S5 E4 - Reuven Cohen – AI, Automation, and the Future of Human Work

In this episode, I have a fascinating conversation with Reuven Cohen, someone who I believe is one of the most important voices in AI today.  

Reuven recounts his journey in technology, from being an early advocate of cloud computing to now working at the cutting edge of AI and reasoning models. He shares insights into how AI is shifting the nature of work, particularly in fields like software development, business operations, and decision-making. He describes AI as "cloud computing 2.0, but with intelligence," emphasizing its role in cognitive offloading—augmenting human capability rather than merely automating tasks.

A key theme of the discussion is AI’s impact on productivity and workforce structure. Reuven shares staggering personal metrics—writing nearly 10 million lines of code in a year, something that would take a traditional developer thousands of lifetimes. He argues that AI is not replacing jobs outright, but fundamentally changing who remains valuable in an organization. He suggests that companies must decide whether to empower their top 10% to become exponentially more productive or replace the bottom 90% with AI-driven automation.

The conversation also dives into reasoning models versus instruct models, discussing when to use each in business applications. Reuven explains neurosymbolic AI, a new frontier where AI models don't just process natural language but interact with the world using symbolic logic and mathematics. He believes this approach will be essential for future breakthroughs in AI comprehension and decision-making.

As the episode progresses, John and Reuven reflect on the geopolitical landscape of AI, noting that China has become a dominant force in AI development. They discuss DeepSeek, the Chinese-developed reasoning model, and how it has disrupted traditional players like OpenAI and Google.

To wrap up, Reuven shares his latest projects, including an AI-driven truth detection system, which sparked ethical debates about transparency, privacy, and misinformation. He envisions a future where AI is not just an assistant but an autonomous force that reshapes industries, economies, and even the nature of work itself.

Transcript

John Willis: [00:00:00] I have another great podcast today. This is an old time friend. You know, just as we, I think we've said this before somewhere, but. On a plane ride, sitting next to him, Reuven Cohen showed me, or told me about the cloud. He was the first one to explain what was going on with Amazon. And very few people had really heard about like, other than just a lot of noise.

But you were like, no dude, you gotta try this, this is this, and let me show you what the, you know, so. So you introduced me to cloud, we've kept the relationship over the years, And I, you know, I'll get you to introduce yourself here in a second, but like you know, right now, I think not just because I'm buttering up because he's on the show, but I think some of the most important voices in AI right now is this gentleman right now, because I, you know, I don't say something like that lightly.

 But he is completely, I started out arguing with him about what AI was going to do. I lost that battle and Reuven, you [00:01:00] want to go ahead and introduce yourself? 

Reuven: Well, Sid, it's never, it's never a battle. And I think we've, we met 19 years ago. I think it was, you know, it, it was like pre launch of AWS when I was kind of like being the, the, the early cloud advocate or whatever you called it back in those days.

And yeah, it was. We were in, I think, Austin at some kind of tech conference. We, you know, I was, I was trying to push the concepts of infrastructure as a service and everyone basically told me it was stupid, never going to happen, which was, it's, it's a kind of repeating theme in, in my life. It, when I, when I know I'm onto something hot, everyone either tells me I'm crazy or it's illegal.

You know, or it's never going to happen and 

John Willis: it's it's weird quote from Jonathan Swift. It's you'll know true genius when a confederacy of dunces all ban against them. So 

Reuven: that's a good that's a really good quote. And I, I think, you know, it's short, long, you know, 30 years, right? I've been built. I've been, I've been [00:02:00] building websites or.

Online, you know, back when I was, since I was a kid, you know, so since the mid eighties, so, you know, I'm, I'm what a, a geeky kid in the eighties grows up to look like as an adult, you know, I've, I've been lucky to be sort of at the forefront of a variety of different sort of emerging technologies over the last 30 something years and, you know, for me, it started with, you know, AOL in 1990, where I was lucky enough to be an early beta tester.

And that was kind of the first. Step along the way. And that led to a whole variety of things. And, you know, freelance web designer when I was in high school, traveling the world, everyone else working at McDonald's and I'm traveling up to London or New York and Los Angeles to build, you know what they call it, cyber designer.

So I've had a knack for being early in things. And for me, AI, although it's, it's 70 years in the making, it sort of had an inflection point. What 2022 when ChatGPT came out and I could see the writing on the wall. This is this [00:03:00] is cloud computing 2. 0 but with intelligence. I'm and I kind of jumped on.

John Willis: Yeah, no, no way. You know, that's kind of fun. Little fact in the middle of there somewhere. You were like one of the top Drupal consulting companies for a little while. I don't know if you even want that on your resume. Oh, 

Reuven: yeah, yeah, no dress and crew. Great, smart, capable team. You know, the tech tech moves things evolved.

Monolithic CMS is maybe aren't so popular anymore. But back back when in the mid early 2000s, you know that The idea of a content management system was fairly revolutionary, and the, the ability to have it run on a lamps stack was even crazier when every other option available was proprietary and, and ridiculous, expensive.

Suddenly you had the systems and a lot of the work that I ended up doing in, in that became known as cloud computing was directly related to the work I did in content management and in the ability to scale these sort of lamp style stacks, right? It turned out. [00:04:00] Separating them and putting them on a hypervisor was the sort of knack for scaling these these systems.

And again, this was 20 years ago, and no one thought to independently scale your your app. From your data from your content, right? And, you know, showing up and saying, well, monitor it as a utility scale it independently and you can scale forever was now it sounds like, well, obviously, but back then it wasn't so obvious 

John Willis: you had me sold on that plane ride.

But so now you know, and I think, you know, there's I think at some point, you know, somebody's gonna have to write a good biography about you. But, but, yeah. But what I really wanted to get into now is you're so cutting edge now on the on what's going on. You know, and, and I periodically sit in, I don't know why I don't do it every Thursday, but Thursday and Friday, but I sit in on your AI hackers and, and it is pretty phenomenal that you walk through stuff that like, it just blows my mind.

But the [00:05:00] thing I, I, I'm thinking a lot about is like, how does, how does executives in sort of a bank or a, you know, the people that sort of grew up. More like me somewhat like you but didn't really jump in as hard as you have and you know, I've jumped in a little bit How do we make sense of you know, we we've got with GPT 4 and then we had GPT 4 Omni And then we had you know The o1 and then we had you know There's you know, the o3 and then we had the deep seek and I don't want to over rotate on deep seek but but I there's this thing right now about reasoning and And I think it's really hard to figure out, particularly economically, as an experimenter, as a problem solver, as a sort of, like, how do I sort of even make sense of the kind of things I might want to do, particularly like in a bank or an insurance, normal sort of enterprises, not so much a very specific AI solution.

I know you work with a lot of clients who are very specific, like they want to do medical [00:06:00] care or something like their health care, but Yeah, can you help me and us figure out, like, where do I start in thinking about these? When do I use a reasoning model? When, when does it suffice to just use? Just a regular LLM.

And does my question even make sense? 

Reuven: Yeah, it makes a lot of sense. I think, first of all, you have to start with, you know, what is AI actually doing to you as an individual? And before you even apply it to your company, your subordinates, you know, either augmenting or replacing your employees, you know, the basis comes down to this concept of a kind of cognitive offload.

What AI is doing is allowing us to either, by crutch, Use information that doesn't exist within our brain, and it is readily accessible, and that can be great. It can empower you to do and understand in ways that were never possible, but it can also diminish in the sense that it makes it can make you lazy and make you not even [00:07:00] wonder or reason yourself about why or how that information, you know, what the meaning of information is.

So, you know, there's two sides to every coin. So when, when you start, when you start thinking about this technology, and thinking is the key part of this, you have to, you know, you have to understand the sort of nuance of the people in which they're consuming that information. And, you know, not everyone consumes or learns or interacts with information the same way.

Some portion of the population is always going to be lazy. And they're going to do the least amount, they're going to put the least amount of effort in and you're never going to change the sort of psychology of those types of people. Those are the people, let's be honest, that are going to be replaced by AI.

There's nothing we can do about that. They're gonna have to, they're gonna have to exist in a world where, you know, some percentage of intellectual capital is now done by A. I. And the people who don't want to apply their own intellectual capital to solve problems alongside [00:08:00] of it aren't particularly needed.

Now, when you look at the rest of the population, which I'm hoping is the majority. Now we look at people who are capable of doing things that weren't possible before. We're talking about movies like Limitless, where you suddenly have the, the understanding of infinite knowledge at your disposal. What, what does a world where infinite knowledge, you know, and abundant knowledge at practically no cost, what does that mean to a business?

And those are the fundamental questions. The CIOs are starting to grasp. Am I looking at this as a way to take my top 10 percent and make them 1000 times more productive? Or am I looking at a way to take my bottom 90 percent and just completely replace them? Is it a mixture of the two? So those are both Those are cultural related, you know components in a business You don't want I don't think anyone wants to work in a business with no people But it's also directly relating to the bottom line if I can get 10 of my [00:09:00] workforce to be a thousand times more productive That is a meaningful impact to whatever it is that i'm doing as a business 

John Willis: You believe you can get the 90?

I mean, let's say that there's you know I will make some form of artificial bell curve and there's a percentage that that like you said is probably you know Not tall enough for the ride when in this new ride But so let's say there's 10 on the end and the 10 on the front and there's 80 in the middle you think that you know that it well, I mean if you never knowing everything, you know now would you for you know, and again, I there's a lot of variables for this question and I'm not giving you but Would you focus on the 80 percent and get productivity out at 80 percent assuming humans do have this innate talent of creativity?

Or would you focus on the 10%? 

Reuven: I think I think the the honest answer is right now. We're in a transformative phase with that. What that means is trying to to maximize the performance of [00:10:00] the human quotient within your business that you have today. And a large portion of what a CIO's job is to sort of maximize it.

That capability right through information technology, obviously, but that's essentially what what a CIO is doing. They're putting the information required to to facilitate the business of the business, whatever that may be in a more efficient and effective means now. In the short term, that likely means rolling out, you know, chatbots.

Nobody wants to use particularly a chatbot, but rolling out chatbots, understanding how those systems are consumed, how they can be, you know, hyper personalized and optimized for particular roles or groups within a business. But what that also is likely looking at is the persona of the role itself. What percentage of the job that those people are doing?

Do they even need to do? And the more CIO can understand, like, what's the value proposition of a given role? And what are the things that a person excels at? Where [00:11:00] some kind of a agent style approach would do a better job is going to be a big part of the sort of calculus that they're going to look at both in terms of cost.

I'm performance, efficiency and everything else that goes into maximizing, you know, the human capital in a business. So short term, it's augmentation, right? How do I augment my people long term? Something a little scarier. 

John Willis: It's pure optimization at that point. I guess once you figure out, you know, where you can, like you said, augment to the value.

Reuven: Yeah, someone in one of my groups did an analysis of the last year of my of my output, and I did a close to 10 million lines of code. So to put that in perspective, and obviously, I didn't look at every line of that 10 million lines of code. It's impossible for a single individual to put out 10 million lines of code, give or take at a couple lines of code a day, which is your typical developer that would take 33, 000 years.[00:12:00] 

You know, many, many, many lifetimes, and so it's just not feasible. But what that speaks to is I'm something like 10, 000, and that's probably an underestimation. 10, 000 times not percent times more productive than your average developer, and that is increasing at an exponential rate that out of those 10 million lines of code.

More than half of that's been produced since September. So in the last five months, so cool. Am I an outlier or am I, or, or am I a leading indicator? 

John Willis: You're a leading indicator, no doubt. But so so I, I do wanna get back to the reasoning model decision, but now this is, I love sort of just coming up with questions basically.

What do you say to people that say, Ooh, you know, but is that code gonna be good? You know? Yeah. Like that. I, I don't buy that argument, but I never have a strong argument pushing, or, I mean, I usually do come back and say, how many of your professional coders that a, you know, 20,000 Java code of bank [00:13:00] are really good anyway, but like, you know, the arguments that you get from people, like, Ooh, yeah, but does it write good code?

Does it write bad code? You know, like, 

Reuven: good, good. Good code, bad code, it's all subjective at the end of the day. And let's be honest, right now, on a first pass basis, a coding bot, you know, run on Sonnet or one of the reasoning models were roughly around 95 percent effective on the first run. So when you look at these models, what you need to look at are two metrics.

One is the pass run. So, you know, we're the state of the arts, give or take 95 to 100%. But let's just say it's 95 percent for the case of comparing against a human. Humans aren't, the first code 95 percent effective. If you've ever developed anything, it's an iterative process. You try, it fails, you try some more, it fails differently, you try some more, it fails in different ways than you expected, and eventually you figure it out and it works.

That is how you program. [00:14:00] Now, there's this expectation that somehow or other AI is going to just be, you know. Pixel perfect the first time out and that's not the way these systems work. They're iterative in much the same way a person is iterative in the development and software is essentially waterfall.

No matter how agile you want your infrastructure to be, most software requires a sequential series of steps to be produced. And when you look at sort of. Agentic style development. It's it's there's two approaches. One is I'm going to throw up the entire code base all at once. And then they're like, Oh, it doesn't work.

You just have put 100, 000 lines and you expected all the work in one pass. The other is I'm going to produce 100, 000 lines over the next several hours, iteratively testing each function and capability until it's perfect. That is where the power is. I can go, I can go to bed at night, have my agent building me things, get up in the morning, have a hundred thousand lines of [00:15:00] fully functional application code that has been battle tested against, you know, zero trust and other known methodologies and know that it's compliant.

Now, do I know every line in that code? No, I don't. And that's, that's the problem. 

John Willis: Yeah. And you know, do we need to know it? And the truth of the matter is most large code bases are not known by really any one person. No, 

Reuven: there's no, there's after about after maybe two or three thousand lines of code, it's really hard for one person to understand all the interconnected relationships of the functions and the types and all the things that make up an application.

Now, you can hypothetically get a, get a feel for if you work on it long enough and much the same way a taxi driver knows the way around London, but it, it, it's, it takes time, but you can do a lot more as an individual now. 

John Willis: Right, right. So back to the the reasoning. So, again, we've got, we've got just a lot of choices particularly, like, I, I think, you [00:16:00] know, you know, I think maybe the, not maybe the, the, the flat out truth is CIOs are going to have to just, you know, get their hands dirty.

And so let's say that a CIO is going to get their hands dirty and they're going to try out some models. But how do they even start? Do they just start off with, you know, so O3, can we, can we At least create a framework of guidance of like, what kind of things I would, you know, from your advice, I would say, you know, I, that could always be done like a document interpretation or something like that versus something that needs intense reasoning.

Reuven: So there's 3 primary buckets when it comes to AI. Right now in 2025. There are, you know, application specific neural networks. These are models that you're explicitly building to solve particular problems. They don't necessarily have to be language orientated. And to be honest, the best models that I'm that I'm creating right [00:17:00] now.

Use things like symbolic reasoning and neurosymbolic structures, things that represent the information, not through natural language, but but through symbolism and and that typically rely on advanced forms of mathematics and and abstract algebra and other structures. Now, those that's where the cutting edges in terms of technology.

The other is a hybrid approach that that uses language as the primary interface. Of that information. And that's where you start looking at LLMs, large language models. And in the LLM space, you're essentially breaking that down into two predominant buckets. You've got instruct models, which explicitly instruct the model on what and how to act, you know, tell me what to do and go do that.

Or, or inversely, I'm going to tell you what to do. You go do that. But I'm giving you explicit instructions on what those things need to be and what the expectation is on that output. And so let's call that. [00:18:00] You know, more of a, you know, an explicit rule set a stepwise sort of approach. The other is reasoning.

A reasoning is declarative. A reasoning model basically says, okay, I've got a problem and and I need a solution to that problem. I can give you some general guidance in terms of what the problem is and what I'm doing and how I might want it solved. And here's the expectation of the output. I need a A file or I need some code or I need a report, a research report, but to get from a to Z, all those steps in between.

I'm gonna let you figure that out. I'm gonna let you determine what the optimal steps to solve that problem from this without me explicitly having to tell you how to solve the problem. That's where the power of a reasoning model comes into play because it can do it. Over a sort of time test their test time sort of structure that meaning that the best results don't happen in 10 seconds.

They happen in 10 minutes or 10 hours. And the longer the I has to ponder the [00:19:00] concept, the better the output will be. Yeah, so you think of the problems that you're gonna solve different problems, different solutions, different models. Hopefully, that's helpful. 

John Willis: Yeah, yeah, no, it's very helpful. You, you kind of went over this the other day. 

And I guess one of the things that is important to know is when do you sort of, I guess, switch between the two? Because there's certain things like, you were given a demo the other day where you, you you used like a reasoning model. It sounded like, well, you actually, it sounds like you were using kind of an instruct model to give me all the things I need.

And then I want to turn that over to some type of reasoning model to say, okay, here's the sort of declarative way of what I want. Go ahead. And, you know, with a lot of sort of information provided, go ahead and make this thing happen. 

Reuven: Yeah, so there's I use instruct models frequently and and the most.

Popular one right now, at least for coding, would be Sonnet 3. 5 from Anthropic or Clod. That is an instruct model. That means it's really good at being instructed on how to do [00:20:00] particular things, and in particular, coding. Now, the, the other approach, again, as I point out, is reasoning. Now, when it comes to reasoning, there are multiple ways you can sort of deploy.

Or I like to say, employ reasoning within your structures. One is you can have a series of recursive steps, chain of thought, tree of thought, other sort of hierarchical thought process that define the idea of logic, reasoning and comprehension as multiple concurrent or sequential steps. And the reason why it's concurrent or sequential is important is whether those things happen one after another.

Or whether they're happening all at once for all the different problems you have to solve and then determining what the optimal solution to that problem is. So when you look at models like deep seek, which is based on when you know, Q. W. E. E. N. or you look at oh, three and oh, one models. They [00:21:00] have an internal.

Reasoning system that gates that reasoning as a sort of, neural network execution process. So think of, you know, abstracting that sort of those, those reasoning steps into a kind of gating mechanism that exists within the language model itself. That's great, but it's not very malleable or guideable.

Meaning that it kind of takes a life of its own. And in order to really get it hyper targeted, you either have to fine tune the model or or given it really, really explicit sort of guidance. Otherwise, it'll kind of go do its own thing. So when you use an instruct model, you can define that guidance independently of the internal gating or neural network itself.

It's slower. It's more expensive, but it's it's a lot more malleable and how and how it can be guided, right? So I can switch between smart models and bad models, reasoning models symbolic, you know, reasoning [00:22:00] models depending on the problem that I'm looking to solve. That's where an agent. Comes in the agent is capable of acting and and requesting things like tools or APIs or, you know, thought processes and other ways to go about doing things.

So, the, the, the important part of this is actually having an agentic flow or pipeline that understands the limits of its capabilities, but it's able to creatively explore where and what. Can be chosen independently of an explicit reasoning model itself, which is generally limited in its ability to interact with anything outside of what it was trained on.

John Willis: Do you have sort of a go to agentic or is that just something that you flow yourself? I do, 

Reuven: depending on what I'm doing. I, I'm a, I've got a, I've got a bit of a bot army so I can build any, whatever I need when I, whatever I need. But that said, it's, it's always good to abstract those things. So when I [00:23:00] have a problem that I need to solve, like, for example, before I hopped on the call today, I was working on drug trial.

Software to re sort of reimagine what, you know, the analysis of hundreds of thousands of pages within various drug trials. And that's a very complex problem. And one of the problems I have, I have with that type of work is fidelity and that with that, a large portion of what being an A. I. Whatever it is that I am a consultant is getting is sort of that evaluation of fine tuning back forth testing that it works.

Checking the expectation is correct. And up until recently, the job of an AI guy was just a lot of manual back and forth of data and say, okay, the numbers are lining up correctly. Now, all I need to do is set an agent and say, run that test. Concurrently 10, 000 times until you get, you know, consistency of 99.

9 percent and then send me a message when you're done. Okay. And then I'm [00:24:00] gonna go have dinner. And then a few hours later, I get a ping in my Slack channel from my bot. I've done it. I ran 10, 000, you know, runs and number 8, 992 was the one. If, if, if I had done that 8, 992 times. Well, it would take me 10 years, maybe longer.

John Willis: So, but I mean, I, I know that a lot of times when people ask you just kind of what tools you use, you usually say, go look at my blogs. But is there a particular agentic tool that you like? 

Reuven: Yeah, I, I like Lang chain. They've come a long way in the last year. You know, ask me a year ago, I probably would have been a little bit harder on them.

I think they've, their updates to Lang graph in particular were really good. This sort of graph structure. I like what the guys at crew AI are doing. They've abstracted a lot of the complexities of Lang chain. Lang graph is based on link chain using yaml, and so the yaml file means that all you really need to do to create an agent once you've got the sort of, you know, [00:25:00] installation in place is basically define it using that the yaml mark up, mark down, whatever you whatever that's referred to, and you're good to go.

Right? And it's easy and quick. The other thing that I do a lot of because I do a lot of autonomous development is I'll deploy in a serverless environment. I getting going fast. I like things like fly dot IO railway. That allows me to deploy these agents for a limited amount of time. Go build me an app when you're done.

Commit it to my GitHub and send me a message in Slack sort of thing. I don't need to, I don't need these things to run all the time. I just need them to run long enough for me to get whatever ideas I need done, done. 

John Willis: The confirmation or the sort of the, 

Reuven: yeah, 

John Willis: what I definitely want to get into the neurosymbolic stuff because I think that's fascinating.

About a year ago I was at a conference where I met a couple of doctors and they were all in on this stuff and I did some podcasts, but the sort of the one last part about the reasoning, I think at one point [00:26:00] you. And one of your sort of blogs or something like when, when 03 came out, you were talking about how, you know, like something like that might actually cost you 1, 500 or, you know, give or take to do it.

And in some cases, it's absolutely worth it. So still, I'm wondering, like, how do I, I sort of like in some of your demos, you start off with like 03 mini, you use sort of the deep research, and then you sort of work your way through. Different instruct models and all, but like, how do I think about, like, you know, I've got this problem, where would I start?

I probably wouldn't start with like the 03 Pro or whatever, right? You know, and what would be the most economic way to sort of figure out where I should start landing in the models I use? 

Reuven: So, one of the big areas that I start with, and let me, it's okay, I'll share my screen quick and give you a little. Oh, we turn on screen sharing and I'll show you an example of something I literally built today and [00:27:00] and the speed and complexity and and what you can do.

And this is this is like as both complicated a concept as you can possibly get and all right

so, so one of one of the, you know, I do a lot of you know, sort of training a model from scratch. So rather than depending on, you know, DeepSeek or some other language model that already exists, I have a preference for just building my own just because I can.

 And so I can, I can move beyond, as I said earlier, language to other constructs, like, like symbolism and, and neural sort of structures that mimic the way the human mind works. And so a lot of what I'm doing is, is, and is. Basically influenced by sort of trying to reverse engineer how we or how I believe we think.

So keep, keep that in mind. So what? So one of the so there was a some research that came out from I think it was deep mind that looked at sort of the the [00:28:00] ability to sort of infuse mathematical structures around. I think it was called like alpha folding and alpha geometry. 

John Willis: Yeah, that's the one that piqued my interest about.

Yeah. 

Reuven: So that came out. That came out last week. So here's what I did. The first thing I did is I, I went into, you know, chat GPT and, and o3. And I said, tell me about the, the alpha geometry latest model, everything I need to know about it. And then it said, here, here's the technical details. Then I switched to, the actual model for deep research and said, Okay, create me a thesis. And this thesis should be embeddable as a Google collab. So and make it PhD level. And I also want the acronym to be dream, right? Because because I want it to be a kind of stream like system that gives it the ability to visualize and take in multisensory information.

So in this case, it came up with a Okay. Creative use of [00:29:00] dream, which actually fits exactly what I'm doing. It's embodied. It gave me this idea of and I guided, obviously I'm going back and forth with, with the, with chat GPT. It uses multimodal reasoning, meaning it uses not only texts, but it uses video, audio, you know sense of smell and taste and other components.

And then, and then it's coming up with sort of theoretical foundation. I want to build something that's never been built before. Maybe that's my ego talking, but. I want to I want to push the boundaries of what's possible. So this is where the theoretical structure comes into place and saying guided symbolic problems.

G. G. S. P. O. which is the basis of the deep seek model and how they do sort of recursive learning embodied integration, recursive self optimization dreaming. It can actually dream and create its own reality. You 

John Willis: didn't ask for the G. S. P. O. It just came up. I did. I 

Reuven: asked for all these things. Yeah. So I'm giving a guidance.

I wanted GSPO. I knew that that [00:30:00] was a structure. I wanted it to be in the form of a Google collab. So other people could use it. So I'm not, I'm not blindly telling it what to do. I'm giving it guidance and the things I want. And this dream was something I actually told that I wanted as well. I wanted a system that mimicked the ability for your To take the information you're you're gathering what I'm telling you right now and then you and then store that information from your short term memory in the front of your brain to the long term memory in the back, but I want to do that.

I want the AI to be able to do that. And I want the AI to not just be able to do that. I want to be able to do that in a way that includes both visual and auditable and other tactile information. I wanted to create a virtual reality that it uses to train itself autonomously. It's crazy, right? It's pretty crazy.

So. Cool. I've defined a pretty, pretty crazy model. Then I use this chameleon seven B model, which allows me to generate and create sort of many brains. If you will. I don't want a giant model. I don't have the budget for that. I want a small group of [00:31:00] concept and then it goes through and it does all this stuff all implementation plans and and recursive dreaming systems and everything you'd ever want.

And then in this case, we can go down here. Sure. And at the very end, after all this stuff said and done, we can go in here and define the actual function. So this is is a something I can actually run. If I take this code right here from my URL, go into collab.

Paste that in there. And, and what this allowed me to run the concept. That's crazy. I built, I built an entire new breed of That's crazy. AI model. From scratch for a freaking LinkedIn post in less than an hour.

And now I can go in here. I I, I'll, I'll connect to my model. Boom. No, actually I, I should probably connect it to a GPU instance, so [00:32:00] I know I, I, I gotta connect without GPU 'cause I have one. But anyway, the, the moral story is it, I can install it, I can set up the, the tokenize, you know, this is using pytorch

it's, you can, I can see my, my training. I can. Multimodal vision you know, visualization. This, I don't think has ever been done. 

John Willis: Guided symbolic 

Reuven: optimization, that's, yeah, it's crazy. 

John Willis: Is I could, like, if I tried to do this, I'd get lost on a couple of those sections, but then I would just go in and say, explain this code to me.

And, you know, like, yeah, yeah, and that's another sort of amazing thing that I love. Like, if there's something somebody does, I can literally just ask whatever model and say, can you explain it? And if I don't understand what they explained to me, I'm like, Give me a little more detail. Yeah, 

Reuven: it explains to me like I'm a 10 year old.

There's nothing wrong. And that's the key to back to the cognitive offloading. The idea that you can, you can train models in ways that, [00:33:00] you know, are novel, never been done, but, but understandable to you. Right. I this what I just showed you this morning wasn't even possible two months ago and in it and if it was possible, it would require the smartest people on the planet to actually do it 

John Willis: crazy.

That's crazy. And yeah, so, so, so on the neural symbolic stuff, what, why, what is it about? I mean, I, I, you know, I've been following, you know, stuff that you can use Symbols and like some of your jailbreaks with symbols were just on the early Quinn stuff was just I'll put a post to that too, but just when people argue with me, you know, like a well known see.

So today was arguing about what's so bad about jailbreaks. I'm like, yeah, you probably should take a look at some of the stuff that, you know, that can be done with this. But, but what, what sort of like, why you and earlier you said you're finding more interest in the neuro [00:34:00] symbolic and. And some of the people that I interviewed about a year ago, just like real scientists, you know, and you know, and they were like, yeah, they, like, they made that conclusion that you didn't need a GPU for everything that you do here.

And 

Reuven: I think, I think the key here is, well, neurosymbolic reasoning separates. The sort of structure of language from the sort of fundamental structure of reality. And I don't sound like a bonkers person when I say that, but reality is essentially math, right? Everything around us is made up of either some kind of mathematical equation or some kind of quantum related sort of interaction.

Or certainly AI, 

John Willis: right? If we don't get into the metaphysical, I kind of lean towards what you're saying in general, but certainly when we're dealing with AI. It's all math. 

Reuven: It's all math. And the, the, the limitation you have with, with the current iteration of generative models is they're predominantly [00:35:00] language.

Now, language is, is, is great for interpreting the reality around us. Tell me about that guitar behind me. What color is it? Right. But when, when you look a little bit deeper within that guitar, it's made up of a lot more than just the sort of superficial description of what that guitar is, you know, there's the, the frequency that the strings vibrate at, there's the, the, the metal that makes up those strings, there's the, the wood within the guitar and the molecular structure of that wood that creates resonance for the, for the sound to flow through it.

All those things are not describable. Through traditional language, especially not English. So when you start looking at symbolic reasoning, you start looking at the sort of, you know, fundamental blocks of everything around us, and you can define those in ways that are more abstract using abstract algebra and other ways to describe those phenomenon.

So the these A. I can now understand the fundamental sort of, you [00:36:00] know, world around it without actually having to necessarily even have a way to describe it. Yeah, 

John Willis: yeah, yeah. So so this is really good. Well, I guess, you know, why, you know, this part, if, you know, if it goes south, I'll just cut it. But, you know, I'm trying to sort of understand, like, okay, we got your open AI, you got Claude, you talk about Sonnet and very good for code and instruct. 

You know, like Google's making a lot of noise, you know, about the, their models now. The Gemini and the Gemini Flash 2. 0, and there's all the benchmark stuff. What about IBM? What, you know, like, what is, what is IBM's, what is your view of IBM? I, you know, at surface level, I think ConstructLab is interesting.

A little less interesting is the Gemini model, but I don't know because I really, I think 

Reuven: if you're an, if you're a fortune 500 company and you're looking for an outsourcer, they're just another outsourcer among a sea of many in, in the world of [00:37:00] true agentic engineering and, and things that are actually on the breaking cutting edge, I don't see them anywhere.

 I don't, I, to be honest, in, in, in the real question to me, isn't just IBM, like, where's Amazon? Like what are they doing? Right? It's the nothing apparently, you know, how did they miss the boat so badly that you know? They with alexa and whatnot now you get google which are you know, they're they're doing amazing work with with Deep mind and their latest models, but they're proprietary And you know, even in the only their only real selling advantage is they're super cheap, which is a great one so they're they're They're an interesting play and then, and then you've got, you know, the open AI sort of Microsoft alliance, which is a bizarro alliance.

I'm not going to lie. Microsoft apparently can't do anything other than what open AI does for them, which is a question like, what have they been doing other than spending 100 billion [00:38:00] on their best friend? And then you have the Chinese. Who are the, by far, the most prolific creators of AI technology on the planet.

The, the, you know, I think some of our friends to the south are, are, are still in, in sort of a, a belief that they lead in AI. I don't know where or how they, they believe this, but it's the Chinese. By far, it's both economically and practically, there's nobody comes close in terms of output. 

John Willis: So how should we think about, so you know, like, because now you brought it to a point where I didn't want to just bring up DeepSeek because everybody on the planet is talking about and has some opinion about it.

And I, as you know, now I, I, I take your opinion very highly. So DeepSeek, right? There are the, well, first off, let's. Agree that nobody should be going to some API or dot com that's hosting Hangzhou, China. But, but I, my fear was [00:39:00] even some of those original models where people were just running them as is.

I thought that was sort of dangerous. Is there a way to think about DeepSeek? Or I can see that you think about DeepSeek in a way that's incredibly productive, but what would be your guardrails? I 

Reuven: wouldn't use, I certainly wouldn't use them directly through an API, you know, they they've proven themselves over the last couple weeks to be, you know, negligent when it comes to even the most basic security.

So right there, and I don't know if that's just moving fast and breaking stuff or, you know, or something more nefarious, but the result is the same now when it comes to using it in a host environment, it's. And I've posted some tutorials and how to sort of retrain it. It's, it's relatively easy to create your own version of their system, you know, and it costs a few few dollars.

Nothing. So, yeah, I think that the, the definitely thrown a wrench in, in [00:40:00] the sort of midst of this market, their, their models are, are AI. You know, so. Yeah. Now, anyone, anywhere can run these models and they've proven that, you know, these sort of trade embargoes and other things, these protectionist systems that we've put in place have no moat whatsoever and no meaning other than being lip service.

And if anything, those protections have only encouraged these supposed adversaries. To find workarounds that are hundreds or thousands of times more effective than what would have been done otherwise. So we've literally created the very adversary that we were worried about 

John Willis: the constraint. And O'Reilly wrote a pretty good article about how the constraints really gave them.

Disadvantage. So, so in other words, what you're saying, though, is if I grab the model and I do my sort of my own, you know, some level of training and sort of run it myself, that might be right now the more effective approach to solving [00:41:00] reasoning problems or agentic. Or is it just, you know, still take the 1 that fits best?

Reuven: Depends what you're doing. It depends. You have depends on your, your technical capabilities in house. As I said. It's hard to find truly capable people in the space there, you know, anyone can can go to chat GPT and, and masquerade as an expert, like, let's be honest, right? You can have a conversation with chat GPT and it'll answer any of your problems where the rubber hits hit some road is being able to actually take that information and do something practical with it.

Like, yeah. Other than maybe a LinkedIn post or something. So the key here is in the practical application of this technology, and that still requires significant amounts of human ingenuity to actually make make it happen, which is good for me, I guess. But 

John Willis: yeah, well, then. So that begs the question, or like, is it just going to be 1% ers that are going to be able [00:42:00] to use this?

Really? Okay. 

Reuven: Yeah, it'll be 1% ers. You're going to have, you're going to have a dichotomy. You're going to have this, this group of, look, we'll call them the 1% ers, but they're probably less than 1%. But let's just, for the sake of it, say there's going to be a group of 1% ers, whether it's corporate or individuals who are exponentially more capable than everybody else.

And I'm not even talking 10, 000 times. I'm talking multiple millions or billions of times more capable than the average individual on this planet, and that's going to change the very nature of work and you know how you effectively create things when you have a single individual that can do the work of a billion people.

What is that? What does that actually mean for the value of that work and the value of those other people doing that work? Well, it greatly diminishes the value of that work for everybody else. If one person can create 10, 000, 000 lines of code, what's the value [00:43:00] of those 10, 000, 000 lines of code? It's not what the interpreter tells me of which is like 200, 000, 000.

Obviously, it's not worth 200, 000, 000. It's probably worth fractions of a penny on the dollar for each line of code, you know, because it's so easy to create. And that is just for one particular sector. What does that look like when it. Architects or call center workers or the entire U. S. Federal government is running based on a I.

Where do those, where those 2 million federal employees go? What do they do? 

John Willis: Yeah, well, and then you get into sort of like, it's sort of like, there's a, Isaac Asimov wrote a trilogy on this topic, right? So you know, the, the but yeah, I mean, because then you get into like, well, then, which would probably don't want to go into, but the sort of the meta of, like, if nobody can, Work than who buys anything.

What is there to build? Right? But 

Reuven: well, there's a fundamental flaw in the form of [00:44:00] our of the state of capitalism in our economy, which was based on labor, right? And and labor essentially has been exploited over the last 200 years. And I'm not a socialist. I don't pretend to be one. I'm making my money in the way I can.

But the reality is, there's always been this kind of, you know. Reverse hierarchy. The 0. 1 percent that has all the money and then and the 99 percent that does all the work, right? And that system worked. It was a kind of trickle down of money enough to keep the lights on. That hasn't worked as well recently, but generally it's it's worked pretty well since the industrial revolution and now we're in a state where all of a sudden, you know, that 0.

01 percent doesn't even need the other 99%. 

John Willis: That's pretty scary or it is what it is. Right. So, and 

Reuven: it's. It's terrifying. Yeah. 

John Willis: Yeah. 

Reuven: It's, it, it, it, it means that you'll have a, a, a, a [00:45:00] a class system of the super empowered, super rich, and everyone else just scrapes by it and, and has no real purpose unless we redefine what purpose even means in a world.

Where we have abundant knowledge and capability now, and that's where I sort of step in and say, well, I can make my stuff really available and make it available to everyone to use it for whatever purpose they want. And it's it scares people, like, for example, the weekend, I put out a. A truth system. I guess you could call it.

I saw that. Yeah. Yeah. And I got a 

John Willis: little bit about that because that's that's 

Reuven: fascinating. It's working. Yeah. And there was two groups of people. The most vocal being I'm evil. I just put out a weapon of mass destruction. So let's take a step back for a moment. I create a system with that can do with and I built it in like a few hours.

So, you know, take take it with a grain of salt. But, you know, I generally know what I'm doing. So. But let's let's just assume that it [00:46:00] does do what I'm what what what I what I say it does and it can with a high fidelity to use multi modalities, you know, text or speech or sound or vision or other forms to detect whether someone is being truthful or not to a high degree and and you can run it through let's say a zoom meeting and anyone can run it and now I'm not gonna get into the specifics of like consent and EU, but yeah.

The ability for me to conjure up this thing just by asking a few of the right questions shows the ease of which others will likely be able to do this in the future. And in a world where everything's fake or artificial, having a mechanism to be able to determine what's true and what isn't is Critically important.

Should that be illegal for me to be able to discern that, you know, the person or the entity I'm interacting with is, is, is factually correct that that should not be illegal. That that should be. You know, that should be the opposite. That should be like the fundamental [00:47:00] right in a world driven by, by artificial intelligence to understand what is and what isn't.

But apparently, according to most people, it's a weapon of mass destruction that will undermine the very nature of our society because, you know, no one will be able to lie. 

John Willis: Yeah, I mean, well, you know, like, I mean, do you, you know, think about how we do truth today. We jury of your peers, which is. About, you know bias, 

like, I don't know what percentage effective, but it ain't it's not 95%.

Yeah, that's 

Reuven: why that's why you don't have capital punishment. Most modern countries and because humans are valuable, right? And and there is no coming back from from a mistake like that. So there there's, you know, we have safeguards in place. And for for those sorts of things. And we know we know that that these things can be abused.

You know, if you've ever watched minority report, you know, a system that can predict, you know, potential, you know, crimes can be used for all kinds of [00:48:00] terrible things as well. You know, can You know, you reach a point where the A. I. controls us in that scenario, right? If you're, you know, whether whether through manipulation or just by, you know, blind bias, I don't know, but we're time 

John Willis: sounds like your truth.

The system is like, 1 step away from an A. I. Turing test, you know? 

Reuven: Oh, we, we passed the Turing test ages ago and everyone was like, whatever, you know, and that's. That's the thing about AI is when you're living through an exponential growth phase, it's hard to, it's hard to actually comprehend that growth.

We've done more in the last month than we did the previous 50 years. 

John Willis: That's what I was trying to tell people. You know, was at an AI plumbers thing over the weekend and, and I and, and you know, so I got my new book coming out, right. It's the history of AI and, you know, part of. What I'm trying to say is, you know, some of what you need to understand is the [00:49:00] history of AI, right, to at least grok what's going on, but if you think about the last, I mean, like, the last two weeks have been insane, but the last two months, you know, three months have been, like, crazier than the last two years, and the last two years were crazier than the last 40 years that I've been doing this.

So 

Reuven: I showed you a self aware cognitive thinking system that I devised in less than an hour just for a social media post. If we're not living in the future, I don't know what, you know, what, what the future even looks like, you know, but the fact that we can create pseudo conscious systems that are able to interact and operate with no human oversight in itself is mind boggling.

Now, and now we have the ability to apply those, those systems to any problem and let it run forever. 

John Willis: I, you know, I guess just since I have you and you're not running away from me, [00:50:00] I, the question that comes up a lot is like emergent behavior. You know, I, I say there are these what I call anti AGI ers, right?

You know, and some of them are just cognitive science people and some of them are, like, there's a gentleman called Eric Lawson, he wrote a book called The Myth of AI, and And I think a lot of their argument is sort of, they base their argument on abductive reasoning versus sort of deductive or inductive and that gets very meta.

But in general, I think their argument talks around emergence and I'm on the fence, right? I don't know. I mean, do you have a sense of like true human emergent behavior or is there a difference? 

Reuven: You know, there's the, well, a critic is generally a critic because they can't actually do, you know, it's easy to criticize something you don't understand, right?

And if you look at a lot of the criticism in the space, you know, especially when it comes to like true capability, almost all the [00:51:00] critics that I see aren't actually practitioners. And you can tell by the way they criticize there. They're using theoretical constructs to define what's possible when all you need to really do is actually go and build a few of the things that I'm putting out there.

Go try my dream a I or my my spark protocol or my consciousness prompts and and then. Let me know what you think. But every single critic I run into is basically saying, you know, I believe this, and therefore it's true. Well, you're limited by the the expanse of your own mind when you when you believe something you've never tried.

And now, when it comes to, like, this idea of, you know, is it, you know, A. G. I. Is it an all knowing system that can answer any question ever for any purpose whatsoever? Of course not. That's not how this stuff works. Now, when you look at Nero, can it do programming better than, you know, [00:52:00] the 99 percent of programmers on the planet?

Yes, today. And six months ago, that was like 50 percent of the planet. We've gone from like 50 percent of the planet to 99 percent of the planet in less than six months. 

John Willis: Yep. Yep. 

Reuven: Yeah. And so now the question of AGI is not one of like, is AI better than us? It's obviously better than us. And no one questions that whether or not AI is going to be better than us in the near future.

It's as it expands into other areas, which areas will it be better than us? And what layers Will it, will it not be? Are there areas that Eliz never will, will never, you know, overcome a human like when it comes to things like EQ and, you know, understandings, you know, the actual sort of human elements of the world around us?

I don't know. Probably not. Maybe. 

John Willis: Yeah, all bets are off on any bet, right? So, 

Reuven: yeah, it's an alien intelligence. It's so [00:53:00] comparing it to a human intelligence isn't a fair, isn't a, isn't a fair comparison, 

John Willis: you know? Yeah. 

Reuven: We've, now it's a little scary that we've conjured an alien intelligence out of our own thoughts, right?

It's weird, but that's what we've done and here we are and it can do amazing things and it can also do. You know infinitely terrible things depending on which way you want to use it It's it's much same the same as a nuclear weapon You can power you can green power the entire nation with it or you can destroy the entire planet with it 

John Willis: Yeah, like everything else science realm.

Well, I think this was exactly what I thought it would be which was excellent I think people are really gonna enjoy this. I'll have a post I would put a final where do people if they were coming looking for you where but I'll have all the links up but just 

Reuven: If you're, if you're into the, the the more, you know, avant garde tech, check out my GitHub.

I'm [00:54:00] always posting all kinds of stuff. It's RUVnet. And if you're looking for the absolute cutting edge, look at my gist. That's where I post The bonkers, crazy ideas at two o'clock in the morning stuff. You know, some of the stuff will scare you probably should. And if you're interested in seeing live coding, we, we have the AI hacker space.

 I think it's agentics. ruv. io. And we hope we do live coding. People tune in to watch me code on Thursdays for about an hour, hour and a half at 12 noon Eastern. And then we have a more show and tell sort of structure on Fridays at 12 noon as well. And we've got our WhatsApp and Reddits and LinkedIn, you know, just search around for Rube or follow the links.

I'm sure John, you'll share here when you post this. 

John Willis: And I wish I would've got JMW when I, you know, when we started. When I first saw it, I went and got a stupid botch of a loop on Twitter. And when I saw your RUV, it was [00:55:00] too late. Like, I was like, man, that was brilliant. Getting a three letter name.

Well, I 

Reuven: went, I, that, I think that might have been right around the same time you and I met. And there was a party in Austin South by Southwest, I think? Yeah, yeah, okay. And, and the crew invited me, and they did their big launch party. And I got the, I could've had an R. Totally. 

John Willis: Yeah, I could have got, I could have got JMW at the time I got botched glue, but hey, what are you going to do?

Reuven: You know what? I'm not even on, I'm not even on Twitter anymore. I'm not being on Twitter. I know, 

John Willis: me neither, but now you're up, right? I'm botched glue. 

Reuven: Yeah, 

John Willis: yeah, there you 

Reuven: go. So we're, we've, we've got our handles and And the funny part of my, just a side story for a moment, my my, my domain is ruv. net, right?

I'm the only owner to ever have that domain. I bought it when I was you know, like 16, when the internet became a thing. And so, so and all my kids and my wife have at ruv, so we'd be like Brenda or [00:56:00] Sam at ruv. net. And it's almost like it's become like our last name almost. That's awesome. It's weird.

And people are like, what do you mean, you know, you're, you're R, right? R U V dot net, you know, anyway. Hey, just one 

John Willis: last, last thing, which is, I've been following your son playing guitar and all, man. I, I am all my nephews and nieces, I always buy them guitars if they get, if they're interested, none of them play it.

So I'm always a fan of a young, younger person picking it. It's been such a gift to me in my life. So to tell him some guy he never heard of or doesn't know is rooting for him to sort of continue. Guitars are. A great thing to pick up. And 

Reuven: he, he, he loves that thing. He plays and it's like, I, he's strum it at like, what is it, two in the morning and he's, it's plugged in and he's got the metal pedal going and he is, and he is doing like, you know, led Zeppelin or something I would expect from a 14-year-old.

It, it, it's, it's awesome. I, oh, he plays that thing all day long and he comes over and shows me the calluses on his fingers. That's awesome. That's awesome. He's like, or, or a new barcode. He learned. [00:57:00] Yeah. There you go. 

John Willis: I love it. I love it. All right, my friend. Thank you so much.