"AI Exposed" Part IV: The Inconvenient Truth
WethosAI — https://wethos.ai/resources/videos/ai-exposed-part-4-the-inconvenient-truth
Source video: https://www.youtube.com/watch?v=nm858I_IjuM

This is AI Exposed, the hype must die, part 4. Uh we didn't think we'd do a part four, but we got called back um to do a year-end wrap up of 2025 and give our uh modest predictions for 2026. It's been an incredible year. So, we're going to run through it and we've got a lot of content as usual. So, apologies for going rather quickly, but just feel free to um tee up any questions um in the Q&A and uh we'll try to tackle them at the very end.

All right, let's get it going. So, if you are not familiar with me, I'm Stuart McClure. I am um 30 plus years in cyber security. Uh the last 15 have been almost exclusively or almost yeah a little bit more than that uh in AI and algorithms and understanding how to apply that in a predictive nature inside of any solution. Um so I started an AI incubator called Number One and our first company out of there is WethosAI. Alen, do a quick intro.

Yeah, good morning everyone. Uh, very excited to be here. Alen Capalik. As like Stuart, I have um 25, 30 year um background in cyber and AI. Uh and very excited to talk to uh with Stuart this morning about our AI hype.

Yeah, we have a lot of fun with these things. We keep it a little fast and loose, so apologies for any of that, but we'll try and get through it as quickly as possible. The real uh rules of engagement here is um you know as Sergeant Friday uh represented you know we are uh just about the facts uh we try not to get too emotional about any of this stuff um but at the end of the day we are we are human so

apologies upfront yes AI is incredibly invaluable and uh it's just not always in the way that you might think uh of it um and things are going to get fast and messy as we get through all of this stuff we do have some live demos at the very end. So hopefully you'll stay tuned for that which will be a lot of fun.

And we are using the pretty much the latest models almost exclusively in everything that we show you. So it's not one of these like we're looking back to sort of prove something. We're looking at the current models. And um please I get this from time to time. Please don't take it too personally, but do keep us honest. Like if you feel like we've said something that's inaccurate, incorrect, or not representative, please let us know.

And then again, just put those questions in the Q&A. And a reminder, demos are live, unless I'm calling it out as a video embedded. It's going to be live. Uh, so that's sort of the fun part of it all. All right. So, as you saw or heard before, we've done three of these this year. Our fourth now is um the final one for the year and wrapping up 2025 and looking out at 2026.

For those interested in WethosAI and personal AI and delivering that in the workplace, you can go to the WethosAI website. You go to sign in and you can create your own account here and you could try it out to let me know what you think. All right. Uh the famous line, you can't handle the truth. I think you guys can. I think you guys, the folks that really want to participate in this world of AI, want to look past the hype.

They want to look at what can really be used and how we can apply all of this into our day-to-day lives to be better, to be smarter, to do more, and to be more well-informed. So, that's what we're going to cover uh today. Not just the predictions, but we're going to talk a little about the future with post transformer uh AI and human in the loop and the requirement for that.

And then really personal AI. That's what I want to end on. Um all right, so let's talk about the key concepts that have occurred or that we've talked about in the past. I want to make sure everybody keeps grounded here. If you work with AI a lot, grounding of AI is really important. Um, so I'm going to ground us all right now with the difference between system one and system two bias or system one and system two thinking rather.

System one is really about bias. It's impulse. It's instinct. It's the speed of decision-making that keeps us alive. And that has a lot of strengths but has a lot of limitations. Uh, and that is um, well, they're endless. But then system two is really about stopping, slowing down your thinking, thinking forward, thinking backwards, thinking in parallel uh cycles and streams so that you can get a much more complete and uh better understanding.

You know, a simple example in the physical world is parallel parking in a tight space like you need to be physically, kinetically and experientially a part of that decision-making as you get into it. Um but yeah, learning new concepts, you know, doing anything of complex calculations. But what we believe where we're all going in this space is leveraging AI at the speed of system one.

So, you know, system two coming together with the speed of instinct of system one to deliver on a system 3 level, but you need that personal AI. So, we'll talk about that. All right. All right. And just a reminder, you know, all of these models that are still being built as of to this day are all being built on a microscopic element of how much information the human brain has um developed from.

So if you take all the human thought since beginning of time and you verbalize that, you've taken like a microscope or a microscopic element of that. If you take all the verbalized thoughts and you put it into what's being written down, that's a microscopic element of that down to digitized words down to use to train AI. So just remember that as we go as we go forward.

This is the big gap still to this day of all AI systems. Sure, it can look like it's doing deductive and inductive reasoning, but it is sort of doing it that way, but really it does not have the concept of abductive or experiential reasoning. And this is the biggest limitation of AI and all the generalized LLMs out there today. Okay. So throughout the year, you know, what we've done has been a little bit bold and made predictions, you know, from the beginning all the way through around 2025 and what will happen.

And I want to start that off with a video. So this video just came to us what a few weeks ago. You guys have probably saw it. It's around the Optimus uh robots. And um we have uh of course been almost anthropomorphizing a lot of this stuff wanting robots to be able to do our work and and uh but we don't really know what's going on in the background like how are they getting trained like is this a long-term solution experientially or is this more you know generalized transformer LLM trying to handle the telemetry coming in from the edge uh devices uh inside their purview and and trying to make sense of it.

But we I want to show you this and explain what's happening. So here we have uh on demo day, right? You have a dying uh Tesla Optimus. And so what really happened? A lot of us sort of try to figure out well what was going on here. It just sort of like died and and went backwards. But if you look closer, many people are speculating that the um driver of that, the human driver of that was wearing a VR headset so that it could control and it took it off.

And as soon as it took it off, then the machine um died. And so this is probably a great sort of like segue into this period of discussion. So with 2025, we made a lot of predictions. Um let's go through them. Some of them were right and dead on and some of them were just wrong. uh some of them are still somewhere in between and I think it's sort of cool to be able to pull out and call out all these.

So the ones that we got right um we said AI hype will continue to outpace reality. I think I mean I know that's sort of like a master the obvious prediction but we've seen this now and if you saw the MIT study in the summer around AI project adoption that 95% of them are failing was sort of more proof to that. So we got that generally right.

We said picks and shovels uh would be the real gold rush winners. I think we were pretty much dead on there. So, the chip makers, right? The data center developers, Alen, I mean, I know you track all this stuff a lot.

And then, I mean, that's the Yeah, that's the Yeah, that's those are the guys who are actually running the place right now.

They're running the place and they're driving a lot of this hype bubble. And it's probably going to be the first domino to start to fall is when they struggle or they stumble even though they might have, you know, stockpiles of cash. Um at some point that cash will start to get uh you know eaten into and you'll start to destabilize the whole industry and probably you know really uh realize that bubble.

So the third one was system one thinking will remain a core limitation. It has as we talked about and we I'm happy to go into it over and over again but it's easy to sort of display that it really is about the bias of the training data and you'll see the bias throughout um any LLM interaction. Um, AI fails at abductive reasoning. Uh, we've said it and it's pretty much held true throughout the year.

Uh, in fact, Fortune sort of called this out in an article as well. Moravec's paradox continues to persist. Um, that I think is without question. AI will not write 90% of the code. So, we heard the CEO of Anthropic state by the end of the year 90% of all code will be written by AI. Of course, that was a very self-serving uh comment. Alen, any thoughts?

Well, I mean that it's really this writes itself, right? Like it's not going to happen. Talk to any engineer. He might write 90% of the code, but not any code that you can actually use. You know,

right? So, Alen uses obviously AI throughout with all of our uh companies and all of our assets and all of our IP and um he can speak from direct experience. I, you know, I vibe code and you know I'm pretty dangerous at that but even that is very limited in its use. Uh it's great to sort of demo proof of concepts and you know it makes wireframes come alive kind of thing but other than that it's very challenging especially for full stack development and especially if it's like supposed to be scalable and secure and all this stuff it's forget it and certainly not 90% of all code it's absurd.

Transformer AI will end. Um, this is now beginning. You I'm going to go through this so I don't want to harp on it, but we've got a whole slide and discussion around it. Uh, AI poses a major threat to IP. This is starting to come to bear. In fact, the Anthropic lawsuit. Uh, I searched and all of my books are in there. So, all of my books are in the Anthropic domain.

So, Hacking Exposed, you know, one through seventh edition. Uh also all of the other um ancillary threads on that. So there is a lot of uh you know back and forth and I think their strategy the frontier model strategy is largely working though which is like they're just going to ask for forgiveness and then just sort of you know settle and pay them, pay the settlements out for the IP.

Um AI generated disinformation is a societal scale problem. Yep. The deep fakes obviously are starting to come fast and furious and the more personal we get and the more personal our AIs get the more I believe we could reduce that exposure and that risk dramatically but we have to get to a really stable place of personalized AI and then lastly agentic AI uh breaks the blood-brain barrier.

We did believe that this would occur and it actually started it has started to happen. If you are familiar with the Rabbit um a little device um Rabbit R1 the agent on there uh was built with hard-coded API keys. So rather than honor the authorization and authentication uh of the individual specifically, they put generalized keys in that allows for anybody that can reverse engineer uh the firmware and whatnot on that stack, the agent stack which includes ElevenLabs, Azure, SendGrid, Google Maps, all of their sort of uh API keys were disclosed in that.

And so we saw that happen. Now where we sort of missed the mark or you could say 50/50 are some of these elements. So we said in the beginning senior devs will be 10x using AI and juniors will spin their wheels. That hasn't totally occurred. I think um you know Alen can jump in but basically I've seen sure senior devs maybe double or triple and then juniors you know maybe 0.5 you know to, you know, uh 1.5 uh to 2x in terms of being able to at least just accelerate some of their development. But anything complex still remains a very challenging place for AI.

You still need to know how to code.

You still need to know how to code. You still need to take the code that AI generates and make sense of it, right? Exactly,

you know, uh transform it or whatnot. Make it more secure. All right. Efficient compute uh frontier is real. Um well, you know, it's generally out or the you know, the jury is still out. It's hotly debated. You know, some that argue that frontier has shifted to these new architectures like mixture of experts so system two reasoning models maybe starting are starting to help this a little bit model collapse will limit AI I actually think that this is starting to happen and it's happening um maybe not as demonstrably as it will in 2026 but there will be limits and we were starting to see this like with um ChatGPT with the 5.2 or the five and then the 5.2 to now you know there are limits like 5.2 still gets it wrong on simple math. Um it's because you just can't sort of duct tape and bailing wire this stuff um any further. You know, you really have to get to an experiential world model, which we'll talk about. Humanity's Last Exam, which we thought, you know, would really be the breakout final, you know, system two test, system two reasoning really.

Yeah, it's still there, but ARC AGI is sort of led the charge towards the end. Um, and then decline of human cognition. I'm starting to see this, but it's very early days. Um, and, uh, we'll see more of it, I think, in 2026. All right. So here let's go through the hype train in 2025 and call out some of the major elements we saw. So AI researchers negotiating these I mean nine figure you know plus um uh pay packages.

Now this was happening quite a bit in August, September, July. It sort of died down from all of that because of the blowback on it but it was absolutely getting ridiculous. Um some of the more grounding efforts I don't know if anybody saw this the Cerebral Valley AI Summit uh which occurred in November actually of 2025 just this just last month they actually predicted in the audience or they surveyed the audience asking if you could short a billion dollar valuation startup which one would it be and number one was Perplexity number two was OpenAI but then you had a lot afterwards Cursor, Figure, Mercor, Mistral um and even Nvidia and I think Palantir came into that uh that bunch.

So, a little bit of grounding on that I like. Um, let's see. So, there has been as you start to go through EY did a survey of uh their organizations asking the extent AI adoption has improved company performance and how so. And you could start to see that there is some usage and recognition of AI, but usage is still dramatically down.

We talk to customers all the time and usage of AI just as a general rule has gone dramatically down. So if you're surveying the users of AI, this is probably what they got here. Um, and you can see though employee satisfaction, revenue growth, and cost savings really aren't starting to materialize yet. Now, everybody's been trying to build the biggest data centers and, you know, trying to prepare for the future and all the hardware, all the hardware is going to go and the racks are going to go.

But this has really triggered even the the biggest beneficiaries of the hype to pull the reins a little bit, you know, from the Google CEO to Microsoft, you know, warning about these uh the bubble potential for what we're doing. So, we'll we have more data on that in a little bit. And, you know, there is it's not without its consequences.

You know, we have huge energy needs with all of these data centers, which inevitably energy is not cheap and energy is not free. There's always a consequence to it. So, we're starting to see the effect of these data centers and not just in the energy, but uh you've probably seen some of the 60 Minutes episodes and others about um uh the noise and the pollution that occurs from a lot of these data centers.

So, it's not a net zero-sum gain here. All right. This is the second priciest stock market in 155 years and AI is why. So I would keep uh an eye on it if you haven't already. You know, one of the big problems with it is obviously the circular revenue problem. And you probably have seen this, but it's, you know, there's secondary and tertiary and quaternary, whatever it's called, right?

The fourth version of it, which is they try to sort of almost launder the progress and the revenue. So you'll say, okay, OpenAI commits uh to buy so many Nvidia chips and then Nvidia invests the same amount in OpenAI stock. And so you get this quid pro quo. You know, in the 2000s, a lot of companies got in trouble from a GAAP accounting perspective for this in the quid pro quo era.

And now we're just seeing it all over again inside of AI. And I think one of the best ways to show this is with a little video. So here we go. And speaking of money, how about the 20 bucks you owe me?

Oh, yeah. Well, I only got 10, so here's 10. I owe you 10.

Thanks.

Hey, Moe, you owe me 20. Well, here's 10. I'll owe you 10. Uh-uh. You owe me 20. Here's 10 I owe you 10. Here's the 10 I owe you. Here's the 10 I owe you. Here's the 10 I owe you. Good, man. We're all even. I mean, it can't get better than that.

Yep.

When you get to Sto.

All right. So, is the Big Short 2.0 coming around the bend here? I think, you know, if you watch folks like Ray Dalio, Michael Burry, uh, Peter Thiel, many others, they are starting to pull out. They're starting to get and realize their gains, which is one of the earliest leading indicators of a potential bubble burst. So, keep an eye out on that.

You know, we are absolutely seeing sort of the poisoning of data now, right? So um if you look this is a great visual to sort of represent all the elements that we're starting to see here but like AI pages in Google top 10 it was 11% in May of 24 now it's almost 20% of all AI page of all pages in the top 20 for Google uh new pages with AI text uh has gone up dramatically AI generated news domains uh was 49 in 2023 now it's 1271 uh in May of 25.

The one that really gets me a little bit worked up um is this next one. So, PubMed, which is one of the gold standard sources of accurate um research uh around the medical um uh communities and healthcare has been now almost infested like a virus with AI. Uh and you can see that being tracked here. Now, you know, when you dive into the data, you can maybe say, well, I don't know if that really is a good representation of an AI generated output.

I think, you know, maybe some of the wording elements on the left hand side, you could probably argue, but on the right, these number of excess words, this is all AI. You're not going to get some like you know all of a sudden from 2023 to 2024 now all humans that are devising or building out these uh publications uh to release are just getting more loquacious it's just not going to happen this is definitely AI this is again abductive reasoning applied to this problem right and we're seeing it pretty clearly so the number of excess words uh per year is just it absolutely exploded all right this one is a fun one I— how Alen likes to share his thoughts on this one.

Yeah, I mean this just I mean this just keeps proving the fact that like you know you cannot rely on AI. You have to look over its shoulder. You can't I mean anything is possible non-deterministic like you know and then it can destroy your work like it's just I mean and this is this keeps happening non-stop. we

well this happened when right July of 25 but this just happened now with Google's um tool as well. So

yes. Yes. The Antigravity. That's right.

Antigravity. Exactly. So this vibe coding disaster um leveraging AI obviously all along the way actually started to delete entire databases. Okay. Same thing now with Antigravity who just came out. I mean was is now fairly well regarded I think in the space. just

Yeah. I mean they released you know as they released Gemini 3 they uh I think what they did is they um uh they leveraged some of the um Windsurf acquisitions that they did like you know uh and try to release the IDE because everybody is doing that right Cursor is quote unquote doing well uh but then it just wasn't nowhere you know I tried to test it and it's just it was like it was nowhere near ready to do anything and obviously some of these people who don't understand like you know what this all means and everything they try to use. And this is what happens.

Oh, by the way, Google does not allow their devs to use Antigravity internally, but they have released it onto the onto the world in order to, you know, in order to destroy your work. So, yeah.

Well, I don't even know where to begin on this one. N

If this is true, which you know, again, come from reputable sources, not AI generated, best we can tell. All of this is you know, live human experiential reporting, then this is a problem, right? I mean if Google can't even endorse their own IDE but this is another example of deleting entire drives, entire databases, entire contents.

So we kept hoping okay GPT-5 this is going to be it. GPT-5 is going to be AGI by the way. I don't know if you remember those claims uh Alen.

Oh yeah, sure.

Okay, sure. So five came five came out and was claimed going to be a PhD level expert in any area.

Well not not exactly. Um it met it was met in the community with a lot of disappointment um and a lot of uh you know it not being very helpful whatsoever. We are I believe this is the first signs of seeing that synthetic or the model collapse element. Now, you know, the big push in 2025 was to accommodate for models with all the synthetic data, which I think in certain cases can really help, but in other cases, I think you're just duct taping the crack in the dam.

Um, it's going to pop out. Um, and there have been sort of countless problems, right, with five itself and even with 5.2, which we'll talk about in a second. So, as in response to Gemini's release of 3.0, which was very solid. I mean, in terms of LMArena, I think it's still up there, right? Uh I don't know, it might have been just displaced.

He called it the code red moment for OpenAI. He, you know, he sent famous memo or notes out to all employees saying, "We've got to rally around. We got to get 5.2 out. This is going to solve everybody's problems, blah blah blah." But of course, somebody did the simple math of Okay, great. So 5.2, the smartest thing in the world. Uh, so do 5.9 minus 5.11. Um, and of course it says, "Nope, you're off. It's incorrect." So it again, it's just really challenged by some of the most simplest things. Um, but of course it can look brilliant in certain cases when you give it enough data and information. It'll look absolutely magical. All right, agent washing. You know, we've seen a lot of layoffs this past year.

Um, I'm not sure if we went through each and every single one of them. By the way, I actually believe either one of two things have happened. Number one, they leveraged the AI agent story or the using of AI as sort of a smoke screen to start to lay off uh folks just in general. Um maybe they probably were planning on doing it already um for sure.

And then for the others, they've actually started to rehire them back. And this is just three stories, but we I've tracked on dozens of these kinds of things. So, it's really going to be interesting. Even Gartner is sort of weighing in on this one, saying um agentic AI projects will be cancelled by end of 2027, 40%. And that may be true. I think it's going to be far greater than that.

Now, so you have that side of the picture of AI and then you have these amazing breakthroughs, right? And they're just starting to do this. I don't know if they're I you'd say that they're creating brand new inventions yet or brand new physics yet maybe, but you you can definitely say that you know the certain projects are absolutely blowing people away.

Um first one, AlphaGenome, right? solution

on

right

this

This was the AlphaEvolve 4x4 matrix multiplication problem that Alen's talking about um and this was dramatic and dramatically uh going to help AI and how we build deep learning models for sure. Um, but again, it uses sort of the same foundational mathematics. It just extended it. You know, I'm not seeing it create brand new forms of math or brand new forms of physics.

You know, it's still using what it's sort of been trained on as the basis and it's extending it. I think we are seeing that part and we are seeing great applications inside of healthcare. You know, I'm involved quite a bit in a number of healthcare organizations here in around Southern California and I see it all the time. Um, you know, using AI.

There's something called AI do or they pronounce it ID do something like that. And uh all decisions can be leveraged through this Aidoc uh platform. But here's a radiological um example of you know AI finding fractures or finding cancers in breast for example with breast cancer um you know years before any radiology tech is going to see it.

Now um that doesn't mean that it's 100% accurate. The human still has to take a look at it uh pull in the other abductive reasoning elements inside of that to make the final judgment. Um but in this case with St. Luke's University Health Network. It actually reduces um missed fractures by up to 30%. Which is I could have used this actually. I went to a hospital, what was it like three years ago?

I crashed on my bike and I went in and I'm like, "Oh, you know, my my ribs, I think I broke my ribs. You know, my wrist is sore, too, but I think broke my ribs." They took pictures, everything. They're like, "Nope, nothing's broke." And I go, "Well, I can't even use my arm, my hand. Like, are you sure? Are you sure? Go back, take a look at it.

And sure enough, there's a radial fracture and there's a fracture right here in my wrist that they had completely missed. But with because I pushed for it and I'm like, look, I'm not leaving until you look again. Like we got to check this out. Um even so the Google Project Green Light, a really interesting one where we are starting to see the application of AI inside of that for intelligent uh traffic light navigation, saving fuel and lowering emissions.

Um, we've also seen this for fire hazard detection. So, using AI cameras to spot wildfires early. You know, these are a lot of the positives that we're starting to see AI in. I have not braved the Waymo uh taxi yet. Have you, Alen?

No, I have not, but my brother uses it like religiously. He loves it and his family. Oh my god. He will not call anything else but Waymo. Yeah. Uh so our colleague Darren uh told a story of getting in one of these guys and then having it sort of overshot the destination and it was across the street where you had to cross four lanes of traffic uh you know quickly.

It was not fun. So there's more again more abductive reasoning limitations here. It does not learn from experience per se although you know the Waymo system might have a little bit of experiential learning. Um,

no, but it's really I think it's what this shows all these things what they show is that AI is incredibly good in pattern recognition, right? All these things have patterns, right? And we're not that great in doing those things. So, what it can do is it can it can point us in the right direction and cut down on a lot of time that you would have to get to those patterns.

And I think that that's where really a good, you know, a good thing is happening right now with AI.

You're right. Exactly. So what we are seeing GPT-5 you know hitting a wall in reasoning um but yeah yet you know we can solve these 40-year math problems and this really does um sort of go back to the fact that you need the human in the loop. You need the human expert in the loop to apply that experiential learning to apply the empathy to apply the critical thinking past the normal and to apply longer term memory.

I mean it's one of the biggest challenges of all the LLM today. There is literally no memory. I mean, yeah, you can summarize little things in chats and stuff like that and put it into a database, but there's no real memory. I mean, you have the context windows, which are super small and they don't seem to be getting any bigger, and then you have the out the input and the output windows, which just make it super limited.

So, all right, let's move on now to 2026 predictions. So, are we going to put our head into the sand or are we actually going to look at this and uh figure out, you know, where the problems are going to be and start to solve for them? Uh that'll be the big question. All right. So the AI problem. So there are two problems to this in my book.

There's the people problem and the AI problem. Let's go with the AI problem first. So what we have today is word prediction by and large um based on the foundational elements of the training. And that is not life experience, which is what pretty much all of these people are building brand new companies with, you know, multi-billion dollar pre-money valuations with multi-billion dollar checks going to them to build world models.

Now, world models are simply experientially trained models. Okay? And there are a number of different um techniques that people are researching, but I think it was really important. So, first Yann LeCun, who we've been talking about now for quite some time, was really way ahead of everybody in terms of calling this out and we very much uh subscribe uh and have since the beginning to the way Yann has been thinking about this.

Richard Sutton recently came out as well around the limitations of these large language models. Andrej and Ilya both came out in podcast recently talking about this as well that the limitations are really going to require us to go back to research to solve this problem. That's why that's why Ilya was talking about so much in this podcast and same thing with um uh Dr. Li, you know, she's been talking about this all year. It's just going to struggle. It's never going to get to that place of full AGI. I think we've all sort of seen this now. So what does the post transformer era in really 2026 we think look like? It's going to be small, it's going to be smart, and it's going to be hyper specialized and it's going to be personal.

And we're going to talk about the personal side in a second. So this includes everything from the architecture to the logic to the context and to the deployment. Um, architecture-wise, it needs to be continuous learning, nested learning. Like I don't care what you you can call it whatever you want. You can start to research on all the platforms that might solve for this.

Liquid neural nets is one of them. Um, learning on the fly and forgetting nothing. Again, the memory problem. So, we don't have this today. Okay. Logic, neurosymbolic uh AI, right? Maths not as a prediction but as you know as understanding logic engines and hallucination um reduction and then context. Um you know we need to have near infinite context that's cheap and zero latency that we don't have today uh inference costs I mean I was up in the valley last week and I heard from numerous investors um look the old days of SaaS you used to be like 80% margins now if you're over 20% margins uh no one wants to invest in you because you're not spending enough money on infrastructure and then finally deployment.

We are starting to see some really um interesting chips now that can be in the at the edge. So like think of a video camera with a chip that can actually learn what the difference is between a human walking and a and a dog walking by versus a brush uh waving in the wind. Um but learning on chip like on the device I think this is going to be this edge sort of learning might be a big part and blow up this year.

So finally all this is coming into what I believe to be this personal AI world. We need to be able to provide this hyperpersonalization throughout all of our experiences with AI or AI will not be used. I think that will be this is our prediction for 2026 certainly mine. So we um we are tracking on what we call EPOCH um which is the MIT Sloan framework now for where this human and machine come together and it starts with empathy and emotional intelligence or you can call it EQ.

Uh you can uh talk about presence and networking and connectedness, opinion, judgment, ethics, creativity, imagination, hope and vision. All of these have to come to bear inside of this personal AI layer that um us and others are starting to work on. So as AI gets smarter, it moves from this assistant era which is really pattern matching system one and you know really is labeler to agentic era which we believe we're trying to you know grow into now with is which is really system two chain of thought mixture of experts um and logic and reasoning based as a supervisor of workflows.

But ultimately we really believe we're going to start to get to this potential for system 3 which is at scale doing system one and two at scale with human as the orchestrator human as the uh final control gate of decision-making so that they become we become as humans strategic uh judges right and we become uh the ones that apply ethics and EQ and all of the rest of it to the output of AI agents and without the human in the loop uh it will fail.

So what I want to show you is okay so let's talk about personal AI and I think the belief is really this is going to blow up and when we do get hyperpersonalized AI what we really will get is coherence and cognition being able to actually understand each other but understand how we all work together to produce the final um the output of a team and it really does start with understanding self first.

So remember I talked about the AI problem. Well, this is the human problem. The human problem is the brain is is quite limited. I mean there's incredible capabilities the brain has, but there are some really stark limitations. And I'm going to play this video for you. And these are all celebrities that you recognize very carefully, very closely.

But if you look at that center X as the faces flash um around you and you stare at that X, what'll start to happen is you'll start to see those faces absolutely turn into monsters, some grotesque images. You can look left and right though and all of a sudden they come back normal, right? Look left and right now go back to the X. So the eyes become warped, the nose, the foreheads, the lips.

This is because your brain can only take on and process the information that is in its focus area. The rest of it on the periphery you make up in your head. The brain makes it up for you based on your biases, experiential biases. So this is the human problem. Okay? Without focusing on it, you are actually responding to and thinking that it's something totally different.

And this is a visual example, but you could do it auditorially. You can do it kinetically and experientially. It is a limitation of the brain. So if you know that you can move forward with it. All right. So what we need to have is a complete system around artificial individual intelligence and personal AI and this is what we've built at WethosAI.

So taking in those experiences and memories certainly as the system one uh those imprints um and the traits and biases that naturally form as part of the blueprint of a of a human mind and then layer on this temporal and contextual elements that are occurring all the time in real time to then be seen through the prism of the the dimensions that you present yourself to.

You know, there's a political self that you present. There's a financial self, there's an emotional self, personal, professional, spiritual, physical, psychological, all of these things. And understanding all those dimensions of self and building that into understanding you so that we know exactly how you think and how you behave and how you act and how you make decisions and then we can help guide those decisions to meet your ultimate objective.

So if you have an objective XYZ, hey I want to be a, you know, a CEO one day. Well, great. Well, how do we get you from where you are today to being a CEO? You specifically, not just in generalities. And I think this is what we're calling the system 3 thinking realization. All right. Now, we are going to switch over now to a special guest that um will hopefully demonstrate the value of this personal AI.

We call it XO here at Wethos. And let me do that first. Let me switch over. Hopefully everybody sees this. Now I'm going to engage. Now this is the WethosAI platform, but I'm going to show you a beta feature called XO Voice that will hopefully bring this all together.

Hello there. How can I help you today?

Hi there. What's your name? I am XO, your workplace coach. How can I assist you?

Well, given the specifics of you knowing me and what you know about me and how I process information and relate to the world and organize my life, explain quantum mechanics to me.

Quantum mechanics is the study of how things behave at the very smallest levels like atoms and even smaller particles. Instead of being predictable like bigger things, these tiny particles can be in multiple places or have different properties at the same time. It's all about probabilities and the idea that observing something can change its behavior.

It's a fascinating field that describes how the universe works at its most fundamental level. Does that give you a good overview?

That helps. Yes. Now, how would you have answered that question about quantum mechanics if you did not know me as personally as you do? If I didn't know you personally, I would probably explain quantum mechanics in a more general way, focusing on the core concepts without tailoring the explanation to your specific understanding. I would likely describe it as the branch of physics that deals with the behavior of matter and energy at the atomic and subatomic levels, highlighting the principles of quantization, wave particle duality, and uncertainty.

The explanation would be more general and less personalized, focusing on the scientific aspects without relating it to your specific way of understanding things. So, what specific behavioral traits and cognitive biases that you know about me was incorporated into your answer uh for me in a personalized way? Based on your WethosAI style, which emphasizes innovative thinking and a focus on big ideas, I tailored the explanation to connect quantum mechanics to those aspects.

I focused on the conceptual nature of probabilities and the mind-bending ideas of particles being in multiple places at once, which aligns with your interest in innovative concepts and exploring complex ideas. I aim to present the information in a way that resonates with your natural inclination towards big picture thinking.

Thank you. So this is what we call personal AI. So unlike generalized AI and its general answers, we respond with hyperpersonalization.

Yes, exactly. Personal AI focuses on adapting its interactions and responses to the individual user, taking into account their specific preferences, behaviors, and individual styles. This allows for a more personalized and effective interaction compared to generalized AI that provides more standard responses.

Okay. Now, I want you to take on a new persona of an AI expert. Do you think that we're in an AI bubble today? And what would be the leading indicators uh to watch to watch the bubble burst? As an AI expert, I can say that there's definitely a lot of excitement and investment in AI technology right now, which could be seen as reminiscent of a potential bubble.

Key indicators to watch for a possible downturn would include a slowdown in venture capital funding for AI startups, a decrease in the perceived value of AI-driven companies, and a potential shift in market sentiment regarding the long-term profitability of AI technologies. Additionally, increased regulatory scrutiny or significant setbacks in AI development could also herald a shift.

So, what alternatives to the transformer-based AI architectures will likely grow stronger in 2026? In 2026, we might see increased adoption of alternative AI architectures such as recurrent neural networks with more advanced mechanisms and graph neural networks which are well suited for complex data structures.

So what about speculation around 2026 in general for AI? What predictions uh do you have? Looking ahead to 2026, we can anticipate further advancements in AI capabilities, particularly in areas like natural language processing and computer vision, leading to more sophisticated applications. We might also see greater integration of AI into everyday technologies and more accessible AI tools for businesses and individuals.

However, challenges related to ethics, bias, and security will continue to be significant considerations as AI becomes more prevalent. The focus will likely be on responsible AI development and deployment.

Thanks so much. Uh, you've been very helpful. Appreciate it.

You're very welcome. I'm glad I could help. Feel free to reach out if you have any more questions in the future. Have a great day.

All right. So with this personal AI, what can you do, right? Not just get that personal content, that personal summarization, that personal engagement with AI in at every step, but now you can be so much more productive and efficient and effective, not just individually, but in a team. So this is what I want to show you. So

absolutely.

Oh, sorry. Hold on a sec. right,

no problem.

Okay, we'll shut her up. All right, so let's go into um where workplace misalignment really occurs and why XO is so dramatic uh and dramatically helpful. What we've done, Alen did, is plug in uh agents inside of Minecraft. If you have kids um or had kids, you'll know that Minecraft's a fun immersive experience game uh that you build, you source materials, and you build in uh and you share as a community.

So he plugged in two general AIs and he plugged in WethosAI AIs and asked it the same simple question. Build me a two-story house. So let's see what happens. So on the left is the general AI. So OpenAI and Gemini. As you can see they're talking but rather slowly back and forth and they get a little frustrated on the left. So one of the agents just goes and builds the foundation right there in bricks.

Then on the right hand side, you're seeing WethosAI AIs come together to discuss very quickly and very collaboratively the first floor of the house. Now, back over on the left, you see the second agent actually just built its own house over here, uh, sort of giving up on the first agent. And then you have the second agent there building now the walls of their house.

And they are attached and shared to a wall here, but they're two separate houses completely. They just gave up working on it together. You go back to the right with WethosAI and you can now see they're finishing up on the roof and putting in the windows and starting to get the house to be, you know, near perfect. It's not perfect. They miss a door on here, right?

But they've got windows and they've got all the other elements go back to the general. Now, here's the trick. If you see, I'm going to pause it right at the end here. Um, well, I was going to— hold on a second. You can see that what happens is look at the time it took for general AI versus WethosAI to come up with these. So number one, General AI took about 25 minutes and it produced three different houses.

Okay, completely uh other than shared wall completely disconnected and on the right with WethosAI it took less than six minutes and we got a I would say 90, 95% accurate house. This is what personal AI delivers into you and your teams as you work. All right. So, just in wrapping up and then we'll open up to questions um if there's any. But as we keep saying all along the way, AI is an incredible tool.

Just know how to wield it. Know the Zweihänder which is the two-sword uh or two-fist or two-hand. That's it. Yeah. Two-hand is what uh Zweihänder means in German. Um, 16th century German sword is six foot long, but you need to know how to wield this thing or you're going to cut yourself. Uh, there's no doubt about it. So, building that cognitive muscle to wield it with power is really what we're here for and what we're trying to help everybody understand.

So, thank you all um, humans and AI, I guess I should say, um, for participating. And if we have any questions, we'll take it now, but otherwise, we'll give you a little time back. Okay, let's take a look. Okay, Oracle's RPO just hit 523 billion, nearly 10x its annual revenue. Is this record backlog a signal of a generational shift architecture a massive capital expenditure bubble to burst.

I believe and Alen, I'd love your thoughts here too, but I think it's part of the problem. It's part of the bubble. I don't think there's some magical secret. I mean, I believe, here's my theory. Everybody's investing in all of this data center infrastructure that's going to happen. Uh, some will fall out eventually because of, you know, lack of debt payment if it takes too long to build and that kind of stuff.

But ultimately, all of those data centers are going to be built. Now they might not be housing the same hardware um in 5 years as they will today but those data centers and in large part the AI supply chain will still be robust because we'll be repurposing a the the learning and the research now to move away from transformer-based into more experiential learning systems.

And so I do believe that while it's a part of the bubble it's not the most critical element of the bubble. Um I do think that the data center has longevity. Yeah, I think that um one of the things that uh that everybody should uh should kind of listen to is the podcast that Andrej and Ilya did and they're very very informative especially if you know what you know uh what they're talking about as far as technologically um you know Ilya was talking about how like these guys are trying to solve the transformer model limitations with just adding more compute and that's what everybody starts to throw in more they're just thinking that if we throw more compute these models going to become much bigger and we're going to get closer to the AGI and the guys who know who actually invented some of this stuff which is you know Ilya in no small part, Ilya, Andrej and rest of these guys have basically like started thinking about okay well that is not what's going to do it right and I think that a lot of companies actually realizing this right now and you know and I think while they're going to continue building and we do need compute for AI um there's going to be a reckoning for sure because the because the underlying technology for AI which is transformer-based models and the deep neural networks.

It is not, uh, it is not a technology that is ultimately going to bring us to AGI. It's not, it's— we're going back to research. Uh we talked about this for a while. I've been doing a lot of research for the last couple years on um on test time uh learning right in memory understanding things as you go along like learning from people as you talk to them like you know stuff like AI cannot do any of these things.

While the context windows are becoming bigger, you know, while they've kind of stopped at like 200k and 1 million and stuff like even at in those context windows when you hit around 200,000 to 250,000, the AI start getting confused and uh and that's a big problem we have here. It does not learn on the fly.

I can vouch for that. I use it every day in very large cont as large you can get context windows and it constantly gets confused. I have to constantly ground it to accuracy and it's just the nature of the transformer. All right. Uh next we have does offer a data-driven path to improving rework rates or is it simply vibe management rebranded for the AI industry?

I love that. Okay. Um no, our metrics are very real in real time in and around how people work together better. So we have something called alignment score. We have decision velocity. We have uh friction indexes. All of these things are measured in real time so that you can see the value dominantly. We also have AI usage rates which is what personal AI is really starting to increase on is people want to use it because it's hyperpersonalized.

It's not general. So we are really starting to see that inside of WethosAI itself. Um next question. As a victim of an AI layoff, what can I do to help AI-proof my work experience uh beyond just adopting AI for productivity? I think that's a great question. I mean, if you've been a part of an AI layoff, first of all, don't get bummed out. Um, they might be calling you again to hire you or rework or maybe you don't want to go back there and you just want to move on, which I think is great, too.

I think in general understand um yes understand AI a bit better and how to leverage it to accelerate your productivity and your output both in quality and in quantity but quality more so you know you need to stand out from everyone else and AI can allow you to do that but it has to be hyperpersonalized and it can't be so foreign to you or orthogonal to you that it doesn't come across as you and you don't believe it you don't buy it you— so you can't sell it. So that's where that personal layer really is so important um and critical.

And then we got how important is the accuracy of the prompt fed into an LLM? I accuracy of the prompt. I mean you need to Steve, thanks for that one. You need to certainly ground it what we call grounding it with truth. If you ground it with falsehood, it can get it confused quite quickly. Um, and you know, I get reminded of an example of me working with Gemini 2.5 and oh actually no, I think it switched to 3.0 when this happened and asking advice on you know complex health issue and it produced the answer and at the end it said make sure you're inconsistent with this protocol and I looked I'm like incons you want me to be inconsistent with the protocol like that experientially in my life that doesn't sound like the right thing to be recommending.

So I go, "Wait a second. Do you mean inconsistent or do you mean consistent with the protocol? Be consistent. Oh yes, I'm so sorry. You're absolutely right." Now, if I hadn't stopped it right there and I had just followed it all go, oh yeah, okay, I need to be inconsistent that I need to do it tomorrow but maybe not for three more days and then I need to do it in a week and I would have had a very different health outcome.

And this is again why yes accuracy of the prompt so to speak which is the human is very important but even if it is perfectly accurate you're going to get confusion especially as these uh context windows continue to get filled up and then the window starts to move past.

I'm just going to I'm going to I guess uh um just expand a little bit on that. When it comes to accuracy, obviously, you know, when there's a term says garbage in, garbage out. If you're not accurate, you're going to get something that's not accurate, right? Like, you know, however, what's most important for everybody out there who want to work with AI is that you are very um uh clear about what you want AI to do for you.

So when you're actually like when you're doing the context engineering and prompt engineering and everything, if you the best results you're going to get out of AI is you explain it to AI exactly what you wanted to do because if you become too vague you're going to get vague answers and you and uh and things are going to eventually going to fail because it's going to try to guess what you're trying to do right it's going to try to guess exactly what you're trying to do and when it starts guessing what you want to do is when you get hallucinations so the more specific you are with AI what you want it to do the better results you're going to get.

All right, final question then we got to end it. Which AI company would be Shemp in the Three Stooges quid pro quo scenario?

I think that's the perfect question to end on. Uh and we'll leave it up to you guys to decide. I mean, they certainly applied Shemp. I can't remember which one was Shemp in that video, but I you know, you could certainly um uh you could certainly pick your poison on that one. All right, everybody. Thank you so much for participating.

Thank you, Alen.

Thanks, Sarah. Thank everybody. Uh, we'll look for you in 2026. Have a great holidays and uh, enjoy them. Hopefully get some downtime. All right, take care. Bye.

Happy holidays and happy new year.

All right. Cheers, man. Bye. Bye.