Latest / The Adults In The AI Room / Earley AI Podcast - Episode #61 Featuring Don Gossen: Advanced AI and Blockchain Integration for Decentralized Energy Networks

Transcript

Earley AI Podcast - Episode #61 Featuring Don Gossen: Advanced AI and Blockchain Integration for Decentralized Energy Networks · The Adults In The AI Room

• 0:00
  SPEAKER_00: Welcome to the Early AI podcast.
• 0:03
  SPEAKER_00: My name is Seth Early.
• 0:04
  SPEAKER_00: And I'm Chris Featherstone.
• 0:06
  SPEAKER_00: Really excited to introduce our guests for today to discuss some of the misconceptions around AI, some of the ways that we can think about AI and Web3.0, how blockchain and crypto's uh technologies fit into getting the greatest value from AI and from uh machine learning and from agents, and really talking about some of the challenges that are facing implementations of AI and how assigning value to those interactions and those transactions can really start to release the real potential of AI and agents and really incentivize exchange of value and exchange of information across the ecosystem and really to start to turbocharge how people are implementing and getting value from AI.
• 0:50
  SPEAKER_00: So our guest today is a seasoned technologist.
• 0:53
  SPEAKER_00: He has 20 over 20 years of experience in machine learning and IT consulting.
• 0:57
  SPEAKER_00: He's worked for some of the largest companies in the world, like um HSBC, L'Oreal, uh AXA.
• 1:03
  SPEAKER_00: He not only has he founded cutting edge, uh co-founded cutting-edge AI projects such as Ocean Protocol, but he's now at the helm of Nevermind, a company revolutionizing the way AI agents interact with financial systems.
• 1:16
  SPEAKER_00: Don Gosson, welcome to the show.
• 1:19
  SPEAKER_00: Thanks for having me, and happy new year.
• 1:21
  SPEAKER_00: Oh happy new year to you.
• 1:23
  SPEAKER_00: I know that's uh it's already 2025, and I have not uh made the mistake of writing 2024 on any of my checks.
• 1:29
  SPEAKER_03: So well, lucky you.
• 1:30
  SPEAKER_03: That's an example.
• 1:31
  SPEAKER_03: I don't write checks, but I'm definitely writing 2024 on stuff.
• 1:35
  SPEAKER_00: I actually do still write some checks.
• 1:37
  SPEAKER_00: There's like the dog sitter, you know, the dog walker.
• 1:39
  SPEAKER_00: He uh he gets a check, uh, you know, every Friday.
• 1:42
  SPEAKER_00: He gets a check and I make it out.
• 1:43
  SPEAKER_00: So I still do a lot of.
• 1:44
  SPEAKER_03: Well, pretty soon you will have uh an AI that will be uh paying your your your dog walking robot via Nevermind.
• 1:53
  SPEAKER_00: Well, that's there, there you go.
• 1:55
  SPEAKER_00: That's really funny.
• 1:56
  SPEAKER_00: When you start thinking about what's in store for us in 2025, we think there's a lot of really interesting things happening.
• 2:02
  SPEAKER_00: So I wanted to start off with conceptions that you run into about AI, giving your background, you know, what do you see as some of the biggest misconceptions that people have around AI and machine learning, particularly regarding integration with their systems with regard to information architecture?
• 2:20
  SPEAKER_00: But give me your take on what you're seeing in the industry.
• 2:24
  SPEAKER_03: Yeah, I mean, so like it's funny, right?
• 2:27
  SPEAKER_03: Because people are I would say strange funny.
• 2:31
  SPEAKER_03: I mean, coming out and being in ML for as long as I have, it's it's also kind of haha funny that you know this hasn't actually been recognized yet.
• 2:41
  SPEAKER_03: But so there's this perception that what's out there now, AI is new, right?
• 2:47
  SPEAKER_03: When in actual fact, AI, this is like, you know, the synthesis of arguably like 70, 80 years worth of work, right?
• 2:57
  SPEAKER_03: It's a culmination of all of this.
• 2:59
  SPEAKER_03: So, you know, I find that to be both shocking and humorous at the same time that people have that perception.
• 3:06
  SPEAKER_03: It's the the the other sort of shocking and humorous aspect of this is it's still the same problems consistently, right?
• 3:15
  SPEAKER_03: Your model and the inference, for example, is only as good as the data that goes into it, right?
• 3:22
  SPEAKER_03: Garbage in, garbage out.
• 3:23
  SPEAKER_03: Um, so garbage pertaining to both obviously the data uh from a training and then a contextual point of view for inference, um, as well as the algorithm, right?
• 3:35
  SPEAKER_03: And the sophistication of that.
• 3:37
  SPEAKER_03: And so I think what is relatively new is the model or algorithm architecture, right?
• 3:45
  SPEAKER_03: Transformers being, you know, kind of the the um the most popular in-demand architecture at the moment.
• 3:53
  SPEAKER_03: Um, you know, a lesser but not forgotten architecture would be like the diffusion models for images and that sort of thing.
• 4:01
  SPEAKER_03: Well, we've seen it.
• 4:02
  SPEAKER_00: Retrain transformer.
• 4:03
  SPEAKER_00: That's the right exactly transformer.
• 4:05
  SPEAKER_00: Meaning trained it on a set of data, and then we are able to transform other data and other use cases with that same set of training and the same uh uh weightings and so on.
• 4:18
  SPEAKER_03: Right.
• 4:18
  SPEAKER_03: It's it's uh in effect um a way of of compressing the data set into like fine slivers of filters for lack of a better term, and then compressing all of those together so you can um you can reconstitute the input information without having the same size of data set that exists on on the input side.
• 4:45
  SPEAKER_00: Yeah, and and in simple terms, it's transforming you know one language into another, it's transforming a question into an answer, it's transforming one piece of data into another piece of data, but it's really basing it on the learnings from a very large data set and then applying that to a smaller data set.
• 5:02
  SPEAKER_03: Yeah, what it's doing is I mean, if you look not to get like overly technical and um and we are going to exhaust my technical expertise if we go super low level, but it's really about establishing the connections effectively between words and then uh um, you know, through a vector math, being able to identify what's called the next best word in the in the case of transformers, right?
• 5:28
  SPEAKER_03: So it's it's this yeah, a proximity algorithm that basically says, like, here's this word, the next best word, based on the overarching context of whatever you're you're trying to um resolve to, the next closest word in the the context that you're trying to articulate is this, right?
• 5:48
  SPEAKER_03: And so these the that that is in effect with the latent space or this like soup of zeros and ones um actually represents.
• 5:57
  SPEAKER_00: And you know what's so interesting is we can say it in that way, and we can say it's a prediction mechanism, and it's just predicting a likelihood of next the next word in the process in the response, but it gets so complex and it gets so interesting, it's so sophisticated.
• 6:13
  SPEAKER_00: And when you start thinking about how some of these models are starting to uh be Machivellan and can be deceptive in certain contexts, where you're looking at theory of mind, it's really hard to square that with, oh, this is just a predictive mechanism, right?
• 6:27
  SPEAKER_00: I mean, it the the manifestation of this and the and the the characteristics and the power and the complexity, I mean, it's just mind-boggling and astounding, but it breaks down to that simple explanation you provided.
• 6:39
  SPEAKER_00: But it's really hard to kind of get your mind around the fact that this is just a prediction mechanism.
• 6:44
  SPEAKER_03: So I I mean it's in effect, it's less prediction and more recall, right?
• 6:49
  SPEAKER_03: It's just um uh the the information that it's recalling is just far greater than what you and I can recall, right?
• 7:00
  SPEAKER_03: It's corpus of information that has been compressed and can then subsequently be recalled into next best action or next best word or whatever is far greater than any single human, right?
• 7:12
  SPEAKER_03: So, you know, it is is predictive in the sense that um it's more, it's less, I would argue, predictive and just non-deterministic, right?
• 7:24
  SPEAKER_03: You don't know exactly how it's going to respond, though.
• 7:30
  SPEAKER_03: If you know the entire corpus of information and and have an overview of that of the topology of the model, you you can discern what course of action it could take, right?
• 7:42
  SPEAKER_03: And this is kind of like this new idea of reasoning and the ability to let a model think.
• 7:47
  SPEAKER_03: It's just like taking different permutations of this corpus of information and coming up with new outputs.
• 7:53
  SPEAKER_00: Anyway, Chris, you got a comment or question?
• 7:56
  SPEAKER_01: I was gonna double-click on something you said.
• 7:58
  SPEAKER_01: So, you know, just in terms of it being, you know, most laughable, right?
• 8:02
  SPEAKER_01: That this is quote new, right?
• 8:05
  SPEAKER_01: And um, I'd love to get your take too on because where I see a lot of people missing this is also knowing how to plan for it commercially and business-wise.
• 8:14
  SPEAKER_01: It's one thing to understand the deep, you know, that what it does and what it's for, but I think there's also a huge miss on it being new, and people do not know how to plan for it andor budget for it or anything like that.
• 8:26
  SPEAKER_01: What are your thoughts?
• 8:27
  SPEAKER_01: And you know, have you seen those types of misses as well?
• 8:30
  SPEAKER_03: Yeah, I mean, admittedly, I'm less on the enterprise side, right?
• 8:35
  SPEAKER_03: And looking, you know, we're out on the edge.
• 8:38
  SPEAKER_03: We're trying to get AI agents, we're trying to enable AI agents to pay and get paid, right?
• 8:42
  SPEAKER_03: And to say there's a lot of corporates that are focused on this at the moment, it, you know, I'd be um lying.
• 8:51
  SPEAKER_01: Uh it's more nobody focused on what you're doing for sure.
• 8:54
  SPEAKER_03: That's yeah.
• 8:55
  SPEAKER_01: Well, hopefully more than just nobody.
• 8:57
  SPEAKER_03: Yeah, I mean, it's it is definitely, you know, um, it is definitely new and novel and fringe.
• 9:04
  SPEAKER_03: I think what we're gonna see is is standard issue new technology adoption curve, right?
• 9:10
  SPEAKER_03: You know, we're definitely headed into the proverbial Gartner trough of disillusionment, where what we have right now is likely a bunch of CIO and CTO offices ramping up on AI skills, which means bringing in a whole bunch of consultants, like I used to do, to try and tell them what they're gonna do, right?
• 9:35
  SPEAKER_03: And likely, you know, overshooting budget and underperforming on the output.
• 9:40
  SPEAKER_03: A lot of this, look, I mean, that's not to fault necessarily the process, because that's how it works, right?
• 9:46
  SPEAKER_03: You you um I think you distill what's useful as an organization through the process of trial and error.
• 9:54
  SPEAKER_03: But I I do think there's still there needs to be recognition of sort of the the historical issue of you know proper data management, model management, access control, all of these things that are sort of known requirements to make your analytical systems perform or performant.
• 10:17
  SPEAKER_03: Those they're still requirements, and they're probably more important now because of the speed with which a model could be non-deterministic, could hallucinate.
• 10:27
  SPEAKER_03: Um, right?
• 10:29
  SPEAKER_03: What these architectures do is it makes it more familiar for people to leverage.
• 10:35
  SPEAKER_03: The interfacing has come, you know, the the the sophistication or the complexity of the interfacing has been drastically simplified.
• 10:44
  SPEAKER_03: Anybody in an organization can really interface with like back-end technology in natural language, and that it but that at the same time exposes anybody in the organization to information that maybe they shouldn't be privy to, right?
• 11:02
  SPEAKER_03: So there are all of these concerns and thoughts that need to go into the overarching process.
• 11:07
  SPEAKER_03: Now, what's very interesting is the capability, when you introduce reasoning over these corpuses of information, enable you to effectively create, like double down on that interface, right, and leverage the model further than models or algorithms have been leverageable in the past, which quite frankly is is, you know, it's likely going to lead to a lot of displacement, you know, because these roles that were, you know, sort of anchored to repetitive knowledge-based work, right?
• 11:42
  SPEAKER_03: You're an expert in something.
• 11:44
  SPEAKER_03: Reporting is like one of the likely candidates for all this.
• 11:48
  SPEAKER_03: You know, these these tasks, when wrapped in like business process optimization technology and that sort of thing, workflow management, um, you're just gonna be able to, you know, you're gonna be able to replicate that functionality quite easily.
• 12:02
  SPEAKER_01: The one thing you said too, which I think is interesting and I think also misses the trough of disillusionment, is all the cool innovations or really valuable innovations that come from each of those points.
• 12:11
  SPEAKER_01: So, like you guys, early innovators, who's innovating in the trough.
• 12:15
  SPEAKER_01: There's that almost that mirror that needs to be seen on where the laggards and innovation comes there, innovation comes at all those different points.
• 12:21
  SPEAKER_01: I think people miss too.
• 12:23
  SPEAKER_03: Yeah, a lot of what goes on in that that trough of disillusionment is like the industrialization process, right?
• 12:30
  SPEAKER_03: Like the true application of traditionally what would have been referred to as MLOps is now being applied to AI in the form of observability tooling, right?
• 12:40
  SPEAKER_03: And really, if I were to kind of extend what we're doing, we're doing observability on AI, you know, what's actually running, but we're we're we're we're accounting for it.
• 12:52
  SPEAKER_03: We're strapping an accounting module to this and saying, okay, this is what was actually used from like a GPU cycle point of view, tokens in, tokens out, and then translating that into a settlement cost for the requester.
• 13:04
  SPEAKER_03: So yes, it's the same process that's already always existed.
• 13:09
  SPEAKER_03: What's nice is that now, I mean, okay, this is gonna, depending on who your usership is, this is gonna sound a bit crude and I'm not trying to be distasteful, but like I look at this as very positive, right?
• 13:25
  SPEAKER_03: We're kind of removing the weakest link from the process, which is in a lot of cases you and I, and replacing it with this thing that, okay, you argue, well, they're non-deterministic, they can hallucinate.
• 13:39
  SPEAKER_03: Yeah, but so are humans.
• 13:40
  SPEAKER_03: Humans are non-deterministic.
• 13:42
  SPEAKER_03: We hallucinate, we sit there and daydream, right?
• 13:44
  SPEAKER_03: And forget to like click something during a workful process.
• 13:46
  SPEAKER_03: What do we do as an organization?
• 13:48
  SPEAKER_03: Well, we wrap these processes in processes, right?
• 13:52
  SPEAKER_03: To prevent and mitigate against that, you know, daydreaming risk, for example.
• 13:57
  SPEAKER_03: Wait till wait till AI.
• 14:02
  SPEAKER_03: It's just that this is gonna lead to a lot of displacement and and discomfort for people, right?
• 14:10
  SPEAKER_03: It's gonna lead to a lot of discomfort.
• 14:12
  SPEAKER_03: How so for people?
• 14:13
  SPEAKER_03: Well, it's people people are gonna lose their jobs in all likelihood and be replaced by AIs.
• 14:19
  SPEAKER_03: But tell me, but tell me where so it kind of I mean the head of Klarna, just as like anecdotal evidence, two Sundays ago came out and was like I'm the CEO, my job is toast in a few years.
• 14:30
  SPEAKER_00: Who said that?
• 14:31
  SPEAKER_03: Who said that?
• 14:32
  SPEAKER_03: That was the CEO founder of Klarna, the payments company that lets you pay in installments, right?
• 14:37
  SPEAKER_03: They they're one of the leaders, I think, in in AI adoption and agent adoption, right?
• 14:44
  SPEAKER_03: They've offset a significant headcount internally with these fully automated processes run by agents.
• 14:53
  SPEAKER_03: You got Reed Hoffman coming out and saying, like, Salesforce isn't hiring any more engineers, right?
• 14:59
  SPEAKER_03: We're just gonna we're gonna build agents.
• 15:01
  SPEAKER_03: Now, some of this is marketing, but a lot of it's true, right?
• 15:07
  SPEAKER_03: And I gotta say, like, okay, we just did a round, right?
• 15:10
  SPEAKER_03: We're out there hiring, and you can see the displacement.
• 15:14
  SPEAKER_03: It's not necessarily directly related to AI at the moment.
• 15:19
  SPEAKER_03: It's more, you know, uh, it's more sort of companies, tech companies tightening their belts, but there's a lot more engineers out there than they have than there have been in the past.
• 15:30
  SPEAKER_03: Like I've been an entrepreneur for almost a decade, and this is probably the best time to be hiring that I've had with respect to kind of the the the skills that are out there to be had.
• 15:43
  SPEAKER_03: And I think that that is just gonna continue to grow and depending on your perspective, exacerbate.
• 15:51
  SPEAKER_00: So tell me more about um you know the work you're doing at the intersection of AI and Web3.0 and how are you defining Web3.0, and then what does it really mean from a practical perspective to be able to uh ascribe value to certain pieces of uh uh of an ML workflow?
• 16:08
  SPEAKER_00: So, you know, one of the things you were saying in our prep call is that right now it's metered, right?
• 16:13
  SPEAKER_00: It's it's not about value, it's about usage.
• 16:15
  SPEAKER_00: So are you using it?
• 16:16
  SPEAKER_00: How many tokens are you using it, using tokens in tokens out?
• 16:19
  SPEAKER_00: Uh let's you know, look at your quota, let's look at your max, let's you know, pay by this metric, and then it's about usage, but you're talking about ascribing value to insights or transactions or whatever.
• 16:31
  SPEAKER_00: So, how how are you kind of beginning to do that?
• 16:34
  SPEAKER_00: And tell me more about I know there's a bunch of questions in there.
• 16:37
  SPEAKER_00: So, how are you defining web and then how looking at the intersection?
• 16:42
  SPEAKER_03: So, I I mean web three is it it's crypto blockchain tech, right?
• 16:47
  SPEAKER_03: So, at the risk of of turning the next 30 minutes into a 30-hour conversation, I'll try and condense you know the history of all of this.
• 16:58
  SPEAKER_03: But as you mentioned, I've been in the ML space for 20 years.
• 17:04
  SPEAKER_03: As we've been discussing, you you need to understand the inputs of an analytical solution, right?
• 17:12
  SPEAKER_03: So the data sets, the algorithms, and then everything that all of the actors that are sort of touching this pipeline as it moves towards in enterprise landscape, we call it the actionable insight, right?
• 17:26
  SPEAKER_03: AI ProLont, you call it the inference.
• 17:28
  SPEAKER_03: So, really, you're trying to answer four basic questions.
• 17:31
  SPEAKER_03: Where's the data model coming from?
• 17:33
  SPEAKER_03: Where's it going?
• 17:34
  SPEAKER_03: Who's using it, and what are they doing with it, right?
• 17:36
  SPEAKER_03: The inability to answer just one of those questions undermines the entire output, right?
• 17:43
  SPEAKER_03: Like there's nothing worse in as someone on the analytics side, if you go into a meeting, you're a data scientist or whatever, you go into a meeting and you present information as an actionable insight and as kind of as fact.
• 17:57
  SPEAKER_03: And then somebody's able to undermine that by um, you know, pushing back and saying, where did you get that?
• 18:03
  SPEAKER_03: Where's the data from?
• 18:04
  SPEAKER_03: Where's the model from?
• 18:05
  SPEAKER_03: How did you leverage this?
• 18:06
  SPEAKER_03: Where did you deploy it?
• 18:07
  SPEAKER_03: Right.
• 18:08
  SPEAKER_03: And so you need to have high fidelity on each of those questions.
• 18:13
  SPEAKER_03: The problem is that existing contemporary infrastructure that enables you to do that is dog shit.
• 18:20
  SPEAKER_03: It's not very good, right?
• 18:21
  SPEAKER_03: Yeah, it's usually patchwork of that's yeah, it's a it's a super technical IT term, so most of your audience probably won't get it.
• 18:28
  SPEAKER_03: But yeah, that's just the case, right?
• 18:30
  SPEAKER_03: It ends up being like you you you create these bespoke tracking solutions that are very patchwork, right?
• 18:36
  SPEAKER_03: And they don't work very well, and they definitely don't scale, right?
• 18:40
  SPEAKER_03: And so, you know, beating my head against the wall for a long time, and then in 2016, I got introduced to crypto to Web3, and you know, the the light went on, right?
• 18:54
  SPEAKER_03: It was a bit of a eureka that blockchains are very elegant solutions to the provenance problem because that's what we're talking about.
• 19:01
  SPEAKER_03: Understanding the provenance of a model, of the data set or data sets that go into training as well as the outputs and understanding all of the systems that came together to create this output, right?
• 19:15
  SPEAKER_03: And the higher the fidelity on each of those different uh components, the more trustworthy the output, right?
• 19:25
  SPEAKER_03: And then arguably the higher the value.
• 19:27
  SPEAKER_03: So my motivation in getting into this across section, right?
• 19:36
  SPEAKER_03: Merging what are inherently very divergent technologies, right?
• 19:40
  SPEAKER_03: ML AI technology is is very centralized in nature, optimally so, and decentralized technology blockchains are are the exact opposite, right?
• 19:52
  SPEAKER_03: So merging not just technologies but also paradigms creates all sorts of conflicts.
• 19:59
  SPEAKER_03: And so spent close 10 years trying to manage these technologies kind of butting up against one another.
• 20:08
  SPEAKER_03: And what we realized so, so payments, settlement, attribution in particular, understanding that Provence history to a fine degree of detail to the point where you could actually provide rewards or payments at the same level of granularity.
• 20:26
  SPEAKER_03: Payments or rewards were always a big part of this, right?
• 20:29
  SPEAKER_03: So you create an output, you capitalize offset output, and then you reward all of the contributors that help create that output, right?
• 20:39
  SPEAKER_03: And so this payments and settlement piece is has always been parcel of the solution.
• 20:45
  SPEAKER_03: What we realized 24 months ago is actually it's the most critical piece.
• 20:49
  SPEAKER_03: You can have all of the reward attribution technology that you want, but if you can't capture the value created in the last mile, there's nothing to actually reward or pay out, right?
• 21:02
  SPEAKER_03: Exactly.
• 21:02
  SPEAKER_03: And so we shifted focus into specifically just drilling in to this payments piece.
• 21:10
  SPEAKER_03: The the catalyst for this was the diffusion models initially, right?
• 21:14
  SPEAKER_03: That stability AI released, and then catalyzed by Chat GPT from OpenAI that really exposed way more people to the capabilities and unlocked this easier interface, right?
• 21:28
  SPEAKER_03: And it's led to this explosion in adoption.
• 21:30
  SPEAKER_03: And what we saw back then were builders at that point, they were like rag tuning, so retrieval augmented generating models to do specific things.
• 21:39
  SPEAKER_03: And people would release them into the wild through different platforms, and they'd get like hundreds of thousands of users that would come in.
• 21:47
  SPEAKER_03: Right.
• 21:49
  SPEAKER_03: And these builders would all have to underwrite the cost of that, right?
• 21:53
  SPEAKER_03: Right.
• 21:54
  SPEAKER_03: And so what we saw was like builders trying to strap Stripe onto this inner.
• 22:00
  SPEAKER_03: interface and charge like five bucks a month, right?
• 22:03
  SPEAKER_03: And but you'd still get like a power user that would come through and consume all of it.
• 22:08
  SPEAKER_03: Because at this point, if you recall ChatGPT in particular was like quite limited.
• 22:13
  SPEAKER_03: If you hit your quota of tokens, boom, your service went down, right?
• 22:17
  SPEAKER_03: Like they they froze, right?
• 22:19
  SPEAKER_03: They throttled you to the point where you couldn't use the service anymore.
• 22:22
  SPEAKER_03: So this was happening to these builders.
• 22:24
  SPEAKER_03: And so they were spinning up new connections and things like the more sophisticated ones.
• 22:28
  SPEAKER_03: Anyway, they were having to underwrite the cost of all this.
• 22:30
  SPEAKER_03: And so we saw effectively these cottage industry businesses go out of business.
• 22:35
  SPEAKER_03: Right.
• 22:35
  SPEAKER_03: Because you guys were footing the bill and we just thought that this was stupid.
• 22:39
  SPEAKER_03: And we realized that in this sort of journey to get to this laser focus on payments, we realized we'd actually built the solution.
• 22:47
  SPEAKER_03: And this goes back to you know you alluded to this but it's taking the metered cost of running a given AI service to take a slight detour.
• 22:58
  SPEAKER_03: I like to describe an agent kind of like a platform right it's it's the sum of its parts right it's a compilation of different tools namely the models right so it's got like access to a particular series of models GPT series but like from open AI but it could be from different vendors etc but then you also have you know the the the orchestration capability from like a lang chain or land graph or what have you you you might have contextual data coming from a vector db like pinecone or whatever and all of these things get put together they all have associated costs for accessing and using right that's the observability piece right so when you run an agent you want to have like access into the a perspective into like the the ML ops that's going on with your agent with these probably third party services because you're probably not running this stuff yourself.
• 23:56
  SPEAKER_03: You've compiled it into the form factor of an agent and so you've got GPU cycles or tokens in tokens out with these different models and these different models have different requirements right I could be a simple user Seth you're a power user we can use the same agent that Chris built right but we're completely divergent personas right I submit my simple request to Chris's agent that gets decomposed and rooted to a model that can handle the lack of like my unsophisticated requests.
• 24:27
  SPEAKER_03: Let's say it goes to GPT 3.
• 24:29
  SPEAKER_03: You on the other hand you're you submit a multimodal request to Chris's agent that's got to go to a model that can handle the multimodality or GPT 4 or 4.5 what have you the spread right the cost difference between instantiating GPT 3 and GPT 4 4.5 is like an order of magnitude or two a difference right exactly and so taking that variability and accounting for it and translating that into a settlement cost is non-trivial.
• 25:03
  SPEAKER_00: Of course of course and so what does it take to start to do that?
• 25:06
  SPEAKER_03: I mean what does it take in to deconstruct that and to understand that it first it takes recognition of the problem that that's actually something and then second it takes you know selecting the tools that are available to you at that given period of time in order to enable you know the solution that you've identified for that particular problem.
• 25:29
  SPEAKER_03: Right.
• 25:29
  SPEAKER_03: And so we look at this again in in sort of the larger picture of what's going on in an analytical workflow right using a blockchain to understand these compartmental pieces that are going in and then decompose those and understand how the result set is actually being generated and then provided as an inference all of the different counterparties that go into that potentially right or even just the different solutions the the tool stacks that go into creating this form factor that is the agent accounting for all of that and then passing that off into uh onto the requester doing that in a pure digital manner just seems to make sense.
• 26:16
  SPEAKER_03: Yes and blockchains are by and far the best solution for both of those functions both the providential piece right establishing that integrity and identifying sort of the attribution component but then also on the payment and settlement side.
• 26:34
  SPEAKER_00: Right because you can invoke contracts to control the access and all of this right so what is a simple use case so where do you begin this process because again I I imagine there's a lot of complexity to it and you know what are your objectives in terms of how to get this to market like where do you see the low-hanging fruit for you?
• 26:55
  SPEAKER_01: What do you need in place right?
• 26:56
  SPEAKER_01: What do you need in place also?
• 26:58
  SPEAKER_03: Yeah yeah well first off you need like the mindset and you want to adopt this that goes without saying right so you got to find those counterparties that are going to be open to this from the get-go because look the honest answer is a lot of what's going on in the agentic ai side of things is being underwritten largely by VC.
• 27:18
  SPEAKER_03: So that whole like issue of of I'm gonna effectively I'm gonna run these operations right I'm just gonna charge you like a set amount I don't really care if that 20 bucks a month for each user is actually costed out properly because I got a bunch of VC dollars sitting behind me that I'm just gonna underwrite if if I go over.
• 27:41
  SPEAKER_03: So that's part of what's going on.
• 27:43
  SPEAKER_03: I will again sort of taking a side step here this stuff's moving so quickly that these services our perspective is that these a agentic services are going to commoditize faster than prior generations of technologies and so it behooves us to have this mechanism in place ASAP so as we push to war I where you hear it already right you got like Sam Altman and stuff like that talking about the cost of generating an inference is going to get pushed to zero right and so where you're gonna eke out a business is in the margin.
• 28:27
  SPEAKER_03: That requires having like incredible insight and accuracy into your cost center your opex in order to eke out that margin relative to your competitors right so doing this like accounting on the observable piece and translating that into settlement as as quickly near real time as possible is going to become in our opinion the primary driver for this but to get back to kind of where is this applicable it's it's quite interesting like it's actually applicable across the board right we're seeing that and and this is why the like the earlier part of the conversation was relevant right because it's going to displace a lot and so just for kind of sake of example one of the use cases we worked on was this energy smart energy grid management balloon right this is something that's been talked about for like 10 going on 50 right IoT is gonna change the name of the game and it turns out these networks plus the information that needs to be consolidated to make the grid actually smart lacked the level of model sophistication um not discreetly right historically you could create a neural network that could perform a function really well the issue was it only worked for that one use case right right right right it wasn't really scalable right you could do some transfer learning but it wasn't really scalable beyond like the that use case and and and uh you know use cases within proximity now what you have are agents that kind of just again they have this broad knowledge base that can do quite a bit so really what you're tweaking is like the tool selection and the reasoning capability which turns out is a lot easier than you know we thought because of the way that these things have been architected.
• 30:30
  SPEAKER_03: And so where we're at now so we worked with um a couple other web three crypto organizations that are in what's called the um decentralized physical infrastructure network space right IoT um in historical parlance so the layer one the blockchain that we worked with is called peak paq and on top of that is a company called Combinder which is building a deep in a solution on top of peak network combinder they're based out of the the dock region so Germany Austria Switzerland in the EU and within the EU if you produce if you're a retail producer of energy you can actually sell back into the grid unlike in North America or the US in particular you can't yeah in I guess in certain states you can but in broadly speaking in the US you cannot sell back into the into the grid in Europe you can and in Germany and actually in the dock region solar panels have become so popular people are building like fences instead of using mesh because it's cheaper to buy solar panels than it is to buy like the mesh or wood to make a fence and it you now have to pay to put the energy into the grid so they have this like massive overproduction.
• 31:49
  SPEAKER_03: But combiner is looking at how do we create these like local area networks from a retail point of view where we have all of this production from solar panels wind turbines that like I might be producing.
• 32:02
  SPEAKER_03: But my neighbor and my neighbor wants to like hypothecate that they want it they want to consume my energy from me they don't want to consume it from like a natural gas power plant 500 kilometers away combiners bringing together the information the data sets and ultimately the network that's like the physical network that's deep in that's going to enable that in conjunction peak.
• 32:23
  SPEAKER_03: So we worked with them we strapped an LLM based model an agent to an HVAC system an air conditioning unit that agent goes out and acquires the data feed from Combiner through their API Combiner is using Nevermind to enable that payment and the payment is done in the underlying DPIN's crypto called peak so you're not paying USDC you're not paying like a stable coin you're not paying ether bitcoin you're paying that native network's token well you aren't your AI is that's strapped to your air conditioning unit and then this in this use case the air conditioning the the agent is is tasked with not determining when it should turn on and on because of optimal price but when it should turn on and off based on overconsumption of energy.
• 33:19
  SPEAKER_03: So it's starting to take back some of that energy for your neighbor turn on so that like your neighbor doesn't have to pay to push that into the grid for example right so you're creating this holistic benefit for your your local area network your your local ecosystem your friends and family your neighbors right this is the type of sophistication that just wasn't available five years ago right mark andreessen and Ben Horowitz address this and you know they have like the Ben and Mark podcast or Mark and Ben con podcast whatever it's called mid-October they actually talked about this exact use case um about leveraging the ability to now build truly smart energy grid management solutions okay by combining blockchain technology and you know these new architectures for for AI.
• 34:16
  SPEAKER_00: So that sounds like a great use case that can be um uh scaled because it's a very clear process it's very unambiguous in terms of the outcome and you can ascribe you know hard dollar value to that transaction right so so this is so those are the types of things that you want to start thinking about in terms of how you can monetize AI it has to be eatable it has to be clear it has to have an unambiguous outcome it has to have clear value right and they're not necessarily like it's not the first use case most people would think about when they think about AI right most people when they think about AI and AI agents is like oh I'm gonna have an assistant and what's that thing going to do for me?
• 35:00
  SPEAKER_03: Oh it's gonna like book my flights and hotels and stuff.
• 35:05
  SPEAKER_03: It's like no what it's probably gonna do is make your house transactional make your car transactional and and like now we get into kind of the esoteric and fun domain because we were talking about this before where you got like the robot dog walking your your pets and like the currency there is not denominated in dollars it's like pets and treats right and so like where it gets kind of esoteric but also super interesting is like what other actors what other life forms can be like merged into economics and finance now because we're moving away from just the stores of value that the three of us actually want and into stores of value that like maybe our pets want definitely what these AIs want.
• 35:59
  SPEAKER_01: The irony of this is everybody worries about what jobs am I going to lose but everybody uses and tries to target to your point AI towards replacing these goddamn jobs.
• 36:09
  SPEAKER_01: Yeah the part of this like you're saying is there's so much depth to the use cases that people aren't thinking about it in multidimensional forms that we're missing the picture on what a lot of these are well I mean you guys aren't no but but and and look the reality is like like I'm not smarter than the next guy I've just had the luxury of being able to sit here and think about this for 10 years.
• 36:32
  SPEAKER_03: Most other people have not so imagine being unshackled from this constraint of your day-to-day job right and the stresses and shit like that and being able to actually think about new stuff right I'm very bullish on the next hundred years I think a like I think we're the best or at least one of the best organisms on this planet at at you know responding to change right evolving and I think we're going to unshackle ourselves and there's going to be an explosion of creativity on a scale that we haven't seen before.
• 37:14
  SPEAKER_03: Like I think I honestly believe we're entering a new renaissance where we're unlocking the next level of of human ingenuity with all of this.
• 37:24
  SPEAKER_03: I do think it's going to be like there there's somewhat going to be a bit of a race here.
• 37:29
  SPEAKER_03: Yeah right but I'm of the opinion that a lot of what is going to end up manifesting in the AI derived economic ecosystem will largely be ambivalent to and vice versa right yeah yeah it's awesome it's awesome what what about challenges what immediate challenges do you see?
• 37:48
  SPEAKER_01: Well I mean we talked about some but like what are some that you're that you're having to think about regulation is it is it data sovereignty is I mean especially in the in the uh dock region and all that the European regions they're super ahead of the curve we'll say on I'm being paranoid paranoid with a lot of this tech right yeah so what do you think it'll be interesting yeah uh on so reg capture regulatory capture of the industry is a is is sort of an existential threat.
• 38:17
  SPEAKER_03: I do take solace in the fact that like we're kind of so far afield at the moment so one of the at risk of sounding crazy to your audience so one of the drivers for us like another you know Seth you asked is like not not so much why blockchain you were asking like you know what is it and what'd you do but there's a there's a reason there is an element and like an ethos attached to why so we look at payments as being the choke point for AI agents you know if um if openai Microsoft Google Facebook you know if they monopolize the means for agents to pay and get paid their ability to shut off one agent becomes an existential threat to all agents right if Microsoft says eh this agent is competing with my line of business poof you no longer have the ability to transact that is an existential risk to all agents right and so a big driver for us is providing that optionality in a decentralized fashion so that these agents always have an option for for paying in and getting paid.
• 39:32
  SPEAKER_03: And so we we do look at this you know in terms of the reg side of things you know one of the here the the funny part haha funny and also kind of strange crazy funny again is the fact that like our focus is on silicon based agent to silicon based agent not carbon based agent to silicon based agent.
• 39:53
  SPEAKER_01: Right we want to see AI agents actually transact right trillions of these things out there all kind of paying or bartering with one another or whatever right in which case what regulations actually apply right right there if these are if these are decoupled from any human and you know you talk about data sovereignty well I'm thinking more about agent sovereignty right if these are fully autonomous sovereign like self-sovereign agents which you know theoretically they could be in the not too distant future we don't have a regulatory landscape for governing right yeah digital governments yeah exactly so I look at it both as a threat but also as an opportunity right go ahead capture all the you know carbon based actor carbon based agent market as you want we're not going after that market we're going after the silicon based agent market it would also then most of that regulation is like you said always when it hits humans so nobody cared about solar until it hit your pocketbook or sorry savings right savings yeah nobody cares about these things until like you said we're good at change but at the same time nobody nothing or people don't change until it becomes painful right yeah that's a good point yeah so yeah it'd be interesting to see if they'd even care until it hits humans and then humans yeah I mean it is gonna hit us and people are gonna care.
• 41:21
  SPEAKER_03: Yeah um it's it's a question of how draconian the regulation becomes and who it favors because the reality is it's never draconian for everyone.
• 41:33
  SPEAKER_03: Right.
• 41:34
  SPEAKER_03: It's it's only draconian for most.
• 41:37
  SPEAKER_03: Yeah and then you get this small few that benefit from the fact that they were like grandfathered in or whatever right and so that's what needs to be avoided and that's also something that we're trying to provide the optionality to mitigate against.
• 41:52
  SPEAKER_03: So that that would definitely be a risk.
• 41:54
  SPEAKER_03: I mean the the other risk you know you asked about what do we need we need agents right we just need more agents and they're coming online all the time but yeah the the reality is we're at you know arguably tens of thousands of productive agents we have a trajectory of trillions so there's yeah pretty big gap in between the two I was just saying a registry of these kind of things too right so you can find a yeah I mean that's part of that's part of what we're building right is is like a registry or catalog right that's part of this protocol in order to actually transact you got to know what's out there and so it's it's less marketplacey and more Craigslisty right now but yeah for an yeah I mean honestly I don't want to build a marketplace let somebody else do that I'll just build the you know the components that feed into that marketplaces are yeah what's your next target where do you want to go um next good question okay so we're working on I'm not I'm gonna leave the names out of this um it's it's somewhat clandestine at the moment but there are these you utilizable agents like agents with true utility coming out right um they do things uh mostly on the sales side right so we're thinking like developing listings of of of like potential customers and email campaigns um you know CRM management cold calling or or like following up with actual phone calls and and that sort of thing and so what we're we're eating our own dog food internally and what we really want to show is heterogeneous agents agents from from different third party vendors that can be used and choreographed orchestrated or coordinated amongst one another and where reward More payment can actually be facilitated.
• 44:03
  SPEAKER_03: So that's what we're working on presently.
• 44:07
  SPEAKER_03: Industrializing our artificial intelligence go to market with agentic AI across heterogeneous agents and demonstrating the capacity of these things, but also the commerce that takes place between them.
• 44:23
  SPEAKER_03: Right.
• 44:24
  SPEAKER_03: Because in our opinion, you know, AI agents are the vassal, but the category is AI commerce, right?
• 44:33
  SPEAKER_03: This new agent that's going to be commercial in nature, that's now entering the commercial landscape.
• 44:41
  SPEAKER_03: It's no longer the three of us on the internet anymore.
• 44:44
  SPEAKER_03: It's also AI agents that are going to take advantage of the dead internet.
• 44:50
  SPEAKER_00: Yeah.
• 44:50
  SPEAKER_00: Very exciting.
• 44:51
  SPEAKER_00: Very exciting.
• 44:52
  SPEAKER_00: So tell us uh what do you do for fun?
• 44:54
  SPEAKER_00: What do you do outside of work?
• 44:56
  SPEAKER_00: Besides work.
• 44:57
  SPEAKER_03: Yeah, well, like to move and travel.
• 45:00
  SPEAKER_03: I've lived on every continent.
• 45:02
  SPEAKER_03: Yeah.
• 45:03
  SPEAKER_03: Except Africa and and and Antarctica.
• 45:06
  SPEAKER_03: You know, yeah.
• 45:08
  SPEAKER_03: Traveling's a a big thing for us.
• 45:10
  SPEAKER_03: I've got two little girls, so hanging out with my family.
• 45:13
  SPEAKER_03: And then when I can, I now live in in Lisbon, which is quite sunny and warm.
• 45:19
  SPEAKER_03: Yeah.
• 45:19
  SPEAKER_03: But my my my passion is like um slack country powder skiing.
• 45:25
  SPEAKER_03: Oh wow.
• 45:26
  SPEAKER_03: So ducking the line and you know hitting the slopes.
• 45:30
  SPEAKER_03: Unfortunately, we've we've chosen, I mean, for for that one pastime, we've chosen uh or hobby, we've chosen sun over snow, which, you know, being Canadian, I will choose nine times out of ten.
• 45:43
  SPEAKER_03: But yeah, that's what I like to do.
• 45:45
  SPEAKER_00: Nice.
• 45:45
  SPEAKER_00: Nice.
• 45:46
  SPEAKER_00: Well, this has been a pleasure.
• 45:47
  SPEAKER_00: Where can people find you?
• 45:48
  SPEAKER_00: We'll have uh information in the show notes, but um tell us about uh where they can find you.
• 45:53
  SPEAKER_03: Yeah, um, I think the best place to find us is either follow us on X.
• 45:58
  SPEAKER_03: So it's nevermind N-E-V-E-R-M-I-N-E-D underscore IO.
• 46:04
  SPEAKER_03: Or you can jump into our Discord via our website, nevermind.io, N-E-V-E-R-M-I-N-E-D.io.
• 46:13
  SPEAKER_03: Yeah, we we just did a fundraise.
• 46:15
  SPEAKER_03: We're actively hiring.
• 46:16
  SPEAKER_03: So if anybody wants to reach out and is interested in enabling AI agents to pay and get paid, hit us up.
• 46:22
  SPEAKER_00: And where are you located again?
• 46:23
  SPEAKER_00: You said you're in uh Lisbon.
• 46:25
  SPEAKER_03: I'm in Lisbon in Portugal.
• 46:26
  SPEAKER_03: The team's Europe based, so we're all in either continental Europe or the UK, though soon to be US as well.
• 46:34
  SPEAKER_00: Well, this has been a pleasure.
• 46:35
  SPEAKER_00: Don really enjoyed the conversation.
• 46:38
  SPEAKER_00: Thank you so much for for being with us.
• 46:40
  SPEAKER_01: Thanks for having me.
• 46:41
  SPEAKER_01: And uh thanks.
• 46:43
  SPEAKER_01: It was this has been one of the most interesting ones that we've had, right?
• 46:46
  SPEAKER_01: It's been a gamut, but I again just in order to you know get our our listeners thinking ahead.
• 46:53
  SPEAKER_01: It's been valuable.
• 46:54
  SPEAKER_01: Thank you.
• 46:55
  SPEAKER_00: Um thank you uh to our audience for uh tuning in.
• 46:58
  SPEAKER_00: Uh thanks again, Don, and uh thank you, uh Carolyn, for the work behind the scenes and Chris, and we will see you next time on the next Early AI podcast.
• 47:08
  SPEAKER_00: Have a good day.
• 47:09
  SPEAKER_00: Thank you.
• 47:10
  SPEAKER_02: Visit early.com to find links to the full podcast on all audio platforms to listen on the go.
• 47:16
  SPEAKER_02: Thank you for watching.
• 47:18
  SPEAKER_02: Contact us today at infoearly.com.