved AGI through continual learning and the other components.
Srini (1:34:35)
Mecognition and concept formation. No,
Aashka (1:34:38)
Yeah
Srini (1:34:38)
what what we say continual learning is also the first point that Peter mentioned. Real-time incremental learning. That’s exactly
Aashka (1:34:44)
Real time incremental learning
Srini (1:34:45)
that’s continual learning, by the way, right? That’s a different way of saying it. And then content formation and abstraction followed by metacognition. These are the three. Yeah, keep going.
Aashka (1:34:53)
Meta cognition Yeah.
So we have reached AGI. Then how do you see humanity moving towards ASI or Super Intelligence or Intelligence Explosion?
Srini (1:35:06)
No, no such thing. It is the pure marketing term, ASI. Okay. AGI is ASI because AGI doesn’t have any biological limitations.
Aashka (1:35:16)
Okay.
Srini (1:35:17)
So the moment let’s say you are Aashka, you have all the biological needs. Let’s say you don’t have any biological need. You are intently focused on just doing the podcast, becoming the number one podcast on the planet with none of the biological baggage. You will become that. You will that then you will be the super intelligent.
In the podcasting world
Aashka (1:35:37)
Yeah
Srini (1:35:38)
You know what I’m saying? Because you know, you have to sleep and you have to eat and you have you know all of the physical needs, financial needs, and ego needs, and all of but you put all of that aside. That that’s what AGI is. That’s why AGI they have said the moment it becomes a financial analyst, it is one of the best financial analysts for your company that you could ever have. And same with every single job it takes, it becomes that.
Right? Because it is active twenty-four by seven. It’s constantly thinking with photographic memory. It c it it it can give you rationale about any particular decision it made at any point of time. Why did you make this decision, you know, fourteen months ago when I said this? AIGO will give you full details of what happened, how it happened, why it arrived at that e even humans can’t do that, right? So for to me, AGI is ASI because it doesn’t have any of the biological limitations of humans. That’s why it will be smarter than us in in many of these areas.
Aashka (1:36:30)
Hmm. Yeah, that’s very refreshing to hear. I have never heard anyone saying this, but yeah, I now understand from where you are pointing AGI equals ASI. So yeah, it makes sense. So let’s get into a deeper philosophical question. Just the closing question. So Rutger Bregman, the Dutch historian.
argues in Utopia for Realists that we already live in Cockaigne the medieval land of plenty that generations dreamed of. Like our previous generations for them, this is abundance, age of abundance, right? We have fast food available 24x7, plastic surgery to prolong youth, etc. And it has arrived. Fukuyama, the American political scientist, calls what followed a bleak paradise.
endless technical progress, no new vision for what life is actually for. King Midas is the oldest example for this. He got his wish granted, turned everything into gold and couldn’t eat. Aigo’s mission is built around four kinds of abundance as you mentioned, planetary, societal, personal, technological. So, what does abundance mean to you and how can humanity ensure that we don’t end up in a King Midas
kind of situation with AGI and abundance that follows.
Srini (1:37:54)
Right. Yeah. So at an individual level, I think we briefly talked about it. It basically enhances your your competence first. Enhances your well being, that is your health and quality of relationships and and all that overall quality of life. Right. so because it helps you think things through, you know, basically, hey, here are some of the actions you’re supposed to be taking. You you talk about working out, but you don’t actually work out. And and that’s why you’re getting the kind of
results and and these kind of things. So you may want to do just start small, you know, take whatever, you know, hundred steps, two thousand steps and and start incre incrementally and keep increasing maybe five hundred steps every day, something like you know, whatever. I’m just using that as an example, right? So are a quality of relationships. You you you the the pattern you fall in in these relationships and that’s a pattern you need to break. Without that you will keep getting the same kind of outcomes that you are you’ve been dealing with. So it because it helps you think through things, that is what to me is
Whatever, you know, what’s your definition of of your purpose, you know, your your your your joy, your happiness, right? So AIGO will actually help you enhance, you know, every single area of your life. The more you share it, the more intelligent it becomes, the more you’re and it is happening in a very, very private manner, right? Talk or text. it can see you and and kind of let you know, you know, all of this stuff, right? that is what to me human punishing is and and the overall goal is.
For it to for AIGO to see you thrive. You know, as an example, if you want to learn something, it will learn your learning style. Are you a visual learner? Are you you know auditorial learner? and it’ll learn your learning style. It also learns the subjects that you seem to understand you know fast, quickly than than others, and and there are things that you are slow. and and AIGO will help you kind of evolve help you evolve into those things depending on your interest. AIGO itself doesn’t have any agenda other than
It wants to see you flourish at an individual level, right? So that’s where it is coming from. Now, the philosophers you’re talking about, essentially they come from the school of thought, and that is something that I think we will deal with, which is when there is you know enough food and all that when when they when the struggle for the day-to-day is not there, you lose identity and you lose motivation. Purpose. Yeah, purpose, and therefore you fall apart. Well, to me, it’s about repurposing that.
You don’t have to fight for the food. You have to fight for something else. You know, there’s going to be, you know, a Maslow whatever in the the hierarchy of needs. You know, they will always humans always are really, really good at figuring out what it is that we want that gives me the the you know that satisfies your ego your this, you know, your you know, your your your fulfillment, your happiness and all that stuff. Now that’s what I want to pursue, right? so those kind of things will will always be there.
I I what what will be those kind of things? I cannot answer that right now.
Aashka (1:40:49)
Yes, yeah.
Srini (1:40:50)
Just as the human nature, we will always, you know, maybe we dig into
Aashka (1:40:53)
figure it out.
Srini (1:40:54)
that. Maybe we dig into consciousness very deeply and we do a lot of research. AIGO will help us do those things, right? there are many, many questions that are unanswered. Maybe we figure, you know, those things out. There are many problems that are not solved. You know, in some ways you know, we dealt with a shock kind of thing during pandemic.
during the pandemic. Some people thrived and and some people didn’t know what to do with their time. And some people were got into fights in the closest relationships, got too many fights and all that. But there will always be people who are needed to bring other people up to up to speed, kind of help them out in any which way, to bring the clarity, to bring that to show that support because some people will not be able to kind of you know adjust cope with the kind of change and adjust and all of this stuff.
but it’s not because not because they don’t have the food or safety shelter. It’s just because of the rate of change that’s going to come happen as a result of this, right? but that’s where other human beings will come into play. uplifting
Aashka (1:41:56)
Yeah.
Srini (1:41:57)
and kind of counseling and any any number of those things, right? Yeah. so I believe vi you know, I I don’t believe that just because we have the food and we’ll become what is that that movie, WALL-E If you remember the movie WALL-E everybody’s sitting on their
you know, big chairs and they’re like three hundred pounds each or a four hundred pound feature and doing nothing. I don’t I don’t believe that is gonna be reality. it may be for some people, it could very well be. but I believe we will move on to the next horizon, whatever that next horizon is, whether exploring space or exploring consciousness or any number of these things and and what else we can do for humanity to make you know everyone that much kind of move them along.
Now that you know we we we we need to get rid of the disease, we need to get rid of poverty, we need to get rid of, you know, all the climate change issues, any other stuff, right? How do we bring planet to, you know, best condition so it doesn’t kind of go through what it’s going through right now. These are all there are many problems that we have to deal deal with anyway. So good problem, there are ch challenges, good challenges. And with the technology like AIGO, you should be anything, AGI, we should be able to solve a lot of them.
And that’s where your energies will go. And there’s gonna be a lot more new things. What does the entertainment look like? What kind of things will get the attention at the time? There’s a whole whole new set of things that we’ll we’ll figure it out. I’m confident humanity will will figure those.
Aashka (1:43:21)
We’ll figure it out. Yes. Yes. Yes. So just one request as a podcaster in the end. So as a podcaster, I have tried building agentic workflows for post-production, editing, show notes, clips. They’re getting better, but they still can’t operate independently like a human editor or a human producer. I want to request
for a discounted AIGO for me whenever it gets to the graduate level.
Srini (1:43:57)
Once it’s available, we’ll figure out a way to give it to you for free for a period of time. Don’t worry about that. Yeah.
Aashka (1:44:05)
Thank
you so much. Thank you so much for answering all the questions very patiently.
Srini (1:44:11)
I enjoyed it. Yeah. you didn’t rush it. We had almost an hour and a half or two hours. So yeah.
Aashka (1:44:19)
Yeah,
hopefully I touched all your points that you want.
Srini (1:44:24)
I think quite
a few pretty much all topics are covered. If not, we can always do, you know, one more later. Happy to do it. and yeah, I look forward to the to the output. Thanks again for for having me. I appreciate it.
Aashka (1:44:36)
Yeah,
yeah. Thank you so, so, so much. And let me stop the recording.
Of course, we may have missed some super important questions. Tell us in the comments, and we’ll make sure to put on our Vasco da Gama hat, go explore, and find the answers for you 🤗
Advisory Note: This is a one-in-a-million serendipity story, so please don’t start accepting every random group invite or falling for spam requests 😄
