Speaker 1Hi, everyone, and welcome to a very special episode of the Dumb Monkey Show. Today, Uwensi is looking directly at you because we are looking at a fantastic tech entrepreneur, really interesting thinker in this space, someone who's worked through the VC pathways, someone who's very good at exits, and someone who understands how AI is disrupting and erupting into our workplaces and into our business ecosystems. Anthony Middlemark, welcome, and thank you for being part of the show today.
Speaker 2Hello, Davina and Amir. How are you? Very good. Thank you so much.
Speaker 3We are really excited to have you on this show.
Speaker 2Thanks. I'm excited to be here. Your audience didn't hear us talking before, but we're already pretty much connected. I think it'll be a good conversation.
Speaker 1It's always exciting when we get interesting thinkers in and for the outtakes that we won't be sharing. We've already had some really interesting conversation, as Anthony points to. Anthony, if we can kick off the conversation, tell me a little bit about how you have seen AI change the landscape in your field in the past, say, six months.
Speaker 2Wow. Okay. This is almost an infinity. Yeah, we're starting with the really tough questions and just working. Yeah, exactly. Let me throw. With the beginning, right? In 2022, 2023, we saw Gen AI start, right? And we first experienced a prompt-based interface, so a chatbot. And it was amazing, even the first models. And the first big disruption was obviously. And I'll explain to you what it was. So way back in the 90s, right, we had search launch, right? Just search. And what search did was it effectively changed the entire way people looked for things. But that was based on keyword debt. So if you were searching for a bicycle, you searched for a bicycle, it looked for sites that had. A lot of mentions of the word bicycle, and that determined relevance. And so intent, human intent, I want a bicycle was linked to keyword debt. And for years and years and years, that keyword debt made Google billions and billions of dollars, right? So that's point one. Then Gen AI launches. Now we can fully. Fully express intent in a natural way. And that intent can be pushed into a large language model. You have semantic alignment. I know you don't want the jargon, but let me just explain to you. Semantic is meaning, right? And so it looks for meaning alignment. So you go red bicycle, red road bike. And then you go red bicycle, red road bike. And then at this price with this gear so all of a sudden your intent your expression of intent is so much more descriptive than it's ever been that means that the you know the 10,000 applicable retailers of that bike now become five um and so so what we saw in the last couple of big fridays black fridays was a huge shift into gen ai search for discovery and research so this has created the need for retailers and financial services and insurance companies and anybody who sells anything to think about how those products and services are described so when a generative engine looks at them it determines do you semantically align to the user's profit this this is one of the biggest things that happen and and and if you look at it next to everything that llms can do now it's it's it's kind of trivial but it's still monumental right and and the companies that i talked to a lot of the companies i talked to they they struggle with this they struggle with this geo generative engine optimization or answer engine optimization and the the best way to understand it is look at your product think about every way anyone has ever bought it create the prompts put them in perplexity and chat gpt and gemini and see what you get back if you don't come up as number one that or or in the list then that's the area where you do not have depth of meaning right and you you've got to fill that you got to fill that gap so that's that's kind of inked to the that's the first part of it and then the second part of it the the answer to your question right then then the second thing is um the way tech services businesses work that all the tech services that that every business in the world consumes they they sell you a service some some go in vertically they sell you a thing that does one thing uh but a lot of them uh sell you services where they go in vertically and then they expand horizontally this this is called systemization right and this has been a major strategy for tech companies forever right so you know i the the more systemized i am in your business the harder it is for you to get me out and replace me with something better you know you gotta you gotta reach the point where you know it's too expensive to operate or you're not your your customer experience is compromised just some other really big metric is being compromised and you pull them out so in in ai now for companies that want to utilize ai in in a materially beneficial way that systemization that position in the customer's technical stack is the biggest contested battle for of all time now they're all fighting for it because what happens is between uh you being able to code your own stuff and you and and and and do it quickly and not even need any experience uh but you also being able to ask for tasks to be done by third-party agents you may not need that stuff anymore and so there's a big war on now to either i get you to keep my stuff and fully integrate it some companies will do that um and uh and and some companies will try to do it on their own and then the the third part the third part answer to your question is everybody sees ai right as money savings you know unit economic advantage reducing the cost to deliver a service to a customer um they they they see it as um you know completing completing tasks automating tasks and and potentially improving customer experience but what they don't look at is things like compound learning where they have some aspect of how they operate that they could compound they could layer learning and they could utilize that learning to provide uh overall value to the entire business including the end customer but what's most important about the idea of compound value is in the next few years investors and acquirers are going to ask you questions about your data architecture and your potential for compound learning right so the these are the three biggest but i literally could go on forever about this about about how how how big uh a disruption it is i think you know we'll look back at this period in time and this will be like an inflection point for for humanity pretty much yeah
Speaker 3yeah i think three great points uh the first point about search actually reminded me of a recent uh interaction that i had with my claude so i in my cloud i've actually not trained it but provided custom instructions to it to be very blatant about things to proactively suggest things and so on so i was sharing something that my lawyer had done for me or shared with me and there was a mistake got it the second time there was another instance that i said my lawyer has said this what do you think so it caught the mistake but said hey this is the second time your lawyer has actually not considered this thing i think it's time for you to switch the lawyers i was like okay uh what would you suggest so it then came up with a bit of list of the lawyers and that would be suitable for me so i was thinking it's when you say uh you need your business needs to be optimized for the prompts that uh people would put to search for it but you would be entering in an era where even if you already have existing customers your ai would be challenging you and saying hey you are buying something from them what about someone else who can provide a better service or a better
Speaker 2product this that is a beautiful example you know i i i talk to companies all the time and i go you know all your customers are using ai for procurement and you know the minute you send your proposal in they're going to put that back into ai too so yeah i mean um yeah the point you just made 100 um you know is is claude being ruthless because two does two mistakes merit merit switching but you know you you kind of do you know pay certain services for you know you have an expectation of certain services to do to do certain things so i mean do you change maybe you do maybe you don't but are you glad that it called it out yes because it forced you to kind of think about what's important to you in the context of paying for that service and and and that point can be applied to everybody who provides a service and ai gives us the
Speaker 1time and the space to do this i mean that being able to check these things critically these things were were constant pain points yes in in business and in workflows but if it's a quick search now that pain point's gone so that opens the field
Speaker 3and well and in the
Speaker 2you go amir now i was
Speaker 3saying in my case i i know that claude can confidently make mistakes as well so i would take it's uh advice uh sort of like i wouldn't sort of fully act on his advice but the normal people who are using chat gpt or claude and they are actually taking some of that advice extremely seriously all the time without even applying their
Speaker 2judgment this is a great point as well so remember when i said semantic alignment so they're so llms wants semantic clarity so when we learned keyword searching it was like one and done bicycle home loan you know new job that that would get you a list and then you would click around so you found the one you want and the the the kind of the the challenge of a large language model is um one and done does not work so there's two strategies you can you can prompt and refine prompt and refine prompt and refine until you get to you get a really good answer or you can write big long massive prompts and uh and you know if if you can connect all the concepts together and and and that also works and you can also uh have control prompts uh which you can include with your prompt about accuracy um you know don't don't don't cite any don't don't create a cited report of something that doesn't exist it it must be true and all the behavior to really refine on stuff that's harder than finding a new bicycle um that that's semantic clarity is it can be very challenging you know um you know normally when you when you buy a complex service like you know getting your house built your your scope of thought around it unless you're a builder is is limited right so um so it's hard to write it's hard to write a prompt that covers that but i think we will see that behavior shift relatively quickly because people will learn that they get way better results with with way better clarity but but we're we're in the shift time you know if you're if you're good at prompting you have a good you have a big advantage
Speaker 1yeah definitely and and steve you i think something that i'm hearing coming up in the in in the messaging and the things that you're saying is this idea of how to how to differentiate how to um be you know provide value to be seen to be providing value how do you be found in a market where everyone has access to to our lms one of the differentiators is of course going to be how you're using them um but another one is going to be are you using just our lms or are you looking at doing some custom ai as well and where's the value lines between that for particularly probably larger businesses but as as bills become more affordable this is going to be an everyone question do we make a custom yeah i am what and where do we use that do we just use yeah
Speaker 2this is the big question you know so there's this concept of sovereignty right where you you run a model for your own benefit i i don't know i'm gonna get jargony again but there's there's a concept called data leakage right so if you're using a commercial llm and you're running your whole business on it you're running you're running a a generative interface and everyone interacts through that your your data and uh all all the loops the feedback loops that need to be completed to buy something or do something that all that all goes that all goes straight to the llm and it and it trains them right um and so that's the potential risk um it it it's very hard not to be jargony if you have stable patterns which uh so some companies have years and years and years of data right and they don't have a big asset but if that data is stable then a very small bit of that data gives the same result as years and years and years it's re-derivable where you have unstable patterns and most businesses do but but you you know you kind of have to be an expert to figure out what they are those those patterns you you want to keep right and so um there's there's two concepts there's two concepts there's two concepts there's two concepts there's two concepts there's two concepts
Speaker 3before you go ahead i i would need you to explain yeah what's the pattern versus unsale button with
Speaker 2an example please okay um i do i do have an example for you um so uh a convenience store right i can think of a convenience chain we won't say any names but we all know the name that comes immediately to mind they have 50 years of trading data right and they think it's super valuable it is not that's stable the patterns repeat um i'm trying not to use the mathematical term for what for for the curve that it represents but it just repeats it's not worth anything and you know what i can just say to an llm uh give me 10 000 records uh based on based on a traditional um uh convenience store and uh and and it'll just do it right and and and i don't i get no extra value from from having the five years just only 10 000 records was enough that that's stable right unstable is things uh where um finance right you you have compliance right so so you you you have to you have to comply in if you have a service that's operating across um uh different geographies you have different uh different types of compliance right and so for the customer those different compliance modes mean different experiences sometimes compliance can make an experience pretty awful as as we all know so an unstable pattern is how do you optimize customer experience for differing from what the customer is going through and that's one of the things that i want to talk a little bit about in the next few minutes and i want to talk a little bit about the compliance model so i still want to acquire the customer i still want them to transact i still want them to do stuff but i have to comply my compliance in japan is different from my compliance in new zealand which is different from my compliance in australia so i want to use ai to manage dynamic customer experience so that even with that compliance in place i still provide a good experience but that that gives me what i need you know turnover basket size transaction volume that's unstable right and that's the value yeah because if if i if i compound the learning on that that means that that's one an asset for my business that's an asset that i can talk to investors and if i'm a public company i can talk to the market about right um but also that layers i get i get more and more and more value and then i can utilize that value and if i if i this is again Becky, but I don't know how you can get away from these concepts. If you separate what the LLM does from memory, you can keep the learning separate and utilize it for your own benefit. Or if you don't think it's a differentiator, you can syndicate that learning to other companies and charge for that. So either way, you get an asset or a new revenue stream. So yeah, but yeah, that's stable and unstable patterns. And I recommend for your listeners, even though it sounds like I'm a data scientist, it's worth using an LLM in the context of your business to kind of figure out where you're at, because if you're not looking at compound learning, context engineering, there's other I won't mention the other ones. But there's a lot of. There's a lot of AI processes, which the objective that they enable is straightforward, but how they do it creates data as well that you can store and also use as compound learning, right? And if you don't know about these things, if you focus only on tactical AI, like automate this. Right now, you kind of open yourself up to people who can build a platform from scratch that does all of that stuff, including the automation, and immediately it becomes kind of hard for you to compete because they can do for five cents what you're doing for $3, you know, and that's a very hard economy to match. And that's the compound learning. And the unit economic difference is what really is the challenge, you know, for AI disruptors, right? That's what they're going to be able to do. If you're a small company in a remote geography, you don't got to worry about it. If you're the only place to go, then just do the GEO bit that we talked about at the first and just be good. You know, don't be LLM. You know, don't be a little bit of a, you know, a little bit of a, you know, a little bit of a, you know, a little bit of a. You know, if you're a small company, you're going to have to be very, very careful about what you're doing. And if, but, but, but, you know, sometimes you don't, you don't really have to do anything because if you're, if you're the good coffee and you're rated well, people, people do find you. What, what we're talking about is businesses that, you know, like accounting firms, financial services companies, banks, you know, big, big enterprises where, you know, they were like, well, we're the only game in town. But, you know, that, that. There's going to be people coming for those factors, you know, so, so it is. Yes. And you would
Speaker 3be aware that Y Combinator, instead of funding just the products, they're saying build rather than building the software for a law firm, you build an agent law firm, and we are going to fund that. So you've got all of these startups who are building AI native companies that are coming for you as well.
Speaker 2Yeah, I mean, what, what would you be doing if you were Y Combinator? Oh, absolutely. Yeah. Absolutely. Yeah. Yes.
Speaker 3Yes. You talked about compounding knowledge, and you also mentioned sovereignty. I have formed a view after a lot of conversation thinking I wanted to test that view is, I felt that sovereignty in terms of us having our own model from the scratch.
Speaker 1When we talk about sovereignty, we're talking about individual sovereignty, national sovereignty.
Speaker 3Yes, it could be so many like definitions, but I'm more talking about like, if you look at it, it's like. If you're an organization in Australia, like having an Australian model, or just having an open course model that you place on your servers or things like that. I found a view that is the race to actually build models and build better models, but there is a is a, if I have to spend my energy and my resources on, it would be actually on building a knowledge that is owned by my organization that the thing that you talk about compounding knowledge or an AI base. So, I think that's a good point. I think that's a good point. And when I'm doing it, I'm saying, it's okay, let's take one step at a time. Let's see the efficiencies. What you are saying is, and this is where I want to go with it as well. And I know the reality. But what you are saying is, this is what's going to happen. Either be prepared, go all in, or, you know, just ride it till the end. It's very interesting to see that approach. I want to hear from you, if when you say this to businesses, what reactions do you get?
Speaker 2Yeah, so my approach has always been like pretty explicit. Like, you know, I look, I know this is going to be challenging. But, you know, how much time do you think you have before you really start to get some real competition? 12 months? Okay, if you think it's 12. If you think it's 12 to 18 months, then we need to get to the facts, right? And I know you're worried about shifting the capability at your business and, you know, general AI maturity and, you know, how positions will change, how role descriptions will change, who may, you know, rise with the company and who may not. But you got to focus on that. Some things which take that long to move and the architecture is one of them that you're going to need a lot of time on, right? So, you know, let's go about architecture. Let's get, let's do the 2030 scenario. Let's figure out what AI first means in the context of your organization. Then let's define the architecture. Then let's give that to you. Then let's give that to you to figure out how you're going to kind of enable, right? Then let's go back and look at operating model and all the other impacts on the business. And we will try to enable you the best we can to handle it. I mean, a lot of times I work with organizational psychologists because the combination of the intellectual shift and the operating model shift is, yeah, a lot. It's a lot. Yeah. Tell us about that, Anthony.
Speaker 1What value does an organizational psychologist bring in this process? Because so much of where we talk about AI ends up just talking about how people think and how we respond and how we absorb information or change or deal with all of those big things.
Speaker 2They just work with, you know, groups within the organization to help them contend with change at this scale. But also, they work with leadership to help them understand how they need to lead through this kind of change. This isn't voluntary change. This isn't the old definition of transformation. Like, you know, we're going to paint some graffiti on the walls and, you know, we're going to shift the cloud. This is a forced, violent, rapid transformation. And, you know, you're competing. I know. I know. I know. There's no end, right? Yeah, that's right. Yeah. You're competing not only with startups, but in many industries, you're competing with the LLMs themselves, right? Because their goal is to monetize token usage. And when you have that kind of power, it's very easy to just, oh, you know, you want low-level research for legal? Boom. There it is. You want health information? You want health information diagnostics? Boom. There it is. So, yeah, there's a lot of pressure. So, yeah, you know, the people part is hard. And even, you know, everyone doesn't know what it means for them. And so that creates a lot of angst as well. But I work with a lot of companies where it is really severely existential. And so, you know, being very explicit is, it's just, you know, you just kind of have to be that way. Because you've got to get up to speed on concepts now. It's just like this podcast. It's like we're resisting jargon, but not understanding the architectural patterns that drive those advantages. That's the disadvantage.
Speaker 3So, yeah. And are you almost saying that understanding architectural pattern is now the role of the leadership rather than some IT people who are there?
Speaker 2You're my favorite person ever. Yes, that's exactly what I'm saying. This is based on what you guys said in the beginning. I don't want to make this too fine a point, but I don't know how you can govern a business without having a rudimentary, understanding of these concepts. Because isn't it your job to focus on customer value and business value? So, yeah, like, I mean, it's not that hard, but you've got to commit to it a bit. But I do think, Amir, in the next few years, you know, we're going to see the most common leaders are conversant in this stuff. Yeah. Conversant. They're not. They're not. Architects. They're not. They're not engineers. They're maybe not even, you know, vibe coding. But they understand. They understand architectural patterns that give advantage and valuation. Right. You have to. I don't know how you can avoid it.
Speaker 3So, I'm coming to a bit of operational question over here because this is how, like, this is where the conversation comes to. It's like, once you have sort of. winster business as well a question that people ask is okay what is the budget that you would do we need to allocate towards this activity or what is the percentage of the budget do we need to allocate towards this yeah have you found an answer to this question
Speaker 2yet no this you got to do the 20 30 thing first right and then you're either looking at like an incremental path which means you want to spend less money but it's going to end up probably costing you more to get there or you're looking at leap leap is risky where do you find the capability blah blah blah blah blah but the money part is very much dependent on what you think the big buckets of challenge are going to be in in the in the 20 30 scenario right uh you you may be able to do it incrementally uh for some businesses especially like fintechs you will not be able to do it incrementally you're going to have to leap because all the interesting
Speaker 1question i think too in itself um and people can just jump in there yeah
Speaker 2what are what are those fields
Speaker 1what are those businesses that need to jump now i mean fintech retail who needs to move fast accounting legal
Speaker 3yeah but who doesn't need to move fast as well like maybe just utility services maybe they don't need to move fast but like a business who is in money of making business and have competitors and have customers there's
Speaker 1no more i'm trying to yes point to destruction though is
Speaker 2there there there are some where where you own the physical infrastructure you're relatively you're relatively you know safe right because if if i need if i need a uh connection to your service and it's physical it's very hard for me to disrupt you however for those companies to be competitive they have to look at iot you know internet of things sensors in the physical environment to allow them to manage asset life cycle maintenance predictive maintenance um where they have a third party workforces oh you know those third parties are generating data that you're paying for but you're not getting it because they're not instrumented to give you that data so there's still a challenge for them but it's not as existential as um as you know pure pure services businesses yeah so
Speaker 1basically if you're if your service value if you the values that you deliver is derived from people sitting in front of a computer
Speaker 2get cracking well yeah or or or if if the value comes from like talking or research um diagnostics those those ones but but i mean the the point you guys made about that that that every business will be challenged to some degree is true because even if the only thing you have to do is geo that's not set it and forget it that that's you got to check and check and check and check and check and then also open ai wants to have an advertising product obviously you know gemini google gemini is a great place to be and i think it's a great place to be and i will as well and if they allow um the semantic alignment to be artificially manipulated because someone's paid that changes the game as well so there's a challenge for everyone it's the it's the level of challenging
Speaker 3yes because we do a lot of work in construction and manufacturing although you would treat them as pretty traditional and archaic industries but it's probably not an existential but it's more about them being competitive and their competitors are actually embracing and adapting ai reducing their costs increasing their profitability improving the customer experience as well and you almost supply chain
Speaker 2optimization using ai just you only order what you need it's ordered at exactly the right time um you know it it automatically manages the compliance or tariffs or whatever happens when it crosses borders you know if if you uh you know it might manage sourcing for you so you never you never get to the point where that one thing that you need to build the one thing that you build gets so expensive that you can't competitively build it you know yeah there if you if you dig down there's always a way and if once again if you're if you're the only kind of supplier of that thing am i going to drive 100 miles to buy it you know uh you know some percentage cheaper i i might but but if you're in a very competitive space like car production um then yeah it's important you know i i heard you know that that um bmw had chips on on every single part they knew where every single part was at all times um so uh yeah there's there's a scenario probably like this for most businesses of a certain size you know where you know once you hit a certain size you get more prone to to you have more competitive levers than than a smaller business and we've
Speaker 1we've talked a bit about and they've been fantastic in digging into future thinking and immediate future thinking in a lot of ways but things that are that are needing to be planned for conversations that need to be had and a whole lot of a whole lot of getting moving but what about things that are happening right now so in the space of you know i'm thinking it could be retail it could be you know fintech could be anything else what are the investments that are that are in ai that are delivering return on that investment right now and what are the ones that are interesting
Speaker 2to you yeah okay so um well the ones the ones that are most interesting to me are what the big some of the the big tech company strategies obviously the the competition between the two big llms is really interesting like what they're doing and how they're evolving i mean they're both about to go public but you you're seeing them start to broaden the service for applicability which which is quite interesting it conversely uh google has connected gemini to literally everything so it doesn't it doesn't matter if you're using their llm because they're both about to go public but you're seeing their llm because you're already using gemini and absolutely everything and i think that's a very interesting strategy because they had so many tools that so many people use they had a distribution advantage they block you from buying a service from someone else because you're you're kind of already getting it from them so i thought that was pretty interesting but um but but the other part of your question is a lot of people as i said in the beginning a lot of companies are um they're just going very tactical right um so they're i i need to improve this you know mostly it's automation i i just want to automate this you know and i want i want to get my costs down and i i want to eliminate a lot of people which which which normally doesn't work and and and for the for the point you made amir it doesn't work because you didn't understand the architectural pattern that would have delivered what you wanted well enough to execute it correctly but i think that's a great point and i think that's a great point i think that's a great point and i think that's a great point i think that's a great point and i think that's a great point if you're if you're if you're working with agents you know you they're and i i'm not talking about like board governance i'm talking about you know agent governance like so what what happens if what happens if what happens if but but the other the other thing about um investment and and and why a lot of it isn't working is because if if you have a value stream and retail is the easiest one to understand you know the customer does discovery then they find your value stream and they find your your then they find your store or your website then they look for what they want then they put it in the basket then you buy it and then then then you pay for it and then it's delivered and then there's some reacquisition activity hey dabina you bought this would you like this um they are doing a lot of vertical ai integration and if if you look at a value stream and you improve one chunk it often negatively affects all the value stream and you improve one chunk the rest. And the rules about improving that one chunk can drive improvement in that objective, but actually reduce profitability, right? So if you are optimizing for churn, you don't want people to churn, it may keep unprofitable customers, right? That's one example. But all those activities in that value stream are linked. So if you do one vertical optimization, it can mess things up. If you do multiple vertical optimizations, they just can compete against each other and drive absolutely no benefit at all, even though. Have you seen this?
Speaker 1Are you seeing this?
Speaker 2Yeah. Repeatedly. Yeah.
Speaker 1Without throwing any names under the bus. Can you give us some examples of, you know, an implementation then where it's gone pear-shaped?
Speaker 2This happens mostly in marketing, right? Because marketing metrics are very. Yeah, they're very explicit and influencing customer behavior is like really, really important, right? So, you know, acquire, convert, and then once you're in, you bought one thing, how can I get your basket size up? What's your total lifetime value? You know, how do I reacquire you for less? And then once you're in, you buy one thing, you buy one thing, how can I get your basket size up? Then I acquired you before. I don't want to have to buy you again off Google. I want to reacquire you myself. And so we see departments that go after just one of those objectives. And then maybe one thing looks like it's working and then they buy different tools that do three of those objectives. And if you want, you can go to Quad even and ask for legitimate examples where this happens. And then you can go to Quad even and ask for legitimate examples where this happens. And it'll tell you. But specifically in retail, I've seen this activity, right? So it's not a mistake. It's a very hard thing to understand the conflict of the optimization, you know, models running in each thing. It's hard to understand it, right? It's hard to understand But, but if you actually want ROI, you have to think holistically, not vertically, right? There are times when there are times when like you see that this terrible and you knew like you really need to fix it. And you're prepared to accept that your unit economic will be more expensive, but you just add to fix your CX. Fine, that's fine. That's practical. You, you can do that. You know, going back to everything we said in this conversation, if you want to use it, if you really want to get advantage from it, you've got to consider it holistically, and you got to do the work to understand, you know, in a system, you know, when you throw a rock in a pond, you know, the ripples, the metaphor that we all use over and over and over again, that those ripples eventually, you know, hit the shore. And so you've got to consider that stuff. And where, where leadership in a business goes, fix it, fix it, fix it, fix it, we see tons of tactical activity, right? And the tactical activity, it wastes time, it costs a lot of money, but it also increases your propensity for disruption, because while you're not actually fixing stuff, somebody else is trying to cut your mustard. And, and, and, and, and in retail and financial services, the cut your mustard is coming from the gen AI searches themselves, because, oh, the other, you asked me the question about stuff that I thought was really interesting. Google and Shopify built the agentic protocol. So they recommend they have an agent, the agent goes out, it buys, and they they own the entire transaction. As a user, it's great, because I told you to find those speakers, you found them, you bought them at a good price, I get a good result. But, but, but for a lot of retailers, that's, that's not going to be, that's not going to be great. And if they're forced to use it, that means that they then have to figure out what they're not spending money on because that's not going to be cheap to deal with.
Speaker 3Excellent. My favorite question, what are your thoughts on predictions about super agents? We talked a bit about aliens before we started this podcast.
Speaker 2It's really interesting. Do I think everyone will utilize agents? Yes. Are you probably to some degree already? Yes. Will you have an agent on your phone that is half companion and half task doer? Like 100%? Yes. Does your phone itself become an agent? So not something that has an agent on it, but the phone itself is an agent? I'd say pretty much yes. Can you transfer that agent into your car? Yes. Does that agent live in your house as well? Yes. Will companies be run partially by agents? Yes. Will they be led partially or entirely by agents? Yes. I don't know when. I mean, a lot of infrastructure has to be built to make agent-to-agent interaction work, but we see the scale of the money going into it. I think that every major tech company wants this. And I don't know if you saw, Jeff Bezos has a new company called Prometheus. And basically what he's trying to do is put agentic engineers into your business, right? And this is the beginning. You know, like, does that take off? Will everyone fight for that space? I think yes. Will he win? I don't know. But in answer to your question, yeah, they'll be everywhere. That agent, at some point, that agent will be able to go into a robotic chassis and you'll be able to walk around with.
Speaker 3I don't even know how far. And your thoughts about the theory about that these organizations are trying to build super agents and they have admitted they are trying to build it, but one of the organizations would actually reach to a level where they would be able to, do the research so quickly that it would outdo all of the other organizations. Have you heard about this?
Speaker 2Yeah, I mean, the reason why Anthropic went from $350 billion to a trillion dollars in like two days, it was like three months, but it was a very short period of time, is because the biggest bets in the world are on who can corner intelligence, right? And everyone has a bet. That that intelligence gets to a stage where it can solve massively valuable and complex problems. The, you know, AWS, massive investments in OpenAI and Anthropic, Google investments in itself, and Anthropic, everybody in OpenAI, but the biggest shareholders in those companies are doing it because, you know, they're not just controlling intelligence. It's going to be the new thing. And on your point on sovereignty, if you're using a third party LLM as your government LLM, then that LLM has the ability to very subtly influence policy, right? So these are all things which. And we have also seen, Anthony,
Speaker 3that if an LLM is developed in US, the US government has the power to do that. The government has the power to stop other countries using it, which means if you have an LLM developed in China or US, founding everything on that basis, they can stop that at any point of time and the businesses can be halted.
Speaker 2We're just at the beginning, right? There's the two bigs, Anthropic and OpenAI, but you're going to see, we're already at thousands of models now. You're going to see like thousands and thousands of models. They're going to become, you have very generic models. Now you're going to have highly vertical models. You're going to have large spatial models, large action models. And I think countries will have to figure that stuff out. I do think though that the turn off of Fable and Mythos was political. I'm not sure that they really did anything wrong. And I'm sure somehow they'll work it out. And you'll see that come back on. But yeah, it is a concern. You know, it's just like the internet, you know, you were running all your store processes on cloud delivered stuff and the internet goes down. What do you do, right? And everybody eventually built stuff that would, if the internet went down, it would still work. I think, you know, you'll see a strategy like that where, you know, you shipped from a third party inference to local. You know, local inference until stuff gets fixed. It may not be as good, but it's good enough to just keep going. But I mean, this is a, this is a question that I don't, I don't have any real answer for. Yeah. You know what?
Speaker 3I sleep at home every, like every night I sleep at 2:00 or 3:00 a.m. in the, in the morning 'cause I'm doing some work. And yesterday I slept at 10:00 because my cloud was not working. So I didn't have anything to do. So it wasn't working like.
Speaker 1And we're in this space, aren't we? I mean, because there's the big tech race happening now and because the, the values that had been driven by, by these leading development companies. So it's Anthropic, it's Gemini, it's OpenAI, it's all of the above. And in a world where it's gonna be really expensive to use AI to do these things, if we assume that there is gonna, you know, it's gonna be a holiday, your budget's gonna go towards implementing AI. How, if you're a startup and you're looking for an exit, how do you structure yourself? -
Speaker 2Oh, this is the greatest question. Well, so, so I, I love this. The two things, one, I think you'll see the prices a bit. They'll, they'll, they're going up, but they'll go down, right? Because it'll, it'll become a utility, right? Just, and it'll commoditize. But if you're a rapper, then that's the term for that, what you just said. So you, you just have a, have a layer over a commercial LLM. There, there's this great example from the past there. Twitter used to have the thing, and it was Twitter. They used to have a service called Firehose, and Firehose gave you all their data, and you could, you could buy access to Firehose. And there was this very popular startup in the UK that, that utilized Firehose. And then one day Twitter just shut it off. And they went from, they went from a $400 million valuation to zero. If, if you are a rapper business, you do have to have appearing, a strategy for your, for your reasoning, for your inference, right? So you, you can start with one, but just like you peer cloud, you know, no really smart business relies on one. You know, if you're small, it's okay. It can go down and, you know, maybe that's not so bad, but big businesses, they peer, they, they, they take, they have redundancy with multiple cloud providers. And this is the exact same strategy that you'll do with, with, with, with models. You'll peer the, and, and, and there, there's another strategy for peering too, which is sometimes you don't want to use the commercial LLM 'cause it's too expensive. And some tasks can be done by smaller local models, right? So you, so part of your, of the control around your agents says agent use, use this one for reasoning on, you know, use, use the LLM, LLM for reasoning, but, but for this task, use the small model 'cause it's cheaper. And, and, and part of that, part of that control of agents is understanding how to navigate the agents so you get a cheap price. So you, so you, not, not a cheap price, but you get a price that makes sense to the task you're trying to, trying to deliver. But yeah, wow, this has been the best. -
Speaker 3That's been an amazing conversation. And I think we can just keep going and going. -
Speaker 1I feel like we just scratched the surface. I think this is a good 150 conversations we need to have on the back of that. -
Speaker 3Yes, we definitely need to get you on a sequel, a second podcast. - I think so. -
Speaker 2Yes, yes. - I love to, because when, when the, like, it's exciting when the, when the questions are exciting, you know, and I think you guys navigated the conversation really well. Like, you know, you, you ask the questions, you ask the important questions as far as I, as far as I'm concerned. And yeah, you just, it, yeah, without any, without any script, I, 'cause I, I've done a few of these now, and I think this, this was the, this was the most natural, but most compelling, because it's, it's not all, you know, jerky because you're, you're, you're, you're like, what's my next question? But yeah, and, and, and, and both of you have, have, have both a very instinctive knowledge, but obviously you're studied as well. So you both understand it, but but but you followed the flow and yeah that's good you guys should you guys should help companies if that if you don't do that already
Speaker 1well he he definitely doesn't i'll have my moments but um but i think it makes it an awful lot easier anthony for us yes it's so great when when you know having someone like you in here because we're learning constantly we're absorbing all this knowledge that that you're sharing with us and absorbing your your viewpoint yes and it's just so interesting yes i
Speaker 3just like it feels like the world opens i felt like just picking you up and just put putting you in front of like every every sort of yeah yeah business leader who's still sort of thinking about what to do next yeah
Speaker 1and if people can take away one thing from this conversation rewind it yeah share it widely is anthony talking about the importance of leaders
Speaker 3understanding the architecture i know i just want to say listen to this listen to this
Speaker 2you - just put it on i love it when you ask that amir i i i i loved it i was like oh that oh yeah i mean you know what you know what's this in the same in the same you know kind of line of thinking is boards i'm like um okay so let me take you through blah blah blah blah blah how are you how are you providing organizational governance that you have fiduciary responsibility when when the agent either eats five or six meals a day or eats five or six meals a day or eats six meals a day five million dollars worth of token or does something really really wrong yes that's that's on you but you don't understand any of these concepts you know for the so yeah it's that that that that's in the same line but yeah i i loved that comment because i i totally believe that and um and i've actually kind of believed that throughout my entire career when when you see a leader delegate the decision making to someone who is uh responsive to the business and not proactive about about building um you know technical infrastructure which supports the goal of both the customer and the business yeah you don't you get a big mess you know and yeah so anthony
Speaker 3how can people reach out to you or your organization how can they learn
Speaker 1more from you where do we find more
Speaker 2people i'm on linkedin that that that that would be a primary contact point um or or they or they can they can ask you and and and and you you can connect me under the
Speaker 1in the comments feed absolutely yeah i don't i
Speaker 2don't i don't mind but i i loved i loved talking to you and and if if we do another one i'll make sure that i that i can get down there although you know we would have probably been wild in person so yes yeah yeah
Speaker 1but but yeah
Speaker 2like uh like a great feel from the two of you to the texture is uh you know it's like um it's really inviting you know that the the two tones the two the two viewpoints on the questions yeah i thought it was good you guys got a good podcast i like it oh thank you that's a good part
Speaker 3of it i'd be someone like
Speaker 1you uh it was awesome thank you so much for for joining us anthony um we will absolutely be chasing this up and coming back to you another time but you've just you've delivered some really fascinating things for people to think about and your thoughts on leadership and structure and how you target ahead to that 2030 and really think about where's the disruption coming and what do i do how do i get organized now that's that is the that is the conversation that everyone who's in business and really in organizational positions everywhere needs to be having right now so thank you thank you for
Speaker 2bringing that to us you you you are welcome i've loved talking to you both and i look forward to talking to you again oh and yeah when when when you're when you're done send me the link and i'll post it as well fantastic yeah i'll give you i'll i'll post it on sub stack and um and linkedin very good but for everyone
Speaker 1who has joined us on the podcast uh thank you so much for your time i hope you enjoyed that as much as we did i'm not sure that's entirely possible but i hope you enjoyed that as much as we did i'm not sure that's entirely possible but we hope it was fun for you too and we'll see you again next time on the dumb monkey show
PodcastTranscript AI