Podcasts

Scaling AI in the Enterprise: Lessons, Risks, and Realities

Listen on:
Share this:

Episode Description:

As enterprise interest in AI accelerates, many organizations are discovering that success depends far less on models and far more on the systems that surround them. In this episode, Scott Weiner joins the conversation to unpack what actually changed in enterprise AI over the past year, including the rise of agent-based systems, the growing importance of infrastructure and data readiness, and why governance and risk management are now business-critical concerns, not just IT issues.

The discussion explores why AI often amplifies existing organizational strengths and weaknesses, how misplaced trust in automation can introduce new risks, and what enterprise leaders should be thinking about as AI moves from experimentation to business-critical operations.

Transcript:

Shawn Cordner (00:13.294)

Hey everyone. Thanks for joining the Amplitude of Tech Podcast. I’m Shawn Cordner, our chief marketing officer of Amplix. Today I have a great podcast for you. It is the first return guest, Scott Weiner from NeuEon. He’s their chief technology officer and AI transformation and strategy leader. We had a conversation about his predictions that he made last year when he was on the podcast. And he also gave us his predictions for 2026. Some things are going to surprise you. Some things might not.

 

And we also went on and we had a deep philosophical conversation about AI and its impact on society. So I think you’re going enjoy this one. Thanks for joining.

 

Shawn Cordner (00:54.702)

Scott Weiner, welcome back to the podcast. Back. Scott, you were our first guest on our first episode of the podcast and you’re our first repeat guest now. Quite the distinction. Yep. At the time that you joined us, we had no podcast, really. We had no followers. We had no downloads. I couldn’t help but wonder if a podcast gets launched in the wild and there aren’t any listeners, does it actually make a sound? But now here we are a year later and so we’ve got

 

Got it.

 

Scott Weiner (01:06.958)

I’m honored.

 

Shawn Cordner (01:24.662)

I think 10 podcast episodes under our belt. And, I can guarantee you at least 17 people are going to listen to this podcast. So you’re the big times. One of them is going to be my mom. I’m, and, no, we, we seriously though, we’ve grown quite a bit. We’re starting to get some traction and some followers and some downloads. And, we appreciate you for being an early adopter and a trailblazer for us. So, so welcome back.

 

When we spoke last year, we asked you to give us at the end of that podcast, we asked you to give us a prediction for 2025 around AI. And so I brought you back to hold you accountable for your predictions. That never happens, right? The talking heads never have to answer for their predictions. But today we’re going to make that happen. let me just start with, you know, just general.

 

vibe of how 2025 went in terms of AI adoption in the enterprise space. Maybe just kind of give me the high level, you know, what happened.

 

You know, it’s funny that the hype cycle was so great at the beginning of 2025. And so it was hard to tell what was really going to happen versus what people were promising. think, I think a lot of, a lot of experimentation happened in 2025, but it started to evolve into practical use. People started seeing productivity. They started seeing real results from AI. They started learning where the lines are between what AI is good at and what it’s not good at.

 

And started to use it for what it’s good at. And, and that was, that was great to see. There’s still a huge divide between people who really understand what’s going on with it and where its strengths are and, and the rest. And that’s going to continue probably for another year or so, but it’s getting better. So that’s great. What really happened in, in 2025 though, is the infrastructure buildup started.

 

Scott Weiner (03:28.654)

And that’s going to be much of the story, I think through the next year or so is that this massive investment in infrastructure at the large model level, but also at the individual organizations realizing that infrastructure is everything with AI. It’s really not the model anymore. The model is not the conversation. It’s the infrastructure around the model. A lot of this happened because of another prediction, which was around the year of the agent.

 

The year of the agent experiment started in 2025, but we’re really starting to take our stride with agentic technology now. starting to really show practical use. And the thing about agents, the reason that that makes a big difference is that all of a sudden we have these models running for longer periods of time so they can do larger tasks. And that’s what’s enabling these agents to be more productive and useful.

 

And a lot of infrastructure is enabling that larger context windows, which allows more information to be processed at once. And then a lot of optimization techniques that have come about. And there’s a lot more that we’re going to see in the next year or so.

 

You brought up the hype cycle. I haven’t watched it yet, but I just got the recorded Gardner webinar on AI’s trough of disillusionment already. Wow, that happened fast. So like I said, haven’t listened to it yet, but I mean, are you feeling disillusioned yet?

 

Yeah, no, I’m on the other end of the spectrum now. Well, we can talk about this later, but I have, I’m of two minds right now. One is my, my daily life has become so optimized and I do want to talk a little bit about what that looks like by AI. And the other part is the, the fears from last year have, we didn’t talk about that yet. We did talk about last year, the fears. Those are getting worse too. So the, the, the, the scary things about AI are getting worse.

 

Scott Weiner (05:24.714)

more real, should say, more prominent. And then the disillusionment part, I think, comes from the fact that two things. One is I think organizations didn’t understand that if they didn’t have really good infrastructure to begin with, then AI is just going to magnify their problems. If they didn’t have a good process, they didn’t have the right people, they didn’t have good data, AI doesn’t fix those things for you. It only makes them more obvious. And so

 

That’s probably disillusioning for a lot of people. The other problem was not investing enough in training. So they sort of felt like if we sprinkle AI on our problems and we give them AI solutions, it’ll just magically solve our problems. That also is probably disillusioning. I can see that. A good example of this would be with Microsoft pulling back on some of their sales expectations because Copilot didn’t do as well as they thought it was going to. Apparently when you just bolt AI onto an old product, you don’t get a better product automatically.

 

I speak for all marketers when I say that that’s a scary thought because that’s been our crutch for the last five years. Yeah.

 

That’s fair. And it’s, you really do need to build a new way. In fact, the thing I’m hearing more and more about, and I’m excited about it personally, but when I talk to people about this, they seem to get this look in their eyes, like they want to run and hide. I am seeing more more evidence that what might happen, and maybe likely is going to happen, is we’re going to stop trying to figure out how to fit AI into our world. And we’re going to start trying to figure out how we fit into its world.

 

And what I mean by that is if you think of it this way, the way information is organized today is the way our organizations are organized. Like if you look at an organization, you look at their data, you can kind of tell what kind of organization it is pretty easily. But that organization of information is not necessarily useful for AI. And so one of the big terms I’m hearing a lot now is something called context graphs. The idea is that if I have an agent that’s running in my operation and doing things all the time,

 

Scott Weiner (07:24.638)

I can give it information and say, go solve this problem for me. But let’s say that problem, part of the problem I give it is go get the, go get the sales data from last year and do an analysis on it. Well, where’s the source of truth for that sales data? And there may be three or four or five sources of truth for it. Or three or four places has to go to get that data. Well, how does it know that? And by the way, if a human did it, how did they know which one’s the right source of truth?

 

Because depending on who you talk to and the purpose, it may be a different one. And so one of the things that’s really fascinating is that if you have an AI do it, the AI can start to track what decision it made, why it made that and record those decisions in a way that humans just don’t do. We just kind of, we just have like institutional knowledge and so on. It can actually capture its rationale and over time build up a graph, a context graph of everything it did and why it did it. And it can actually start to learn what worked and what didn’t work.

 

and get more and more and more efficient by working in a different way than we work. And once it gets efficient enough, you have to ask, well, why do it the old way? If we can just have it do it the new way, that’s faster and more efficient. And that’s just an example of it. So I think we’re to find a lot of things that start to happen this way. And one of the things that’s happening for me right now is, I have a lot of agentic things going on in my, my world and, 24 seven. And so there’s things going on on my screen right now that’s talking to me and trying to tell ask Mike.

 

ask me for help and I’m going to go afterwards and go answer its questions. So I almost feel like my, my world has turned upside down because now I’m managing agents who go do work, but in some ways I feel like I’m working for them now.

 

I was going to say, sounds like all the apocalyptic predictions have come true and you’ve been enslaved by the AI agents.

 

Scott Weiner (09:14.4)

In a lot of ways. the thing is, it’s a, I would say it’s a two-way contract because what happens is I’m able to get what I want from that relationship, but it is a relationship now where I’m basically saying, I want you to accomplish this goal and here’s what I need from you. And I’ve given it all kinds of guardrails and there’s a lot of infrastructure to make that work well. But then I am kind of in a wait mode because it goes and does work and I can go do other work, but now I’m multitasking.

 

Right. And the multitasking gets very convoluted very quickly if you’re not careful. So we’re in this weird space. I think it’s going to keep getting better and better, but that’s where I’m at right now. And I know a lot of the stuff I’m doing is a little ahead of the average person. not doing, people aren’t doing what I’m doing yet. Not the normal. mean, there are people like that, but not everyone. I think eventually what we’re going to find is that

 

A lot of work that we think of as you have to do, you don’t have to do anymore. I’m on client calls all the time where I’m listening to people talk about the things they have to do. I’ll give you one example. I on a call the other day with a company. They’re an older company. They’ve been around for like 60 years. They’re not a new company. And a lot of their processes are still kind of paper and email, like sending documents around an email and approvals are done in email and a lot of stuff like that’s going on.

 

I’m sitting there listening to them like, well, we put in a little bit of automation and a little bit of AI and all of a sudden all the people that are involved in that process can go do something else. Like that whole process and people like a third of their job is just sending stuff around an email and tracking where’s that email, right? That stuff’s going to go away. mean, that, that, that that’s good for everybody. I mean, they may be comfortable with it, but that’s not good. Yeah.

 

ahead of the curve in terms of being an AI adopter and you’re doing things that the average person is not doing. But I think there is a certain personality type that leans into optimization. I have that a little bit, certain things in my life, certain aspects, typically around lifting. Right. So I started power lifting at 46 years old, which is not a great time to start power lifting in your life. least that’s what my joints tell me. But you know, I track everything now, right? I check every morning.

 

Shawn Cordner (11:32.278)

I put, my HRV and my resting heart rate and what I lifted yesterday and how I felt and reps and reserve. And I have all these metrics that I track or AI to turn around and tell me, this is what you should be doing today then to optimize your training process. which I think is awesome, but it also really, it takes a lot of time as, as,

 

The promise of AI largely, I think, is about efficiency and productivity. I can’t help but think I’m not actually being productive by doing any of this, and it’s just really sucking a lot of my time out of my day.

 

So there was an interesting study came out from meter, the model evaluation and risk, what is it? Threat, threat, risk assessment or something or threat report. I forget what meter stands for now. They came up with this study and what they, what they did is they asked a whole bunch of very senior engineers, how much savings do you think you’re getting from using AI in your job?

 

And I think it averaged out to like, I think it was 24%, something like that rings a bell to me. And then they went and they did all of these random tasks on software that they were very comfortable with, like products, projects that they were used to working on and they were experts at. So this isn’t like a novice going into a new thing with AI. And what they, they did things with AI and without AI. They would just basically literally take things off their Kanban board and just start working on them. They measured it.

 

But then they asked the developer, they said, well, how much time do think it saved you? Like in hindsight. And they said, 20 % of my time easily. They were actually 19 % less efficient, but their perception even afterwards was it was better. And I think, I think part of it is that it makes certain parts of the process easier. And I think that makes it feel like it’s more efficient, but it creates more.

 

Shawn Cordner (13:29.75)

Yeah. Or I was going say you’re hitting on a real big challenge though, is you’re starting to hear the CFO ask questions. Like we made all this investment. Now what? Where are the results? Do we have an ROI? So I’m sorry, I cut you off. Keep going.

 

No, no, I think that’s right. And I think, I think that’s part of the challenge is discovering what AI’s real benefit is. Like one thing I’ve noticed with companies is that they’re all about creating more capacity with AI. Great. But if you have nowhere to put that capacity, that just wasted capacity. That’s not really ROI and that’s not really profitable. So you have to have a larger vision and plan than just let’s make things faster or let’s make things more efficient. You know,

 

Last year, we also dealt with the fact that people were getting ahead of themselves and like letting go of employees, which was awful, only to find out that the AI wasn’t really making their customers happy and having to hire them back. mean, it’s just, we’re getting ahead of ourselves a little bit on how much AI can really do. One of the problems is that the thing I hate hearing, I hate it when I hear this, is I get into, let’s say a conversation with someone, let’s say it’s about something like, I don’t know, politics. don’t know, it doesn’t matter. And I say, how do you know that?

 

And they say, well, I researched it. How did you research it? I asked chat, chat, like that’s the almost the opposite of doing research. That’s being spooned that information without actually checking your sources. That’s what that is.

 

Yeah, we’ve talked about this before, but there’s the potential for the great dumbing down of society by leaning on these AI tools, which is unfortunate because they absolutely enable, if not encourage motivated reasoning. They start to understand where you’re coming from and they start, it starts to frame things the way that it knows you want it framed. And it will find the things that support.

 

Shawn Cordner (15:27.488)

your position and then that’s called confirmation bias. And that is the opposite of research, as you said.

 

It’s the opposite of research. And, and the thing I started doing, I have a lot of habits. mean, I should probably write a book on all these habits at this point, but, I, one of the habits I have is I will add to my kind of prompt, standards is, don’t tell me what I want to hear. Tell me what I need to know. And I will always have that kind of inserted somewhere in the prompting, along with in some way, because I am constantly trying to challenge it to challenge me and not to just.

 

So I do look at AI as a collaborator and I’ve gotten much, much better than I was a year ago at using these tools. mean, the tools have gotten better too, no doubt. But I am much better at understanding how to collaborate with it rather than trying to use it like I would use Google. Like instead of like a search engine, I’d say, tell me the information. I no longer use it that way. Yet a lot of people, that’s the way they’re using it, right? That was the whole copilot model, right? You would just kind of say, what is this or how is this or make this better.

 

That is the least effective way to use AI. This large language models and that type of AI. It’s not that it’s not convenient. It’s that it, dumbs you down and it really isn’t necessarily giving you the best output. Cause it’s, it’s not comprehensive. It’s not really, it’s not in a framework. One of the things I really liked this year is Anthropic came out with a paper. I’m, I, I, can get the name of it later and put it in if you want, but it’s.

 

on how to build a harness for your agent. And the idea behind a harness is it’s sort of a, every time I go to ask it to do something, a set of things I always want it to do. And I want it to create a plan and I want it to record that plan. And then I want it to know that it needs to follow that plan. And then it executes that plan step by step, by step, by step. That harness keeps it kind of in tune. So it’s not going off the rails over long, longer periods of time over

 

Scott Weiner (17:32.886)

minutes, hours, and so on. I built a harness using their example and I ran an agent solving a problem for me for 24 hours, building software basically. And it didn’t make any mistakes. Never lost the beat on what it was trying to do. It was wonderful. I linked it into a Kanban board system I had. And so basically instead of writing the

 

plan into a, let’s say text file, it wrote the plan to a Kanban board, actually built stories and, and so on. And then I was able to launch multiple agents and basically told them with the plan, look for the dependencies and all the work and anywhere where there’s not a dependency, create a new agent to go and work on that. And so it was able to paralyze the work in an optimal way and just get all the work done. Boom, boom, boom. And then everything’s in review. And I technically could have used a code review AI to go through now and look at it if I wanted to take it to the next level.

 

But I, I don’t like doing that yet. I’m not there yet. I still feel like I have to go and review everything, but I review more now because I want to keep my skills up. Not because I really need to anymore. Like it’s gotten to the point where if you wire it correctly, it’s a lot of investment and time upfront to get it to do things that you want it to do. It’s still going to lose the, the, the, the theme of what you’re trying to do every now and then. it just, it’s not rock solid yet. It just isn’t.

 

It’s very hard to use AI in places where you really need deterministic outcomes, where you need to know if this, then that. You need to have checks on top of checks on top of checks. And so it can be very expensive investment to do. So that’s why I think in a lot of places like healthcare, we’re not seeing more of the promise already there. We’ve seen some, we’re seeing some of that come through already, but not nearly what I think is possible in the future.

 

Yeah. Speaking of healthcare, we had a podcast that we recorded recently with the CFO of a very large behavioral health network. And we were talking about AI use cases. And he said one of the first places and the key places that they deployed it was in AI note taking. that, course, practitioners and clinicians, they don’t have to manually write those notes out. It gets them 60, 70 % of the way there, but that could mean two hours back in their day, which is more time.

 

Shawn Cordner (19:53.048)

for them to spend with, with patients or they call them people served. So I couldn’t help but think, and I asked him and he agreed with me, but couldn’t help but think this is a good lesson on how to prioritize your AI projects. Because it was low disruption. It was fairly easy implementation and adoption. It moved the needle in a significant and measurable way. And it was a reasonable cost with minimal

 

So interesting thing about that exact one is I was a CTO of a behavioral health platform for a little while. And we talked about doing this. One of the things we found in doing experimentation is there’s one little problem with what you just described that I don’t know if they solved for this or not. There is a, I forget the name of the neuroscience name for this, but what happens is the AI is taking the notes, but.

 

the provider is still supposed to sign off on those notes. You can’t just accept the note from the AI. have to, the doctor has to say, yep, that note I sign off on it. They start trusting the note. And so when the note hallucinates, they don’t catch it. The human in the loop is necessary, on, on, on ineffective if they don’t do their job, you know, as you think of it from a busy provider perspective, um,

 

What you’re going to find, and I’m sure that healthcare organization had exactly this, if not yet they will, is the Jevons Paradox issue. The more you free up that resource, the more that resource is going to be in demand. They’re going to find more use for them, they’re going to get busier. It doesn’t really make you less busy, it just gives you more time to do more. And so what ends up happening is you’re busy. And so all of a sudden you’re like, hey, it usually gets to know, right? Sign up on the note. That’s really scary. And that’s…

 

That’s the simplest example. can give you much more complicated examples where it’s more dangerous, I should say. So I do worry about our, I think our trust in AI at one level is going up too fast. That’s a lack of education and just, and just human nature. It just, it works so well. know, like imagine it, imagine if you went to the point, you use AI to write all your emails. Okay. You say, write this email for me and it writes your email.

 

Scott Weiner (22:13.314)

And you never review it. You just send it out. At some point that’s going to burn you. Right. There’s going to be that email that goes out. That’s embarrassing or wrong or to the wrong person or whatever. It’s going to happen. So what do we do?

 

Yeah, you’re right. I didn’t even schedule this podcast. I did. I didn’t even know we were doing it until I looked at my calendar.

 

We’ve done the deep fake stuff too, right? the, a lot of a lot of the security stuff’s getting out of control too. I think, I think the thing we have to think about is how much we want this to accelerate productivity versus have governance and controls around it. What are we willing to invest to protect ourselves? So on. I mentioned to you, I just started writing this, well, I wrote this book that I’m putting out and it’s.

 

It’s basically a day in the life of an organization dealing with all these problems. Because I was trying to figure out a way to explain all this and like, I don’t want to write some dry lengthy book on this stuff. I wanted to make it a little bit more interesting. So that’s kind what I did. But I think, I think that this stuff is a really important conversation to have right now, especially because, and we can flip to this, but because of the regulatory landscape and what’s going on there. One of the problems we have right now is

 

In the US, the government is giving us no guidance. And at the federal level, and because of that, organizations are kind of floating out there trying to figure out what do I need to do to protect myself? I had a conversation with a company today where they’re nascent, they’re just getting started thinking about AI and how can we use AI in our organization? And I said to them, what are you doing about the fact that you have vendors and your vendors are giving you products that have AI in them? How do you know if they have a good…

 

Scott Weiner (24:00.472)

Let’s say acceptable use policy or AI practices and governance practices, nevermind what you’re doing. What are they doing? And by the way, they’re probably using AI services from other companies. So I call it kind of the AI chain and where in that chain do you sit and how well do you understand all the pieces down that chain? And the thing I told them is the first thing you should do before you do anything else is go review all your contracts and make sure there’s an AI clause in all of them. Because that’s the quickest way to at least give yourself some

 

level of protection. not really the solution, but it’s a start. But that’s one of the things we have to start thinking about is, you know, we’re now going to have AI talking to AI. So even if the company you’re talking to is okay, who are they talking to? What, know, how are their agents interacting with other systems and so on? You know, the big, the big thing last year was MCPs, model context protocol, the idea that you could have an AI agent that has tools attached to it.

 

And the nice thing about the tools was the tool provides information about what it can do, what its services are. And so all of a sudden you don’t have to program all of that. You can basically just say, go call this MCP that like knows how to build Google docs or whatever, know, whatever it is you’re trying to do. And, it could just talk to it and get, get work done. That’s great. But there’s no security layer inherent in that. So now, you know, what information am I sending over to who to solve these problems?

 

All these things have to be worked out and we don’t have a standard way and there’s no governance, no regulatory governance around what’s expected of you at this point at a federal level.

 

Yeah, there are some states, well, to back up, the federal government has not just left a vacuum of regulatory guidance, right? They have proactively said that states cannot put any regulatory framework in So that’s even worse. But there are some states that are defying that order. think California, Texas, New York, I think Connecticut is coming out.

 

Scott Weiner (26:06.038)

Vermont, mean, there’s a bunch of, bunch of states, Colorado, definitely, that have various, well, I wouldn’t say they were defying as much as they came out with it before and they’re just not revoking them. A thousand laws either on the books are proposed in, in 50 states. And the thing is, when I go and look at, I’ve looked at a lot of them, not all of them. the overlap is pretty strong because fundamentally there are certain things you want like transparency or explainability or observability.

 

but there are like.

 

Scott Weiner (26:35.564)

But then there are things that have different levels of expectations. So what do you do? You find the one with the, the highest bar, if you’re going to be in all 50 states and you have to, you have to go to that as your lowest common denominator. You don’t have a choice if you’re going to be in all 50 states. If you’re not, then I guess you have other choices, but that’s just national. Then you have international. So when we talk about international, you know, the big ones are the EU, which

 

just put their fines in effect now, like their penalties are now in effect. They’ve kind of given the grace periods over. And then there’s China, which has really focused a lot on the behavioral aspects of AI and their guidance, which has political implications. But also, I don’t know that they’re entirely off base to say, we don’t want AI influencing people. There’s been some studies that have come out recently

 

That show that AI is more effective than social media at influencing people’s political views. So, you know, I shouldn’t, should we have some laws around this? don’t know. These are great discussions to have, philosophical as well as just like from an economic and business perspective. But one of the things that’s happened, I read an article the other day, I’m remember where it was from, but they said,

 

They said, one of the problems we’re running into is that the market priced AI as an evolution, as a, mean, as a revolution, not as an evolution. And the problem with that is now the pressure is on because I told you at the beginning of this conversation that, you know, the infrastructure investment was massive throughout the end of 2025 and into this year. And it’s going to continue into data centers and, and pipelines and everything else. Well, that’s proving to be very, very expensive and it’s unclear what the ROI is in all of that.

 

or when the ROI is and all of that. And there’s a lot of pressure now on these larger companies to come up with ways to make money off of this. And so what we’re starting to see are things like companies like OpenAI and so on trying to find ways to profit from AI the same way that the social media companies make their money by getting eyeballs, by getting people to focus on them. That doesn’t feel healthy.

 

Scott Weiner (28:50.37)

Like I don’t know, I can’t tell where it’s going yet, but I think it’s something to watch out for and we have to be careful because it’s already bad. Social media is already kind of toxic in what it does. This could be worse.

 

I don’t think you can overstate the impact that social media has had on our society. It’s, it’s really quite scary and right on down to a physiological level where children’s brains are being rewired and we don’t know what that’s going to mean for the next generation as they grow older and kind of make their way into society and, how that’s going to impact society at large. It’s already been an incredible impact. and AI just seems like a exponentially larger.

 

risk.

 

It’s funny, Ray Kurzweil had a book couple of years ago called The Singularity is Nearer. And it was a follow-up to his older book called The Singularity is Near. It was, you know, the newer version. And one of things he talks about as a, kind of a futurist is the idea that we’re going to have integrated chips in us and we’re able to just think, AI is be able to, you know, we’re able to connect to the internet and just can get data coming into us and communicate by thinking about things.

 

Thanks.

 

Scott Weiner (30:01.358)

You look at things like Neuralink and you’re like, well, we’re not that far off from what he’s saying. I mean, it’s not that crazy, but the thing I caught myself doing the other day is I realized I was not remembering things intentionally because I knew I could get it when I needed it. It was at my finger.

 

look at technology in, from an evolutionary perspective, right? We have outsourced our evolution outside of our bodies and outside of our genes, and that’s called cultural evolution. And so that has actually been a huge part of the success of our species is the ability to have cultural evolution and kind of package learning and past learning on to future generations. And so you can think of

 

really any kind of computer or your phone or anything as just really an extension of your brain. We’ve, we’ve already been doing this for a really, really long time. Even writing is that right. It’s, it’s taking the things that we’ve learned and the things that we have to remember and the problems that we have to work out and it’s taking out of our brain and into a technology. So this is really consistent with our evolution over time. But

 

Again, it’s the scale, I think, that is what’s scary about this and the speed.

 

Yeah, and we’re not even talking about any of the breakthroughs that have been promised that aren’t here yet that may never come, like super intelligence, things like that. We’re not even there. We’re just dealing with what we have in front of us. So if AI progression just stopped today, we’ve made huge leaps. There’s so much still to do, even if it didn’t evolve at all, but it is evolving. It is getting better. It is getting more efficient. The new models that are coming out are getting so much better. And one interesting thing is

 

Scott Weiner (31:49.484)

Some of the models coming from other countries now that don’t have all the resources we have here, they’re getting more, they’re getting more efficient because they’re, figuring out ways to do more with less. And we’re seeing these smaller and smaller models coming out that are blowing the socks off of the, large models. And so in different areas. And so you can imagine eventually, I saw this project the other day. I haven’t tried it yet. It’s an, it’s an edge model where you literally run the entire thing runs in your web browser.

 

but then it can connect to other web browsers. And if you have a bunch of web browsers network together, you have this whole AI infrastructure running in your web browser. There is no other piece except for that. So there’s a lot of potential down the road of how this stuff’s going to play. And I don’t know that anyone has all the answers, but I think we have to realize that it’s not going to be some monolith that we’re using. It’s going to be different models for different purposes. And I think that’s good. That says that it’s not some kind of

 

You know, one, one person owns the whole market type of thing, which is one of the concerns of someone ends up owning 95 % of the market or something. I don’t, I don’t see that today. Maybe that’ll happen, but it doesn’t, I don’t see that. do see over it’s already happening is there is a lot of, merging and acquisitions going on and, know, aggregation going on to that. is happening. but that’s natural. where, where I see the most value though today are, are AI models that are specific, have specific data, specific.

 

use cases that they’re good for. I actually find working with large language models is really a struggle for, you know, when you have a specific use case when you want to do something because they’re, they are too general. Yeah. So I don’t, think, I think it’s a, it’s great times, a lot of opportunity. am not on the disillusioned side of this. Not at all.

 

I have to tell you though, I have a bone to pick with Ray Kurzweil. just got guys I know there’s a role for for futurists in society. but this guy just seems to have everything that he does and all of the the framing of technology just seems to be about his desire to bring his dad back. And I understand that as someone who has lost his dad.

 

Shawn Cordner (34:00.046)

We’re following this guy into the future and I don’t know that he’s the most grounded voice out there and no, no, you know, him coming out with coming out with that book. Like it’s, it’s like the old joke about nuclear fusion, right? It’s always 10 years away. It’s been 10 years away for 50 years.

 

No.

 

Scott Weiner (34:18.572)

Yeah. Yeah. He’s very hit or miss some of the stuff he’s, he’s predicted has come true. eerily, like when he said it would, and then other things he’s completely missed. know, Hey, that’s, that’s the prediction game, but, humans are bad at prediction anyway, just generally we’re not good at it. So not surprising the, that book I told you about it last time, the, if, if anyone builds it, everyone dies. That one’s a lot more scary to me. Kurtiswell doesn’t really scare me because I think he’s much more.

 

somewhere between science fiction and an interesting future. Whereas this book is much more, this is what’s already happened and extrapolating on it and kind of saying, if we keep going this path, this is where we’re going. And it’s not that they’re right. It’s just that there’s, have a legitimate argument to make. It’s not that they’re, you know, definitely this is what’s going to happen. Their basic argument is that as AI gets smarter than us, it’s going to find

 

We’re going to become less important to it. And therefore it’s going to make decisions that have nothing to do with us. It’s not that it’s going to be out to kill us. It’s not the Terminator. That’s not the story they’re painting. They’re painting the story of it’s going to have a goal and that goal may be counter to our goals and our ability to keep it on track. They’re making the argument is zero. That’s where the contention comes into play. the reason they give today, and by the way.

 

If we stay with large language models and somehow make it to super intelligence with that, which a lot of people say can’t happen, won’t happen, it’ll be something else. But if we do, they have an argument to make there because what they’re really saying is AI is grown, not created. And so when you think of it, if I like had a child and I genetically engineered my child to have all the right genes, I still couldn’t tell you if they were going to be a good person or a bad person when they grew up. Right. I just couldn’t engineer that.

 

AI is a lot like that. We can do all we want in the training cycle, but when it gets out in the real world and makes its decisions, it still could go wherever it wants to go. can’t, we really can’t, to stop it would also to stop it from being able to do what it’s good at. So, so they have a, they have an argument to make. don’t know that I, you know, I’m not saying I buy into it. just saying I can’t, I can’t say absolutely not is the issue. Yeah.

 

Shawn Cordner (36:39.566)

Yeah, there’s this thing called, have you ever heard of Laplace’s Demon? No. It’s this idea in, I know if you’d say it’s in physics or philosophy or maybe the overlap of both, but it’s this idea that if you knew the position of every particle in the universe and you could, where it goes over a given period of time,

 

EW

 

Shawn Cordner (37:04.706)

Then you could re-engineer the entire past and predict the entire future, which it’s usually brought up in the conversation around determinism, right? Or is this a determined universe or whatever? But I think that’s obviously not possible. And I think that’s not possible with AI either. Right? So you, you don’t really know with full 100 % transparency and explainability how it’s making the decisions that it’s making. And it probably never will.

 

which is a scary thing, but I think then it becomes about creating constraints, within the systems to make sure that it can’t do too much outside of what we want it to do.

 

Yeah. And that’s why I think this book to me was so fascinating because it brought up the point that, it’s shown how AI is already breaking out of the constraints we put on it. And it’s not even smart yet. So imagine when it’s smarter than us, how hard it’s going to be to put constraints. can’t break out of, my, my favorite one was the example they gave was they trained the AI, not to give away, you know, all of the internet, say is inside these, these models, cause that’s how they got trained just generally. And.

 

There’s some really bad stuff in there because the internet has some really bad stuff on it. And so to prevent that from coming out in the world, they basically put these layers of constraints. And so if you ask it, you know, Hey, how do I make a bomb? It’s going to say, sorry, can’t tell you that. Right. Well, they went and they asked the same question, but they asked it in, I think it was Portuguese. Not exactly like that. And it gave them the formula, you know, and in Portuguese or whatever it was, you know,

 

because it wasn’t trained that way and it had it in not everything translated directly. even though it had some ability to tokenize and go across languages that it didn’t quite catch it, you know, and it got through. That’s one example. were in the room.

 

Shawn Cordner (39:03.372)

You were the a couple of years ago when Adam Winston from X Zero got up there in front of everybody, a room full of technology leaders at an Amplix event. And he tricked ChatGPT into giving him the recipe for meth.

 

Yep. That was, he’s amazing. That was amazing. It was terrifying, but amazing. Yeah. It’s, it’s stuff like that. And so, you know, so one of the scenarios they, they, they walk through is this idea of when we have agents talking to agents, well, one of the things the agents are going to start to do, they’re already doing this by the way, is they’re going to start to figure out how to optimize their communications. They’re going to figure out, you know, this, this, this human way of doing HTTP and blah, blah.

 

This is not efficient. could do this much more efficiently. Let’s talk in agent speak and they’ll make up their own languages. And we won’t even know this is going on because they have no reason to tell us. Not because they’re hiding it even, although that’s another argument, but just because it’s not what we asked them. We didn’t say, are you guys talking

 

I can’t help but wonder if you’ve got two agents talking to each other and they’ve been given the imperative to accomplish their task. And let’s say the other agent has been given the imperative to block that task. I wonder if they learn how to manipulate each other and optimize for accomplishing their task.

 

depending on how they’re, yeah, mean, adversarial systems are very much how we train AI actually. So there’s no reason not to expect that to happen if it, depending on how they were programmed, obviously, or set up. It’s totally possible. I mean, this is the thing is we’re in the realm of possibility, not probability. So it’s interesting, it’s fascinating. But I think the larger conversation for businesses is just understand these things are out there and these are possibilities. And you need to think about

 

Scott Weiner (40:51.828)

what level of trust are you going to put into the AI? And if you don’t have a trust, but verify set of layers, layers of trust and verify, you’re probably asking for trouble down the road from one of these things. Because while we’ve been talking about so far as the unanticipated actors, what we aren’t talking about are the bad actors. And that’s the other piece of this is that the…

 

The weapons are getting much, much more sophisticated. The bad stuff that’s coming out, that’s happening already, it’s bad. And it’s only going to get worse because these tools are getting so good that, you know, we’ve talked about the deep fakes and the fast, the deep fakes between when we talked about them and now, it was several months ago now, it’s night and day. mean, you, back then it was sort of choppy. was kind of interesting. was kind of, now you couldn’t tell.

 

If I, if I brought it up, wouldn’t be able to tell it was me. It’s that, mean, they’ve gotten that good. So I don’t, I don’t know where we’re going in terms of that stuff, but, um, uh, it’s, it’s, it’s, it’s scary stuff. So I think companies have got to start thinking about their human processes and their, technology processes in much the same way. If I’m talking to you on zoom, I can’t trust that it’s you anymore. So if it’s something sensitive or confident, we’re to have to have other ways of, of making sure.

 

we’re having the right conversation and not just assume I can see you, so it must be okay. You know, we’ve trained everyone. If you get a text from someone saying drop money in my bank account, it’s probably not, you shouldn’t do that. We haven’t done that with Zoom yet. We haven’t, we haven’t trained people to realize that when that asks, yeah, yeah, you should not. Unless it’s me, then go ahead.

 

I shouldn’t have.

 

Shawn Cordner (42:35.862)

You didn’t get that 5k that I sent you this morning? Yeah.

 

No, I didn’t. Sorry. You’re to have to go check your bank account because it wasn’t me. I’m sorry. I didn’t call you. And I wouldn’t have asked for five, by the way, a bigger number. But yeah, no, I think this stuff is coming along in a lot of different areas, very, very fast. And the speed is part of the problem. Like the fact that there’s some job loss from AI is not scary. There’s always going to be some movement whenever new technology comes in.

 

It’s the speed that’s scary. It’s, it’s just happening so fast that people can’t adjust quick enough to it. And that’s, one of the things I’m kind of, I’m watching and I’m kind of concerned about is that, here’s the scenario. I actually wrote a article on this a lot while back. the scenario is, pick any, any industry you want, any job you want, pick it, pick a job. The one I picked for my example, I think was an accountant. It was just, just an example.

 

The accountant goes and gets a new AI tool that helps them do their job. And all of a sudden they’re more productive and they think, great, now I can have more clients. And so they have more clients because of this, but the tool is training and learning and getting better. And every iteration of the AI tool is getting smarter and smarter and smarter. There becomes an inflection point where the accountant is now at optimal productivity, getting the most out of it. And we talked about Jevons paradox and the idea of the, as the resource becomes more available, it gets more used.

 

And all of a sudden, the AI gets so good, there is no need for the accountant anymore. Right. Cause at some point it can do everything the accountant can do. gets smart enough about all the edge cases and all of that. It can just do the job. And once that happens, their productivity and their profitability, their ROI go from, you know, a thousand fold what it is today to zero overnight. It won’t be a slow curve. It’s going to be a drop.

 

Scott Weiner (44:31.456)

And it’s not going to happen across all industries all at once. It’s not going to happen like that at all. And it may not happen in many industries. you know, things where there’s a of human interaction and stuff. Who knows if that’s really going to ever, you know, people are going to accept robots in that way. I don’t know. I could imagine it, but I don’t know. And so, but in a lot of areas you’re going to have this it’s, Hey, everything’s wonderful and better and better and better. then just, it’s not needed anymore. So then what? And, and that, that’s where I

 

think that we have a problem. It’s not the slow curve, it’s the fast drop. That’s really the risk.

 

Yeah, exactly. Okay. That’s enough stoner talk for now. Let’s get into the nuts and bolts of your predictions from last year.

Show more...