"AI Exposed" Part V: The Failure of Generalization
WethosAI — https://wethos.ai/resources/videos/ai-exposed-part-5-the-failure-of-generalization
Source video: https://www.youtube.com/watch?v=22NfbV3JHsI

Welcome everyone to AI Exposed. We are kicking off in just a minute, so hang tight. Really glad you are all here. We are going to have a lot of fun with this one. All right. Hello everybody. Hopefully you can hear me fine, see me fine. We're going to get this party started, kicked off. Um, just to remind everybody, this is AI Exposed. The hype must die.

We've been separating AI fact from fiction since 2023 and we are on our fifth installment here and I think maybe the crescendo um although we keep getting pulled back in. Topic for today is the failure of generalization. Um and with me I have my fearless co-leader Alen Capalik. Um just a quick backgrounder for you on your host today. So I'm Stuart McClure.

I'm CEO and founder of Cylance. Uh we sold to BlackBerry. I've also led Quiet and now I have my AI incubator Number One and our first launched company, WethosAI. And Alen, why don't you do a quick introduction?

Good morning everyone. Uh Alen Capalik. Um my background is a cyber uh like Stuart. We know I know Stuart about 15 years. We uh we've been in a cyber trenches together for a long long time. Um I was the founder of a company called Counter Technologies called GoSecure in cyber security uh endpoint detection response space. And then I uh I was a founder of a company called FastData that worked kind of early on with Nvidia and uh on some of the data processing things that kind of end up being all of there is today for um AI.

I'm very happy to be here uh talking about how incredible AI is at everything seen before.

Yeah.

know,

yeah.

Problem is universe has an annoying habit of doing new things.

Yes, that's right. We're going to have a lot of fun. We're going to go quickly too because we don't have a lot of time. We got lots of content as usual. But together we hold uh 38 uh patents in cyber and AI and over 10 billion exits. So we you know we've been doing a couple things in our career, but Alen's the smart one just to be clear.

Uh I just try to ride his coattails appropriately. But um anyway, let's get it kicked off. So, just as a reminder, this um webinar is really trying to it's not trying to shame anybody or shame anything. It's really just trying to tell the truth, trying to expose the truth and the reality of AI today and where it is going and where we believe it can be uh you know put to use the best.

Think of it as that. And if you've missed the prior ones, feel free to find those um on YouTube on our YouTube channel to check it out. One of the core anchoring tenets of everything that we're doing here is u really a call out to the different ways the human mind thinks. There's really two systems of thinking by most measures, although you could argue there are um ones that go past this, but but at the core there's really impulse, which is system one thinking.

It's just triggering on our cognitive biases that we may or may not be aware of. Then there's system two, which is what we think of as critical thinking. It's thinking forward and backwards and in parallel to try and derive the best answers to whatever challenges you're facing. And I believe AI is one of the few platforms where we will be able to get to system three level thinking, which I call system two at the speed of impulse or instinct.

So, we'll see how that flushes out. But I wanted to get that topic in uh early on. So here we are. The failure of generalization. Before we get going though, I want to encourage everybody. You've probably got an invitation to get onboarded onto the WethosAI platform. Part of the demos here require your involvement. So if you can um go on in, click through on the registration, get onboarded quickly.

You can either do the quick rapid assessment. And it takes literally 1 minute to answer those questions. Or you can simply just post your LinkedIn, but that will get you onboarded and allow us to go through a proper demo with actual registrants inside of the discussion. Also, if you have any Q&A or questions or topics or suggestions, we want you to go on in there and make those suggestions if you can.

If you have any problems, feel free to uh just open up the Q&A chat here. All right. One of the best ways to kick off uh any uh discussion around AI is with an Office reference.

Everyone meet Claude. He's an AI which stands for absolutely incredible.

Corporate sent us a subscription. Michael bought it a chair.

Claude is here to make everyone's jobs easier. Watch this. Claude, first task. Research this.

That's a stapler. Can Claude track, kill, and dress a mule deer? No. Then I'm not worried about it. All right. No better way to say that. Okay. So, as you've probably been noticing, there has been quite a backlash around AI, especially around university system and education because of the fear-mongering that occurs in the space in the industry that humans will be changing.

Human the human role inside of work will be changing. Of course, no one can really articulate what that means. So, the fear machine continues to thrive. So, we've seen quite a few um out there getting booed at commencement speech addresses just by saying the letters A and I together. But I encourage you to watch Steve Wozniak's version because his version of AI is actual intelligence.

And of course, he gets rave fan reviews on that one. And I tend to I tend to agree with Steve. I think you'll probably catch that as we go along the way. All right, let's let's start from the beginning. Well actually before this uh you could look at companies like Cylance that employed deep learning models at the edge in compute but before or after that you started to get into what we call this transformer-based architecture which is taking words and predicting the next words or letters

from those words and that is basically what we have today right this what I call gen one of AI which I also call AK which is artificial knowledge it's not wisdom it's not intelligence it's not experience its word prediction.

So, it's artificial knowledge. Now, it could be very powerful to have that kind of knowledge at your fingertips, at everyone's fingertips. And um it certainly does level the playing field of knowledge. So, we have to think about all of that. But um I've often called this a four-year-old uh drunk genius. And one of the great uh cartoons that came out recently sort of shows that and these are the usual suspects, okay, these transformer-based models.

And I want to make sure we're clear on what AI can and can't do right from the beginning. So as an AI model, it can literally just predict the next word. So if you have something that's dynamic in a request like, well, what is the date today? Okay, what you'll see is it presents literally um all the thinking in the world, but it searches for other tools like system time um and uh you know state um properties and tooling to be able to to understand and answer that question.

So the AI just to be clear is not answering that question. The AI cannot answer that question. So what I'm showing here is Ollama's loading of Gemma 4. And of course, no tooling wrapped around it. You can just ask it. Now, I asked it the second question, right? Who is the president of the United States? And of course, it is Joe Biden. Now, why is it Joe Biden?

Well, because the last training of this model was May of 2024. So, again, what you see a lot of when you ask questions, context, and things of that nature is all the tooling that gets put in front of it. So, there's no better way to express this in in this with this video. Okay? So, just know this is a bit of a long video. I'm going to walk you through this, but it is so funny we have to just play it.

And if you haven't heard of the car wash problem, you should you should try to look it up. Um, basically a user prompted, "Hey, I want to wash my car. The car wash is 50 meters away. Should I walk or drive?"

Of course, the model scraped 15 sources and wrote a detailed essay, including that the user should walk. So, So he says it's a minor glitch, you know, or decimal number. It's like the decimal number comparison. The intelligence explosion is imminent.

They have to tell them the truth.

This is important. We didn't solve them with the neural net, right? We hardcoded them. Whenever anyone asks either of those which was the strawberry and the decimals and now the car wash a Python tool call activates and injects the answer into the context. I'm going to jump ahead. He he he gets everybody out of the room. SHER FIGHTERS WOULD HAVE BOTH SURGEON IN 98.

OKAY. OKAY. It goes on. It goes on. I have to jump. This is our house. is

All right. you you just got a window into the the real world and how this whole thing works. All right. Now, why is this? Okay, it is a static system of data. course,

but there is extreme limitations compared to the human brain on training that uh model. So, I've often brought this up um in priors, but it's important to talk about it. It's that look, our brain has been developed on all human thoughts since the beginning of time. all experiential, kinetic, visual, auditory, you name it. But only a fraction of that has been actually verbalized.

And then of course only a fraction of the verbalized thoughts have been written down. And then only a fraction of the written down thoughts have actually been digitized. And only a fraction of the digitized words is actually used in training. This is the limitation of the transformer model. But don't believe me and don't believe uh what I'm telling you.

Let's let's hear it from Yann. Let me give you a very simple calculation. A typical large language model is trained with something on the order of 20 trillion tokens, right? Or 20 20,000 billion tokens. A token is like a word more or less. A token typically is represented on three bytes. So 20 or 30 trillion tokens each on three bytes.

That's about 10 to the 14 bytes. A one with 14 zeros behind it. This is the totality of all the text available publicly on the internet. It would take any of us several hundred thousand years to read through that material. Okay? Okay, so it's an enormous amount of information. But then you compare this with the amount of information that gets to our brain through the visual system in the first four years of life and it's about the same amount.

In four years, a young child has been awake a total of about 16,000 hours. The amount of information getting to the brain through the optic nerve is about 2 megabytes per second. Do the calculation and that's about 10^ the 14 bytes. It's about the same. In four years, a young child has seen as much information or data as the biggest LLM.

What that tells you is that we're never going to get to humanity.

Okay, so that and that's Yann's just optical nerve uh calculations, right? Quantifications. What about auditorially? What about kinetically through feel? Maybe we're giving too much credit to the four-year-old uh reference is the point. Now, one other sort of um great mind out there talking uh the real deal is Gary Marcus. He's a cognitive science scientist that has been fairly verbal on it that AI does not think abstractly like a human.

Let's hear world models or cognitive models is absolutely fundamental. Having ability to generalize abstract knowledge is fundamental. What we see in LLMs is always piecemeal. They don't really understand an abstract any abstract principle. So this is why they have trouble with arithmetic or lately we've seen these river crossing problems.

they have a lot of trouble with. So, a man and a woman in a boat have to get across and they just confuse it with other similar problems that um are in the database. They never really abstractly represent the notion of a boat or crossing a river um or what a you know wolf or a cabbage is, right? So, you need a much higher level of abstraction.

Um you need something like human values or Asimov's laws, something like that. I mean, we haven't gone into that kind of AI risk. But, you know, we can't even right now say something like be honest to a system or don't use copyrighted materials. Like, even if there there's a list of what is copyrighted in the training, if you put that in your prompt, don't use copyrighted materials, the systems don't actually have a deep enough understanding of those terms.

And so, they they can't follow it. So, we need, you know, to be able to follow those kinds of um explicit rules. algebra is about abstract relationships and the kinds of stuff you guys are doing are compatible with that and turns out that most neural networks as we currently build them are not compatible with that. So there may be some neuro-symbolic hybrid much better than we have right now that really kind of smoothly transitions between arbitrary data in let's say um a corpus database and also having those abstract representations.

All right. So let's get into that concept this what we call gen 2 AI model building and that is really around a number of the preeminent founders of the transformer space but before that deep learning and convolutional neural nets for example with Yann LeCun um so a lot of people have gotten a lot of funding around this space so advanced machine intelligence with Yann, thinking machines for mirror uh human DeepMind of course uh SSI with Ilya and then Fei-Fei with World Labs.

They're all trying to solve for this problem by building 3D world models and then training the the systems inside of them which is a valid approach I think for sure and will certainly bridge us. But where we see the puck going in the game here is around Gen 3 and what we call thinking or as Gary said cognitive models. Now to do that you need a number of factors.

One is infinite memory. Uh and two is to be able to change your weights so that you can self-improve. You can self-learn. As you learn more information, you update things and only through that world we believe you can actually build a thinking system that's anywhere close to a human brain. So there are a number of folks that have just started in the last few months honestly to get funding in the space.

So you can track on recursive inherent engram and merendal which uh actually engram is the asterisk. I misspoke on that one. I'm going to show you that in a second. And then of course you can track on us a little because this is exactly what we're doing and what we're building here. So gen one as you know is not general knowledge. So here's the core of the pro I mean is general knowledge.

So this is the core of the problem. Yes, it's general knowledge. You know maybe you could call it near infinite knowledge. That's fine. Okay, but it doesn't have any wisdom, doesn't have any real world experience. Um, and so what happens is all the answers because it's generalized, you get a regression to the mean problem, which means you ask the same question or similar questions to all of these AIs, no matter which ones you pick, and you're going to get basically the same answers.

Okay, now this is really important because if you're all getting the same answers from all the same um LLMs, it doesn't matter if you're picking all the LLMs from different vendors, they're all built the same exact way with transform based models, then you're never going to get true uniqueness. You're never going to get novelty. And that's exactly what we're going after here at WethosAI is we're building this artificial individual intelligence to create to discover and create that unique alpha wave for you individually.

And that's the big difference. So even Ilya sort of admitted this just a few months ago I believe that was about six months ago I guess but models generalize worse than humans. So all the pre-trained models are pretty much the same because they pre-train on the same data, right? These models somehow just generalize dramatically worse than people.

It's super obvious. Okay. And so he sort of threw up his hands. Hey, it's back to research. Of course, he was sort of promoting his own company, but he was not wrong. So let me give you a real world example. This just happened. If there's any World Cup fans still tracking uh the FIFA World Cup, um you'll see this. you'll see everybody pretty much on the pitch uh with these pink shoes of some sort or some variation.

Well, there is a story behind that. Every single one of these brands relied on a common brand analytics prediction system that leverages AI to predict what would be the most popular visually appealing colors and patterns for a brand at any particular time and you could go into the future uh with that model. So, of course, all of these vendors, um, I'm sorry, all of these brands used WGSN as their vendor to do this.

Well, of course, that meant that there was no distinguishing factors, no differentiation. Everybody looked like they coordinated, but that is not the case. So, everybody was trying to predict and because remember this takes, I think it's 12 to 18 months to build a shoe like this. So they were predicting it 18 to 24 months ago and all leveraging the same um uh the same vendor which used all the generalized models.

Now for example the only thing that really stood out is Messi's specific shoe um the Adidas F50 uh L Ultimal Tango. So, the last tango and they didn't follow the report. Obviously, they leveraged the Argentinian colors and sort of a throwback to a shoe and so now they stand out, but that that was just sort of dumb luck. There really wasn't too much prediction.

It was humans in the loop on that to actually make the decisions live. Now, here's the asterisk on engram. So there is a number of companies out there right now claiming they can do sort of they're learning with infinite memory things of this nature but there were actually four engrams all with the same language doing the same sort of thing.

So, we're not sure who the true engram is here. And we are wondering if everybody about 12 months ago started to ask AI, say, what's the coolest name out there that we could try to get a hold of that's going to represent true thinking? And of course, all the LLMs probably gave a similar variant of this word engram. So, something interesting to think about.

All right. Now, here's Sutton, which I think is this is pivotal change for us to think about. The difference between novel and good. So, let's hear it.

I'm sure you've all heard it before. This is the one about the researcher whose work is being evaluated and the review comes back and says, "This work is both novel and good." Unfortunately, the parts that are good are not novel and the parts that are novel are not good. My first point about AI is that this assessment applies exactly to large parts of AI as we know it today.

When we ask an AI for an answer from the internet or to summarize a document, we don't want it to be novel. We're happy if the if the quality of the answer, the goodness comes from the source material. And when we ask for a fiction or a novelty, the AI can give it to us because its processing is in part stochastic. Every decision that the AI makes can go multiple ways and it will go different ways and produce a different trajectory every time.

The trajectory can be random and thus novel or it can be based on the training data and thus good because the training data is good sourced from people or reality. Thus the trajectory generated by the AI system is either novel or good but it is never both. It's never both at the same time.

So important point, right? It can never be novel and good at the exact same time. So what the industry has created because of this token system is really one of the first cracks that we believe that we should start tracking on which is this compounding token environment or ecosystem that's been built. So Alen, I know you get passionate about this as well, but I call it token tsunamis.

Well, yeah. I mean that's what they are or uh I like to call it a refeeding tax right that's the you know that's the it's why Anthropic and the rest of these guys you know their revenues are growing like crazy you know they found um a use case which was great for coders right they created Claude Code and now suddenly everybody's doing this the problem is that about 70% of your tokens that is being generated by Claude Code is basically done so either for Claude you know Claude to um to reason.

Okay. Okay. So for every for every one that that gets generated for the reason you're actually paying for that and then also you you're you basically every time it makes a mistake then you have to reprompt it to basically like you know to fix the mistake you just made and you know and and it keeps repeating these things from one agent to another because of the context issues which is why uh you know we are very passionate here I know I am and been for been there for a long time is that you know without AI having actual memories like the way we Like we're talking to everyone right now and their brain is actually remembering this, right?

They'll be able to recite this to somebody else and stuff like that. And when I say actual memories, I'm not talking about writing into a file and then injecting into a context and all stuff. I'm talking about actually being able to remember inside a deep neural network. know

steal

okay I won't go ahead go ahead I won't I won't and And Alen will probably remember more of this than I will be. He's a little younger than me, but yes, we that's how our brains work, right? We remember as much as we can and then we uh iterate on that memory to ensure that it is accurate. And a lot of people are starting to call this out.

And this is I would call the first crack in the dam where people are starting to pull back really quickly. And quite honestly, if you use uh you know, let's say Claude Code, you know, /clear is your friend, man. /clear you're just clear it.

All right. So this to the industry is starting to call it token maxing. Okay. Which is really starting to just get out of control. And this out of control nature is starting to really pucker up uh the enterprises who was sort of like letting it go uh you know wild west style and the uh those in charge are really not allowing that infinite budget scenario to work.

So, you know, these AI cartels, I think, are really really worried. They're worried about a couple things. Um, of course, this context window and the compounding token problem, but they're also really worried about the open- source uh model problem. So, as you're starting to see, and we've, you know, we've been seeing percolations of it, but now it's starting to get a wave of momentum.

A lot of the models, especially out of China, for example, are all open- source and they're incredibly cheap and incredibly good. So GLM 5.2, for example, in some of the I mean, honestly, a lot of the benchmarks, it is right there, number four, sometimes number one or two, right behind Fable in terms of overall efficacy and accuracy. And this is an open source model.

You can download it today. Uh run it on your B. Well, you got to have a beefy box, but you run it on your box. Another data point is how many people are downloading these models right now on Hugging Face. It tends to be an early canary for where is the market going. And you can see it's all of these open source models from China and specifically 5.2.

So, you have to think of open source as as almost a correcting moment here. So, here's another thing just came out today as you probably saw in the news. I had to add it in. But you know, genie's out of the bottle in the sense that, you know, China has been um crawling the world's compute for intellectual property for a 25 years to be fair um on the internet and and really almost their sole purpose from a nation state adversarial uh perspective is always intellectual property.

It's not so much to you know take stuff down or you know really hurt um or ransomware things like that. really around just getting intellectual property because they do not believe in it. They do not believe much in patents or anything like that. So, so you can't really blame them. Okay, we've known this. This is their MO since for a long time and they are now of course been caught going after for the last few years going after um so Alibaba's Chinese uh company going after Anthropic's models to uh what do they call it brazenly and illicitly attempting to extract AI capabilities.

This was an inevitability of course it's been happening for years and they're starting to um leverage that and put it into the open source. Remember, Anthropic brazenly and illicitly extracted all the data off the internet from minds like you and me, right? So, I have like 12 books. They're all in Claude's model. And actually, if you go to um uh it's in theweights, is that it?

Oh, no. In theweights. org, I think it is. Shoot, I should have done this. Um check it out. It's or in theweights. org, I think. But you'll see you can query your name. you can query the works of Einstein or whatever and you'll see the actual um all the models and where your data is in. All right, but this is what Alen was trying to uh talk about was around what Satya has been talking about the last few days or week here is really comes down to at the end of the day a frontier without an ecosystem is just not stable.

just to you know read exactly what he said. Importantly, human capital does not become less valuable as token capital grows. It only becomes more valuable I believe uh I believe human agency oh becomes more I believe human agency will be the driver of token capital growth. Humans will set ambitious goals, connect dots across domains, build relationships and recognize patterns that matter most.

Without human direction, you have compute running in circles. And this is really the core of the issue. So where are we seeing AI being used in the enterprise? Okay, in business today it is almost universally inside of decision-making. So you can see the latest Microsoft uh report uh released in May uh specifically uh identify decision-making as a core part of it.

But to just make decision-making in a generalized world means what? Right? Generalized output. That means that you can never differentiate between you and a competitor. You saw that in the um the World Cup pitch shoes. So what you really need is a driving force around cognition. You need to accelerate your own cognition in the space. So this is what we're going to show you here inside of Wethos.

If hopefully you've gotten on board and are into the um WethosAI platform. So we're going to go and jump on in. I'm going to show you what we've done. So, inside of WethosAI, when you jump in, um, you get a nice sort of clear dashboard of yourself, your traits, your biases, how you actually think, and you can start to simulate decisions. You can do Monte Carlo sort of run through on, hey, I need to be able to achieve A, B, and C with XYZ people.

What's the best approach? You can roleplay a conversation that you might have. You can get any sort of coaching. But where it really gets fun is inside of the flow or the Brainstorm. So in the welcome to WethosAI Brainstorm, you'll see yourself um here. And uh some of the early folks that came in yesterday, I really appreciate it. If you haven't already, please try to get on board and get in here.

If you have any questions of me, I'll try and answer them as we go along the way. And I'm going to show you how to do it even though I'm not even in the Brainstorm. Okay. So first of all, I ask a simple question. Hey, Marty and David are uh you know on the um uh on the webinar here. Let's let's see how do they think and make decisions. So, as you can see, you could categorize by core thinking style, decision-making, execution, structure, and by going through that onboarding process.

We actually understand how you think, how you see the world, and how you naturally gravitate to certain decisions. Now, it doesn't mean that you're 100% predictable. Of course not. It means that these are your natural styles. It doesn't mean that you can't think in another way. Of course, you can, but it might create cognitive load for you.

And these are all really important factors to remember. So, if you look at Motti, we would call him her uh the anchor uh and we would call David uh the catalyst. And they are different and how they see the world and how they make decisions and how they work together with other people. And that's how we infuse it inside of the Brainstorm.

So I can ask simple questions like what kind of projects would each person truly thrive and then likely struggle and we can go through it. So Marty would be you know in a high structure high stakes environment but might struggle with ambiguity and rapid pivots and that's okay. Uh that it doesn't mean it's right or wrong. It simply means that's the more natural style and it depends on the project that they're working on.

David, high innovation and high momentum. Well, you might struggle with high routine and rigid constraint. By the way, David and I are, you know, brothers from another mother because this is me, right? I struggle with, you know, routine sort of rigid constraints. All right. So, give me a scenario where these two might not get along. All right.

The last minute feature pivot is a great example, right? So, David's going to be energized by the new vision. He's going to sketch out the concept, rally the team, start building it right away. But Marty might say like, "Wait, hold on a second. Hit the brakes." And you need people of all different thinking styles to be able to um determine the best decision-making to get the best outcome.

And that's what uh WethosAI guides you through all along the way. So I can get okay, guide me on coaching my to be more flexible and I can get specific bespoke coaching. Um I also need their opinion on launching. Okay, that's for our next step here. Hold on one sec. All right. So, with that, we're going to walk through a system three thinking example on the platform, which you can see here inside of the Brainstorm, which is, hey, I really need uh Motti and David uh their opinion on launching this Cognitive Twins capability to the public.

How would each receive the directive to launch the offering? So, now we can see, ah, I I know exactly how they're going to think and how they might manage through the launch. Um, okay. And then so I can ask explicit questions. Hey, what would David say about the problems with our launch strategy? Well, we're moving too slowly, losing momentum.

This sounds like you, Alen. Uh, the user experience is better. It's getting buried under tech constraints. We're losing the core narrative. We aren't being adaptable enough. These are all great. Okay, so um, let's do this. Let's actually challenge the whole group with a question. So, I'm going to say, "Hey, pick two members of this group that would act as strong devil's advocates because that's what we want.

We need something to keep us grounded." So, here's a here's sort of a live poll. Now what it's doing, it's taking everyone that's on boarded inside the brainstorm, understanding their unique cognitive patterns and then applying that into this brainstorm for us to get the best outside in view of the actual launch plan. Um, all right. So here we go.

Stephen Jess Moore would I can vouch for that on the Jess Moore side. Um, uh, as a CISO, etc. , etc. He brings in um he separates emotion from hype isn't easily swayed by the excitement of a new idea. This is fantastic. So his devil's advocate style, you know, he'll ask the hard direct questions that others might avoid to maintain harmony.

If an idea lacks a clear step-by-step process or poses security and compliance risk, Steve will flag it immediately. He thrives on ensuring. Okay, wonderful. And we need a Steve in almost any conversation. And we need a Jess. All right. So that sort of gives you this sense of okay, can we make decisions at the speed of compute here by including different perspectives and ways of uh seeing the world and solving for problems.

So now what we're going to show you is what we call our auto Brainstorm. So I'm going to take this exact Brainstorm. Now I did not this is all purely live and I have not actually tested this out. So we're going to see how this works. We have 130 something folks inside of this Brainstorm. So, let's see how it performs. Um, we're going to go in and I'm going to create what I call a Simulate Brainstorm.

We're going to do an auto Brainstorm. And what I'm going to tell to do is I'm going to say, hey, act as a supportive guide to blend all the opinions and insights from each active member to come together and agree. Use an ultra summary format and always anchor on a win-win mindset to resolve conflict. The objective is to agree on a location budget for an AI conference to be held in Orange County somewhere here in California.

We are calling the conference the AI supercenter conference of 2027 and we're planning on having our guests in this group join. Uh but we need their input. So we need David, we need Motti, we need Jessica Molstead if she's on. Uh and then include the WethosAI employees as well um to make sure that we get their voices because we know that they're on in a minimum of 10 turns.

So, I won't clear the decisions, but we're going to let we're going to kick that sucker off and we're going to come back. Okay. So, this is automatically pulling in everybody's opinions and thoughts and perspectives about um the conference uh the location and uh you know where we should do it and a budget for it. Okay. So, we're going to let that run and we're going to come back to that.

All right. Final thought here before we get into the um the big crescendo at the end. And we get this question. I get this question from all my peers uh CEO peers or C-level peers and I probably um you know ask Alen too many questions about it too. But are we in a bubble? Are we in an AI super bubble? most

No, we're not. We're not in a bubble. Actually, it's it's different this time.

Okay. Tell me how it's different. It's not that's I'm being sarcastic. That's exactly what they were saying for the you know for that's exactly what they were saying 1999 2000. It's different this time you know. know

right and I'm going to tell one quick story. So I was fundraising in 1998 for a little company that I was starting called Foundstone. And I went through the valley and to and met with dozens and dozens of potential investors. And almost every single one of them looked at our pro forma uh plan and said the same exact thing. You're not spending enough.

And so, you know, after about our 10th or 12th word on that, we knew we didn't need that much money. Okay. But but the valley was hyping the the dot e-commerce bubble so strongly that to even play in that space, we had to uh spend more money. So, as soon as I started to open that aperture and say, "Well, look, we'll spend more money in marketing."

Sure enough, we got funded. Now, um, fast forward to March of 2000. That was 3 weeks after we closed our funding, and I got the dreaded call from the investor, right, that that gave us the money and said, "Uh, Stuart, I only have one question for you. Um, did I back the right horse? I'll never forget this moment. And of course, I said, "George, you backed the right horse."

And luckily, of course, we survived that, but it was an incredibly painful experience. You know, after you raised that first round, the bubble burst on the dot, you're, you know, literally out of money in a year, year and a half, two years, and and they they have deep pockets, but short arms. They're like T-Rexes when they start to when the bubble bursts.

And so trying to get money out of of anybody at that point is incredibly challenging. So you are starting to see all of that happening again. Now you could argue there are some differences on the positive and there are some differences on the negative if you're trying to compare it with the dotcom bubble. But we are absolutely in a bubble.

And so what does a what does that mean for it to be in a bubble and how how is everybody looking at it? A lot of different ways to look at it. Um, so what I'm going to show you is uh a dashboard that I created called the AI crash dashboard. People are going to hate me for this, I know. Um, I'm about to release this on my public GitHub as well.

If you have um interests in it, let me know. I think the more interest I have from you guys, the more I'll probably release it. Um, what looks better, light or dark on this mode? Is it white? Probably white.

Well, you know what I mean.

I know. You always like dark though, man. I don't know how you do it. Sit in the dark all day long. I think the the white looks a little better anyway. Okay. So, what we've done and I've put in my own basket. Okay. I have a an AI basket that I put in a while ago and as you can see, it's up very healthy from the beginning and obviously I got in early, right?

So, it's about 82% on the return. Um, but I can see day-to-day how it goes. And I can also create like hey I'm hyper sensitive to it. Okay then I need to sell uh but I'm you know I'm low sensitivity I can ride it through especially with what I know about the markets and each one of these elements from all the canaries that I have a financial canary and adoption canary hardware and GPU glut canaries etc etc.

You're seeing that these also used to be green about a month ago by the way. So this was all green. Now we're looking at watching. So the sigma values are what shows the risk is increasing. So as you can go go through all of this now the fun part becomes let's compare it to other bubbles. So I have the housing bubble in here which isn't too relevant.

So I keep it um unchecked here but I have the dot bubble I have the AI bubble both in NASDAQ S&P and I run a CAPE analysis on it. Now if you look the yellow is the predicted path. This is with um basically at its base, okay, of zero. I don't play with it too much because I've I've already gone and played a little bit and I feel as though we're at a pretty good prediction path here because look what's happening.

So from the very beginning on my model here, I have predicted this sort of first uh bubble heightened and the first bubble drop. And that's exactly what is happening, right? So, you're going to start to see, if this is accurate, um, a a drop down to the first bubble, um, sort of call it a mini step, um, and then back up into the peak with a projected peak of February of 28.

So, you could continue to ride this or, you know, maybe I'm completely off completely and and but you can play with it. You can come in here and do that. And then we also have the other broader markets um as an overall calculation around a risk-off score which is pretty interesting sort of day-to-day seeing all of the variabilities token costs over time I think is variable or very important rather to track on because as as costs start to go down you're starting to get commoditized and as as prices go up you're going to have a big push to go to open source so they have to respond but they can't respond because they cost too much and then you have other tiles that go into it.

I think the one of the most interesting parts probably relevant for this conversation is the layoffs versus hiring discussion. So I have a tracker that measures the layoffs versus discharges um since 2020. And as you can see um we are pretty much totally in the normal. I mean there is no you know more layoffs than hiring. In fact it's quite the opposite right now which I'm going to show you in a second.

So, so that's my quick AI crash dashboard. And if you have interest, let me know. I'll add you to it. All right. Now, Alen, I know you're passionate about this one.

Oh, yeah. That's right. Uh I guess we've seen in the past few months uh you and I you know uh going through the the valley again is you know they have this new bias which is you know you come out of Stanford you have a PhD you know you you you you basically can you know can call however amount of money you want to basically to work in AI you have you come out of uh you know you're an intern at Anthropic for two months you

come out and get $200 million like I don't like I I don't understand like why like you know Um but but you know but there there is definitely uh there's definitely a bias against you know what I call hackers like you know and but but if anybody has ever if anybody ever knows the history of the internet the history of everything we're using today and everything it was built by hackers it was not built by PhDs it was not built by it was built by guys

who were doing stuff and failing and then doing stuff and failing you know um you know from Linux to you know to uh you know Ken Thompson Ritchie like you know C and Unix and you and uh and all these things and and I mean like these are all built by by hackers and all the tools that AI is using today you know a lot of the CLI tools and everything came out of you know um uh Free Software Foundation and the internet like you know

uh and all this stuff like you know so so um the fact that the fact that that that today's investors are thinking that we are not going to be the ones and I say we because I consider myself a hacker like you know I don't have a PhD and everything but like I've I have about 35 years of you know hacking things up and everything like and doing doing really some uh some crazy stuff um that we are not going to be the ones to actually

solve the transformer problem.

They're completely wrong about that. They're they're absolutely wrong about that. They're you know um and you know and and and there is some really great stuff coming out. all

Yeah. Exactly. And you know they're you know they're trying they're trying to like uh optimize things and doing stuff because they don't have the compute and you know and they're you know and and somebody asked a question actually can we trust the models out of China and my answer to that is it's extremely hard it's it's becoming almost impossible to ignore them.

Okay that's the problem. The problem is like you can't ignore them. And the issue is also that they are actually uh you know they're open sourcing the weights. MIT

licensed the weights. You see it right there. I mean, you can even query the bias that's in there. Yeah.

I I I think I think there is a nefarious reason for that, right? Like, you know, I think uh there's a cultural reason behind that, right? opinion

is that if they flood the market with if they flood the market with the models that are as good or almost as good as these models from Anthropic and OpenAI and everything suddenly Anthropic, OpenAI, and Google and these guys are not worth as much money and you know and as we've seen like you know even Microsoft now is looking at these open source models saying like listen like we have compute like if we stick these models on there we can have we can use them.

They're really really good at what they do. actually

well that was the other post I forgot to put in here but it's Satya saying yeah they're moving away and going into DeepSeek model to use.

Yeah. Yeah. And you have now four incredibly incredibly competitive labs out of China. Z.ai which is the GLM 5.2 model which is absolutely insane. Like you know as a matter of fact um the founder of Z.ai basically said that you know they're going to have a mythos style open source model by December of this year. Okay. Now remember US government actually banned mythos right or or or you know as well as that Anthropic actually didn't want to release it because they're quote unquote it was too dangerous right?

Well you're going to have a model like that that's actually open source and anybody can run it like you know basically by the end of this year. And I think also the MiniMax models uh you know coming out of uh China Kimi 2.7 K2.7 and uh and obviously DeepSeek you know and uh and Z.ai just raised what 7.2 billion at a 55 billion valuation like no so they're you know they're and remember Z um DeepSeek for instance like

came out of the actual hedge fund that was in uh in China right like you know and they were doing they were doing actually like you know financial uh modeling and quant stuff and they you know they figure out okay this is a much better like way to make money you know because there's just so much money around it and I think that I don't think you should ignore Chinese models they're all on Hugging Face you guys go to Hugging Face you know.

my

okay

right

so the final thoughts um around this concept is it's being proven out already in terms of Jevons paradox where it really does sort of debunk the AI job apocalypse problem um in the sense that you know it's real simple as technology use goes up so does consumption that's sort of the model behind that and as you as you see um human capital go human capital becomes more more valuable this is Satya's words okay so all of that is super powerful you have to

keep anchoring on that and just know that the more that technology we who's the more valuable the human mind in terms of being that strategic partner, creating that strategic judgment, owning the accountability for the decisions, infusing ethics and EQ, all of that has to be a part of our job force and our minds in effect.

And you know, I I almost put some studies in here that are showing that the more AI usage, the less critical thinking we have. So, you have to continue to engage. In fact, if you were to do anything today you with AI, go and ask it for a Socratic method um to help uh empower you on critical thinking skills. Just do that and and and walk through a first principles approach on how to think about the problem.

And lastly, like I said, you know, humans will be in the mix for quite some time. Uh they're the only ones that can be held accountable. you can't go fire an AI that that answers something incorrectly. Um, you don't have the real world experience yet. Even with the 3D world models, you might get sort of visual and and connect or non-kinetic um uh or or maybe they'll go beyond.

And of course, all of these models today have just very limited context and really no mission or purpose. Okay. Now, as you probably saw in the very beginning, if you were uh the first to show up, I've actually built a complete model on myself that come is the closest thing to true cognition that I've seen out there in the world. So, you've seen some other uh twins out there like Reid Hoffman's and such, but they'll all admit that these twins are really just um sort of memory recalls.

So, if I train it on something, it'll recall it. If I if I ask a specific question, it'll answer a specific way. You can't get an unknown answer, which means I haven't trained it on this uh question, and the answer has to come up from its own cognitive map. And that's where you get to true novelty and newness. And so, let's uh let's give this a shot.

So, I'm going to go ahead and log back in. Uh, we'll call this Alen. Uh, and we'll log in. Didn't like that. It's already rejecting me. Hold on. Hey, how's it going?

I'm doing great. Thanks for taking the time. I'm excited to tell you about WethosAI and what we're building. Where should we start?

Well, first tell me your journey in cyber security and now with WethosAI and cognitive AI. I've been building in cyber security since the early days starting Foundstone in 1999 and later leading McAfee Intel as global CTO where we pioneered using deep learning in security at Cylance. That path led directly to WethosAI, where we're now applying that same expertise to build truly intelligent Cognitive Twins.

What part of that evolution interests you most? Well, let's switch gears and tell me what you'd like to do for fun on the weekends. I love getting out on my bike. I've been cycling since 1982, and I even went to the Olympic Training Center in Colorado Springs to train. Do you ride or have any other favorite ways to unwind?

Well, I want to know, give me some advice as an entrepreneur. My biggest piece of advice is to focus on building true thinking machines and not just a wrapper. The real value is in proprietary cognition. You also need to build deep moats with patents like Alen and I have done rather than just chasing trends. What's the biggest challenge you're tackling with your venture right now?

Well, give me some job advice on how to thrive in the new world order of AI. Focus on building a Cognitive Twin that maps your unique memories, experiences, and biases. You need to leverage system 3 thinking for high-level instinctive decision-making rather than relying on generic AI. How are you currently adapting your workflow for AI?

Well, I'm focusing a lot on critical thinking skills and capabilities for sure. That's the most important area. How are you currently empowering your team to leverage those critical thinking skills amidst the rise of AI tools?

Well, I'd love to engage more, but unfortunately, we are running out of time. So, will you now say goodbye to our AI exposed audience? Got it. We'll wrap up. To all the AI exposed leaders watching, stop settling for generic tools. Build a Cognitive Twin that's truly yours or get left behind. All right, there you have it. Thanks everybody.

Um, let's go back and see what happens in our final seconds here around the simulator Brainstorm. All right, here we go. So, I went in, of course, and asked for the simulation to begin. Darren Thomas started it by welcoming everybody, okay? And setting the true tone of where we're going and what we need to achieve. And then Alen came in and started to give the opinion we need to stay away from traditional stuffy hotel ballrooms.

This is pretty much what Alen would say. If we want if we want this to be a super center, we should look at the Irvine Spectrum. Oh my god. See, that's where the Cylance offices were. You didn't even know that. uh or a custom uh tech campus takeover. Action is better than deliberation. Suggest we aim for $1.5 million budget ensure we have the best hardware network.

Okay, now Jenny comes in. Now she's our grounder, right? Wait, 1.5 million? That's a significant jump given we're just um started the planning phase today, June 25th. I want to make sure we aren't falling into the planning fallacy here. We need a structured approach, the budget. Okay, that's pretty much what uh Jenny would say. And then uh Motti jumps in.

I agree with Jenny. They're pretty much cut from the same block. From a security infrastructure perspective, tech campus takeover is a nightmare for data integrity. We need a venue with established high-grade physical and cyber security controls. Couldn't agree more. Jessica jumps in. David jumps in. Goes on and on and on and then final decisions are made and automatically uh assigned.

So we've decided on this AI supercenter conference to be held at Ritz-Carlton in Laguna Niguel with a total firm budget of 1.1. All right. Ensuring a balanced luxury security. Now I can simply ask say uh everyone on board with this and they giving their thumbs up then gives us accountability and action steps for the next move on all these things.

All right, so with that said, thanks everybody for joining. I think we'll get to a couple questions if we have just a minute or two. Um, let's go back and see if there were any questions in this mix. Uh, let's see.

We have a few questions on the Q&A uh on Zoom.

Okay, let's see. Do you want to call out a few?

Yes, I already answered one of them. Can you trust the models out of China? So, I think we we kind of covered that. And then, uh, do we see AI evolution more similar to EVs or quantum? Um,

oh,

just

I I I just want to one of my closing closing thoughts here is that like we're not anti- AI at all. As a matter of fact, we're completely opposite. Like you know what what we are anti is AI hype. Okay. And uh you know one common denominator that you see from all the things that Stu has showed you especially when you see researchers and scientists and hackers basically like going out there you know from like you know u basically saying things that these people know how AI works.

They know how how neural network works. They know all these things and they're basically saying listen you got to be very careful thinking that this thing has a soul okay or it has all these things like you know. So um we we are very very bullish on AI. I think it's it's an incredibly powerful and everything but I think we are just on a precipice of of what's actually possible.

We this is a first step of many steps in front of it and I think that yeah yeah I think there's there's a lot of evolution here to to be to be had as with EVs like EVs first started like you know in the '90s right like you know and then you know somebody had to actually scale the whole thing and and understand like you know the battery technology and everything else.

So, so yes, there is we are kind of in 1990 I would say 1995 of EVs with AI like you know what I mean like so so that's kind of my my uh you know my position. I don't know what you do.

Yeah. No, I agree. Look, I think even the current AI has plenty of usage um as a knowledge system. Okay. Just not as a real world system of implementation as of yet. And that's what Satya is noticing and calling out. You need a full system. Um I thought Steve I love to see Stephen here. He's probably not here anymore but um this is a great question.

Can this approach be affected by observer effect? And so just to remind everybody observer effect the scientific principle that the act of observing or measuring a situation, phenomenon or system inherently changes state or behavior. It's a really great question and I think that um there is potential for that. Um, if you build a true learning system, if you build a true learning self-learning system, uh, which is what we're building, you could have that observer effect in play.

Now, the trick is just being aware of the observer effect and incorporating that into how the decision-making is done. To me, it's awareness that matters, uh, not that the effect is present. So, anyway, that's uh, simple, but of course, you could always text me on that one. All right. Well, I think that's it. And um how was your video created?

Oh, yes, that's a good question. Uh lob that question into an email to us and we'll answer it more appropriately. But we are wrapped up on time. I want to thank everybody for joining and uh look forward to the next time. Cheers everybody.

Thank you all.