2025 AI Year in Review: Why Gemini Won 2025's AI Model Wars

Transcript

and the power of the agents. How do you make them autonomous but but also within control? Have like a measures in place to make sure they don't go off rail and drive things off cliff, right? >> You've got to have some like artistic talent and be like a singer, songwriter, dancer or you've got to be a plumber. Like there's no like there's no white collar jobs that are left. >> Are the people putting their money on this, right? Um how do you guys think about this? What is your favorite model? What's the best AI at the end of the year? >> Welcome back to AI Supercycle, our premier AI show airing every single week

presented by Near. We cover the ins and outs of decentralized AI, privacy, and the future of this massive technology. Near is the blockchain for AI and the execution of AI native apps. You can check out Near's latest AI product at near.ai. Now sit back, relax, and enjoy the show. All right, guys. Welcome back to episode number 14 of AI Superco powered by Near, our friends over there with their private AI chat product. They recently launched absolutely epic as well as some other really cool features in their latest launches with Near Intense, too. So stoked to have them as our partner for the what I think is going to be one of the best episodes of the year. Here we have Pay Chen. We're gonna do a 2025 AI year in review live on Rollup TV. Hey, it's good to have you

on. Thanks for joining. >> Thank you for having me, Andy. Nice to be here. >> For sure. Uh Rob's here as well, obviously. No cam. Yeah, shout out to Rob for making it happen today. >> Good to see you, Pay. >> Good to see you, Rob. >> Thanks for being here. Yeah, I'm excited. Uh I think you have a very unique and and interesting perspective on the market in general, but also uh with AI. So very very excited to talk about everything that has happened this year in AI. >> Definitely. >> Definitely. >> Let's kick off with your role CEO of the Theor Foundation. Huge week for you for you guys. Token launch, TGE, tons of community initiatives, just a lot of stuff happening. What is your role with um this company? How is this week? Give us the intro and then let's kick this

off. >> Yeah, definitely. So, hi everyone. I'm P. I um I'm the CEO and executive director of Theoric Foundation. So I on a day-to-day basis I oversee the um commercial side of things managing the foundation very involved and did a lot of work along the way uh leading up to the TG of THQ. So we uh went TGE on Tuesday and December 16th and it's been non-stop since then. Obviously the market is super challenging and as a company that is super um focused on building the actual tech and the product there's a lot to prove right and we are just trying to kind of keep our heads down and get through the TG and now look at the stack where we are but prior to

the TGE we already launched the the alpha vault which is a AI managed um uh D5 vault focused on east growth strategy it's a very unique product that it has a AI layer and also had it has vault of vault infrastructure which means under the main vault there's different strategies we are able to work with both human and agent um curators of strategies and making it just more fun. So I think yeah let's uh we can get started and can uh talk a little bit more about that.

>> Awesome. Yeah, I guess um just a couple quick things to clear up for our audience. You guys frame this uh company as an intelligence layer. Kind of walk us through exactly how you guys are using AI today. I know you guys are are doing policy cages. You have this Alpha Swarm product. Um talk to us just a bit more clearly about what the unlock is with your product stack using AI. Um and just kind of you know what exactly is happening under the hood. I know there's a lot. So just kind of a high level overview it would be great. Yeah.

>> Yeah. So I you mentioned like the uh the policy cage which is essentially a cro uh programmable guard rails that we put in place right to define exactly what AI agents can do and cannot do. So we think like we're not trying to limit uh the AI here like the brains is still there but we wanted to kind of limit the um the what what actual things that the AI can touch and handle and control. Um primarily the reason for that is to keep everything that's like stay at state safe right so let's say onchain capital um the key management stuff around that and when the agents can sign transactions all of that is super super sensitive and important to get right from the get-go. So, it's kind of trying to, you know, use analogies like if you

have a very um welloiled and high performance car and you want it to be running as as fast as possible, but you don't want to you do want to define the track where the car runs on. So, you wanted to lock them into safe track and that's kind of like how we want to uh think about when designing the product with AI at the center. But just going back to like generally what we're building as like intelligence layer it's essentially is um thinking about that we're um uh theoric is unique because it's building not just like one product but it's building a stack right the stack has like the top layer of the AI um we call alpha protocol it's essentially is agent to agent

communication and coordination layer so it allows agents to register discover each other and communicate and coordinate pay each other, all the basic functions that you will see in in economy, right? And the second, it's almost like a layer one, but we're not really a blockchain. >> Yeah. Yeah. >> Um >> and in the middle um is alpha swarm, which is kind of what you're referring to as the intelligence layer, right? That's really important because those are multiple modular uh it's like a modular a like agent swarm but has multiple agents to super specialize in different areas. So we use um train training those agents making sure each one of them is particularly strong at doing what they

do. So for instance there's a policy agent understanding all the rules. There's execution agent super fast on onchain execution automation. There is the knowledge agent that can grasp everything about a project, a token, a product, a use case, um a D5 protocol, a platform, right? like all the knowledge specific agents can come together. So the intelligence layer is really describing um the alpha swarm layer where um the swarm of agents can come together and complete very complex uh tasks like providing liquidities to different dex pools across different chains or different platforms and being

able to find the best yield in a specific asset like ETH or BTC or stable coin and they're able to do it simultaneous at once. So there's no single point of failure because it's very scary to think about >> December 2024 we got a lot like lot of like trader agents like singular agents that >> issue their own tokens and say hey please park your capital with us we'll make you best yield right and when that agent goes down >> when the agent goes down or the chat uh GPT like AI chat goes down >> that's you're done right that there's no like accountability built into that >> well I think Okay. I think that like this kind of guard rails, it's almost compliance, but

it's like this guardrails, this really safety security focused mindset towards the evolution of this technology. It is and to do that enforcibly, verifiably in terms of like commitment schemes on chain, it's a really good match. It's really good like match made in technology heaven. Uh, which is really nerdy, but I think it's like a it's a good application. And when we zoom out and look at some of the stuff that that happened this year as we kind of roll into this year in review, the agents are just getting better. OpenAI launched GPT5 over the summer. It's it's, you know, at the time it's strongest AI model. Now we've got 5.1 I believe or 5.2 and double check. There's tons of improvements. 256,000

token context windows. You can read and process entire books, articles, conversations. you had these huge generative AI to transform operations of government um with Elsa and Intact. Microsoft had this project. So just kind of like giving you guys a bit of a sense of how much actual innovation happened this year in the category. So as we kind of zoom out of crypto AI, zoom out of of kind of the intricacies here and look at to what happened. I guess let's just start this off. Robbie Pay, what what is a a couple words or a phrase or or a sentence that you would use to describe 2025 in AI?

Um, I can give it a shot first. You know, you know, there's I I would just say extraordinary in terms of growth, right? How would you describe the evolution of AI in 2025 and then let's get into some of the specific events, prediction markets, and more. Um Rob, I'll pass it to you actually first if you have something to uh add there and then let's go to pay. >> Uh it's difficult to sum up in in entirely in one word. I I almost want to use two words like one to describe the first half of 2025. Uh and then one to describe the second half. The first half felt pretty groundbreaking uh as we you know took a lot of the

machine learning a lot of the uh the deep neural network math that was used but it finally became tangible. Maybe that's even an even better word is finally LLMs became tangible in the first half of 2025. Um and that's when we got LLMs and that's when they started to do things on chain and and a lot of that was really exciting. um these things have gotten a lot more uh usable and a lot more um they they've gotten a lot more context and I think they're a lot more believable and useful and effective. The second half of 2025 um is when I think we got uh much more incremental returns on the performance of these things. Um, I guess uh I I

would probably use a word like integrated the the second half of 2025. I personally I didn't see a lot of model improvement. I don't think these things got uh drastically better with each of the releases. We saw that in Chachi BT5 it sort of fell short of expectations. 5.1 5.2 these things are just very minor increases in the performance. But what I do think uh was um I what I think happened more so in the in the second half of this year was that we started to get these AI products in integrated into the existing apps and systems that we use on a day-to-day basis. We saw Gmail like now has actual AI uh responses baked into it. One click and it's

actually pretty good. Uh you see this in Google Sheets. You see this across the entire entire Google suite. were replacing search across the board. So I think the first half of 25 had drastic improvement in the models. The second half we're starting to make improvements in the UX and we're starting to bring them more or make them more integrated into our everyday uh systems. Those would be my two words. Groundbreaking and integrated. Hey, uh I'm curious. I'll kick it over to you. What do you what do you think if you were to come up with a word or two to describe this year in AI?

I I definitely echo what you said Robbie like all of all of this feels like a much closer in reality. I think I would um use the word uh this year we witnessed AI going from capability to um operational reality or operationalization right so you start seeing large enterprises and companies are putting AI in front of the audience the users they're stepping to demos they're showing up at workplaces they're handling much more aggressively and directly with customers and doing research writing codes and managing even now like capital right so they're really coming into operations so the second

point is now how do you address the trust issue because capability wise AI can do all the stuff but are they reliable can we trust them how much capital can you um put with them and other information and data right so I think you start seeing the qu the focus of the question start shifting from can they do it to like should we trust them doing it and from my corner of world, we're just so focused on the niche of AI and crypto intersection. Um, which is where I think the gentic AI is being mostly discussed, right? Like we're now web two SAS company trying to bring AI into the product and day-to-day and automated things, but we really think of

AI as um a gentic individual like an economic actor. So I think it is um where actually I think you start seeing this year more of a um kind of the the re the rise of agents and autonomous agents is now becoming a real thing. Um so yeah so like we we have a lot of kind of um definition or work around how do we define the scope and and the power of the agents? How do you make them auto autonomous but but also within control have like a measures in place to make sure they don't go off rail and drive things off cliff right um the final thing I would say is you know this is

just like where I feel like AI is now meeting the real um economy having them managing capital coordinating resources uh writing legal contracts and negotiating everything right this is this is where you you move um the economy from softwarebased to uh AI and economic agents based economy. So it's it's super exciting. I think >> doomed. Yeah, I think we're being replaced in in certain areas, but it's for all for good reason. >> Just something to add real quick there. Um I I saw this tweet from CL 207, the OG cat on crypto Twitter. He's like, I'm getting bullish on Amazon 2026. And I'm like I go down and I start to read why.

And he's like, why? because I think that Amazon's going to put out a a release saying we're going to replace half of our employees with drones and robots by September and and then that then their their margins are going to look a lot better. Um Rob, did you have some of that? >> I I mean just on that, I think hardware and robotics are uh very intertwined with the AI rise as well. And I think that's when a lot of this and efficiency and a lot of these new things start to get unlocked when we when we you know use AI as the software in a robot's brain if you will and then you get the best of both worlds. But I wanted to ask Pay a question actually because you know we've seen um you know how difficult it is to bootstrap liquidity in DeFi and

theoric is kind of out this cross-section of DeFi and AI and you know something actually occurred just a quick a quick tangent here is that you know something occurred in my mind like prediction markets um are becoming very very popular. there's a lot more teams that are building on these prediction markets, but it's very very difficult to bootstrap your own markets just like it's very difficult to bootstrap your own uh DeFi markets. And so I wanted to ask Pay, are are you guys um looking to generate liquidity in your own smart contracts, kind of rise that through, or are you guys mostly focused on look, we're going to we're going to uh be like an aggregator or a portal, if you will, to existing DeFi apps. we're going to

allow you to deposit in those and we're going to provide some of the AI capabilities to coordinate amongst those apps. Do you that that's you know I I I do think that um you know there's kind of a distribution play and then there's a tech play. I was curious which which side you're taking. >> It's spot on. We So the interesting context here is that we we start with a letter. We we really wanted to kind just like dive deep into agent being the best alpha um generator of the industry entire industry like forget about I'm not not calling names right forget about like math edge and all the R7s right out there like we want agents to be the curator and the generator of the the

strategies. So that was actually when we started building Alpha vault. We really thought about the first um move would be doing this hatched LP strategy vault curated by agents purely and we actually used AI to do all the back testing work and prove that very substantial amount of alpha can be easily generated in a very small amount of TV let's say like 25 million can generate like 25% of that um as an alpha very easily as AI but >> when it comes back to >> that's a crazy claim 25% returns like or alpha on $25 million is just easy for this AI. That's >> if you can really deploy that agents and they fully trust it to do, right? Um

that's back testing and I'm >> not giving any agents by private keys guys. Sorry about that. >> Yeah, you can you can you can manage the key though. >> There you go. There we go. >> Or at least be a sign majority signer. >> Um so we we thought about that and then there's um a bunch of >> obviously like complexities going into that. you need to uh if it's east you know growth growth strategy you need to constantly hedge so you arrive at delta neutral in east terms east is falling in this market so that is not you know even if you're like a super maxi of east you have some concern about like accumulating east right so I mean based on like kind of market conditions we uh we did a lot of pack testing we know it can be done this is something we're

going to build in Q1 2026 but we decided to before the TG we decided to take the second round which is what Robbie described is an aggregation layer or distribution play uh which is essentially why we did the allocator agent first in the uh currently live in the alpha vault. So alpha agent is kind of like a portfolio manager and the risk uh manager in digital economy. So it sits at the main vault level and it monitors it takes in the TVL and makes a decision how much of that TVL it's going to put in each of the sub vaults that has distinctively different diversifying strategies. So right now Alpha vault has

two subvault and two curators. One sub vault is the uh is Lido strategy vault. It's a um leverage uh looping strategy and then curated by Melo who happens to also be our uh vault infra provider. So they've seen a lot of different strategies in their time of building Melo um core vault right and the second is uh the um course one uh me max. So this is using like os and deploying into a ecosystem. They use a different strategy also with ETH as a core in the based asset. Okay to we're trying to use those at least those two vaults to uh

really perfect u allocated agent capabilities and over time in Q1 we're going to add a third vault and that will be the hatched LP vault managed by agent and that's something that we're putting a lot of effort in. I I think it's honestly kind of hilarious that the first industries to get innovated uh by AI are these like white collar jobs like >> hedge fun managers taking financial advisers like this robo advisor is going to be so much smarter so much safer so much >> so much better than any financial advisor and all of the tokenized funds are going to be the tools by which these robo advisers are going to offer their customers.

>> It's insane. It's insane. >> Um, >> it's like it's like you gota be like a in 2035 you've got to be like a you've got to have some like artistic talent and be like a singer, songwriter, dancer or you've got to be a plumber. Like there's no like there was no white collar jobs that I left. >> I'm glad we didn't choose financial advising as our category of choice for the career uh all of us here. or not, there's anything wrong with it. But yeah, Coinbase just rolled out their advisor product on their announcement on Wednesday. And that thing is going to outperform and really be there in that kind of assistant AI fashion. Guys, let's get into some of the prediction markets that we have for for models, for AI training, etc. We'll be right back.

Quick break and we'll bring these up on the screen for you. Um talk talk the AI models um main best model and and a couple. Let's go to a quick standby, guys. Guys, we'll be right back. >> Havachi is fast and private per trading secured by ZK with Celestia underneath. Go to habachi.xyz to start trading today. Not your keys, not your coins. From the team that pioneered cold storage, Treasure has just released its new wallet. Guys, if you are still using hot wallets in 2025, 2026, you are missing the entire point. Secure your coins. Get a treasure at treasure.io. Ye is built for big moments. the most fun you'll ever have with your crypto. Go to

yeet.com to yeet into your next game. >> There we go. Alrighty, guys. AI year in review. Here we go with some of the prediction markets. Rob, best AI at the end of 2025 from our partners at Kelshi. Right now, Gemini is absolutely crushing it. 92% chance of being the winner because of their uh GP Gemini 3 Pro launch. Grock is 6%, chat GBPT is 3%. $15 million of volume on this market. Guys, I must say I spent most of my year using Chat GPT five when they launched it. Since then, before that, it was kind of a mix. I've been using Notebook LM on

Google and Gemini notes. Gemini notes are incredible. better than Fireflies for calls. The Google doc, like the Google integration here for Gemini, plus this new model for videos and images is insane. Gemini 3 Pro, if you can see on Kowski, right at about August, September, you saw how Gemini just started to fly. You see the the second half of this chart, you see in September down below, Gemini is the green one. It went kind of like parabolic run from about 50% all the way to 90% where it is now starting in about August and I that's very much aligned with the GP with a Gemini 3 Pro launch. Guys, what's your take? Pay Rob, what do you guys

think here? Is this right? Is Kelshi right? Are the people putting their money on this right? Um, how do you guys think about this? What is your favorite model? What's the best AI at the end of the year? >> I Yes. So I I think like um it this looks kind of it looks right to me because from my personal experience we use AI a lot of work and you know not only just for product development and back testing um but also communication marketing and personal messaging posting right like like instantly you you realize like after using for quite a long time and then you see the grog and and then you experience Gemini Gemini is like kind of like by far the the best of all like just overall experience and

overall output. Um because I constantly compare the outcomes across all three of them. I would say occasionally I go to Croc, I like the sense of humor that Croc has like the others don't. It just has like that like Elon Musk print footprint there. Um it's very funny and so like if I want a good joke or when I say like please mix things like a funny joke in there, I go to Croc. Um but um Gemini consistently I think better above um than others. Robbie, I don't know how you feel. Um yeah, I primarily use chat GPT as my main driver because it kind of had the first mover advantage in my in my own technology. Uh and it had some like it just took up real estate on my phone and

then I just went to it to it and I had that behavior just baked in. I think um what's interesting about these markets is that I actually looked a little bit deeper into this and and the way in which they resolve which model is the best uh is according to this website called LM Arena. Uh and so what you can do is you can put in a prompt into LM Arena and it is essentially a uh a rollup or an aggregator of all of these models and then you choose which model you think gave you the better response.

So it has this mode called battle. You put in a prompt and then it gives you two two responses for each prompt and you pick which one. And so what this means when you say best AI is that uh according to this website more people are choosing the response from Gemini than any other model. Um and I think this is what we're going to see. I can even anticipate like in theoric you have robo advisors using different models and they're competing for your business and for your capital. And I think this is this is the schematic that we're going to see a lot more of. Um, personally, I like to use Gemini for any kind of deep research. If I know that there's something that I want to do like just

one, you know, one back and forth, like one cycle, I want to ask one question. I want to get one answer like like search something I would normally Google. I use chatbt for that. Anything that I know, I would probably click multiple links. I would probably check multiple sources. That's when I go into Gemini and I do a little bit of deeper research. But I do like Grock, too, because Grock I always I do find myself pressing the little button at the top of people's tweets just to like the little searching is super searching on Twitter is where I like to use GR.

>> Completely agree, Rob. I think you're actually spot on and um pay this probably is very applicable to how you guys are thinking about adoption from theoric's perspective, which is Chashbt. The app was on my phone, Rob, as you said, early and it was just there now and it was like the I was speaking to it. Great. Uh, notebook LM got on my phone. Grock got into Twitter and just clicked the click. And so it's like if there was an AI Grock, for example, a let's say let's say a Theoric agent model that has their automated agent stuff that was in Rabby wallet and I just could click it and like talk to it and use it. that distribution integration is like incredibly useful because I mean I mean obviously right there's for obvious reasons there but

like I think Rob like it would be funny to actually ask that that website LM Arena which model it thinks will like be the best at the end of the year and watch it kind of have different models show their biases if they think they're the best or if they think somebody else is the best. And anyway, uh >> we'll do that now and let's check back in at the end of the show. P pay, anything to add there on uh on what Rob and I just kind of said there? >> Yeah, I think um I think it makes a ton of sense. I mean, we it's I want to comment on your what you just said about the distribution and kind of baking into the um the stickiness user behavior

platform, right? Um obviously there are two products we're talking about here now since TG we have THQ is actually a product, right? Right? Like we need to now think about the stakeholders for that token itself. But then on the other side um of a vault as a where agents and AI is used a ton and we want to push to the real human users and attract TVL continue to build on top of the 24 million that we currently have. We want to surpass 100. I want to go to next. The more strategies you deploy, the more TV you need. All of that stuff, right?

So we we definitely right now we are um foundable I guess like accessible through like our own UI but also Melo front end where uh yield hunters can go to Melo look for all the vaults they'll see us there's turtle club who is our TVL bootstrap partner so they have a liquidity dashboard for us they have the the campaign for the TVL where you can see more of the the points kind of how everything's calculated and the accumulative yield so so to speak like not only the eield but also others. I think we need to do more. Um I think we need to get on all the dashboards for vaults. We need to get into the apps that live in u people's phones, right?

So the Phantom, the MetaMask, the all the other um Coinbase wallet. We're in the base ecosystem, right? So we need to be in that app. >> Um so the app will be naturally come >> and that's why I think Gemini is winning. Like sorry to just cut you off there about like the app thing, but you just like spurred me off where it's like Gemini has a really good model and like the Gemini 3 Pro that they just launched is great, but once again on Google Docs it's there on Google Meets it's there on uh Google on Gmail it's there on Sheets it's there. just like it's just there and it's so deeply integrated into Google's stack right in this example in your example in the theoric stack and it's like >> it's just there and it's so useful because it I guess like the big thing

here Rob is this kind of concept of this M MCP evolution too it's this cross multi-platform kind of like API calls that is really great so it's like for pay it's like can theorex agent work across one chain one app and one and another app on a separate chain or you know X Y and Z right in in Google it's like notes sheets meet it's kind of like this multi-platform multi-use case approach that is making these AIs better than others in my opinion >> well I think that's why the the word for 2025 and AI is integrated like that is at least for the second half of 2025 because that is what I'm finding myself use AI this way the native integrations

I think is where we tend to go I use Grock on X. I use Gemini on Google Docs and and Google Chrome and Gmail. And then Jad GBT, I do think has started to take on search. I I think Google was late to the party in integrating Gemini into Google and so it became more useful to use chatbt especially on voice mode. That's how I like to use it. And and use it as my search modal. So I I do think um now that these uh these LLMs are getting integrated into social media and uh browsers that is going to be the moat and it's really an extension of their existing product suite. So in order for

Grock to get more distribution and and distribution and activity is going to be what separates the performance of these models because the more data that they have and the more prompts that they get, the better trained they are. So I think Grock as a model is dependent on how successful Twitter is in gaining more social media market share and Gemini as a model is is dependent on you know the existing Google product suite. Um, and these things are going to start to diverge. I think right now they're very, very similar in their models, but because they're being trained on very different data, I think they're going to ultimately uh become more uh niche and optimized for those use cases.

>> Yeah, love it. Love it. All right, guys. Let's move to the next one and the and the last one. The second and the last image uh to talk about here for the AI year in in review 2025 is which AI company will have the best coding model on January 1st 2026. Now guys as you can see here when the distribution part of the equation when the multiasset the multiplatform multi-functionality kind of concept is stripped away from you know different platforms and distribution and get directly down to coding Mr. Sama is in the lead. 85% say Sama, 16% say Anthropic, 3% say Google.

Guys, I have not really used much of these uh coding models. Rob Pay, I'm not sure if you guys do a ton, but how would you guys rank this? Is this right? 85% OpenAI. Are they just that much better at coding than Claude and Anthropic? Is are they just that much better? >> Um, pay may have more than I do on this. I just I did type this into the uh the notebook or excuse me the LM arena battle mode and it's pretty clear that each of these are talking their own book. This it it gives you two. It gives you assistant A and assistant B gives you two different uh responses.

Assistant B says you know clearly clawed three and a half sonnet uh is the best. The other one uh says you know GitHub copilot if you're already using VS Code uh for more powerful general purpose coding. No IDE integration needed. Sign up for GPT4 via Open AI. So there's definitely um a sense of uh competition. >> That's so funny that that they're biased. >> They're absolutely biased. They're talking their own book. Pay, what do you think? >> AI for coding agents.

>> Yeah. >> I mean, we're just one of the many companies that use heavily use AI, right, in day-to-days and research and product development. I I can vouch for Claude being kind of the primary cloud from anthropic. So I guess anthropic in this case. Um but yeah, I'm not sure the the number but like anthropic definitely stands out as >> the coding side. >> Correct. Coding, back testing, uh reasoning, training. It's the mostly um I heard um I've been hearing the most from my engineering team, research team, right? So I know that they're using it heavily for developing alpha swarm and different models and thinking about um the trading strategies and back testing the outcome of the yield and all of

that. Um yeah I I feel like pretty strongly about that. Um >> interesting >> I'm sure yeah open eyes like overall in the rapid prototyping and coding is probably still mostly used or heavily used. >> Looking for yield that stands out? Infinifi gives you just that. Deposit stable coins and earn far more than the current savings rate. Infinify is a battle tested, reliable DeFi platform that is ready to go. Head to infinifi.xyz. Holiday is unified crypto payments. Give your users more wherever they are. Go to holiday.xyz to never write a smart contract again.

Trade on anything, economics, weather, sports, crypto, and more. all on couch.com 24/7 for the world's biggest prediction market. Isn't it kind of an interesting phenomenon where your engineering team at the Oric is going to be using these open AAI anthropic primarily as you said models to code to help code and then what they're trying to code and create is another separate AI model agent. It's like AI inception. It's like what's going on here? Like we're using the LLM to create another LLM, but it's very specific use case type of one. It's very interesting, isn't it? It's weird. It's weird.

>> Yeah. It's a we're entering a gentle economy right? >> Yeah. >> Both sides like a full loop. >> Yeah. >> There's no no 401k. Over at Theoric, you use Theoric to manage your 401k. >> Yeah. And theorics 41k management tool is brought to you by whoever trained anthropic and how good obviously their engineers are and everyone's has engineering work. not like just, you know, press a button and things built. But obviously, I mean, let's just say if we're going to make predictions about kind of on the same train of thought around like the jobs and the Amazon and the drones, you ought to think junior engineering roles and that the idea of doing a compai major in university

has diminished significantly. >> Oh, I I' I would push back on that. Okay. I would say look, AI is a calculator. like you don't see like less mathematicians and statisticians and computer scientists and all these all these uh STEM oriented uh careers diminish as a result of a more powerful calculator being uh created. Go back to like uh the printing press. Did you think you know authors and journalists and writers were going to go out of business? No. It actually gave them more tools and a more powerful means of distribution.

>> And I I I think that's what's happening. I think we have a fellow math major in the crowd here. Pe I'm not going to say who it is. A math and philosophy major here from the University of Florida who's really upset that his math major is not really worth it. He can just ask AI anything at any time. Like why do I why do I need a math major? >> I use my math degree every day as a podcast host. >> Hey, what do you think about junior engineering roles going forward? Uh h how do you see coding evolve with AI in 2026? And honestly, I think it'd be cool to talk g get your perspective too more specifically on some of these solidity and EVMbased programming coding agents which are really firmly able to now do smart contract development, white hat

exploiting, if you will, white hack uh uh vulnerability searching, etc. How do you see the evolution of programming with regards to the broad scope and then in the in crypto native land as well going forward? So I think um I still see the the great need for smart contract developers and the talk about the human ones not the agents. Uh we need them because like I think because the this uh space is evolving so fast there's always integrations with each other. there's nuances, there's edge cases to test like you just can't um predict what kind of plumbing and you know smart contract level integration will involve and what kind of scenarios and we're creating new

use cases as we move forward right as as an industry in general. So I I do think there's we're desperately, you know, looking for smart contract developers. I'm not going to lie. We're looking for strategists and we look for all the technical talents um as we kind of move forward as a as a project. Um but I think yeah like junior engineers I think will enter the market and with a very different outlook and they need to understand that um their strength needs to be a combined you know how to work with AI and work with AI tooling right so meaning like you got to be very good be good at like vibe coding because it it kind of improves your efficiency and your speed of turnaround of of the tasks but it still needs someone you're not

going to ask someone in marketing major to do engineering work um because it's like so now it's so trivial. You still want to someone have the back the background of and you know coding background understanding how to audit and how to interpret codes and be able to capture the hackable like areas like look for the bugs right that is still very much needed um and I think a math background is very very relevant in this uh in this world like as we move forward everything is kind of algo driven and math tied right it's going to become a lot more numbers relevant versus like content driven and and all of that. So yeah, I mean um >> insane

>> the job market will be very different but you you need to be able to do the job but with like an AI in the back in the back back of your mind basically know how to use them to accelerate your work. >> Hey Chen, it's been an absolute pleasure having you on today. What a week for you guys at the Ooric. Fun conversation today. This is this has been an incredible year in AI as well. Guys, it's AI Supercycle episode 14 signing off here. Pay, thanks for your time. >> Thank you, Andy. Thank you, everyone. >> All right, guys. We'll be right back. Quick, standby. Thanks, We'll say bye and get our 2025 year in review tokenization segment up next. See you guys in a sec.

>> See you.