Teng Yan: How To Capitalize on The Biggest AI Bubble Ever (...And Who's Winning Already)

Transcript

I think for me personally in terms of like using in terms of like prompting models I think what's important is to understand what you [music] want to produce ultimately what is the outcome define the outcome very clearly for the AI to follow and then all the intermediate tasks I think that is really um up to you is this a bubble tang is this a big freaking bubble are we in way over our heads right >> [music] [music] >> Welcome back to AI Supercycle, [music] our premier AI show airing every single week presented by Near. We cover the ins and outs of decentralized AI, privacy,

[music] 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. AI. Sit back, relax, and enjoy the show. >> Yan, welcome to Rollup TV, man. Good to have you. >> Yep. Hey, thanks. Thanks, Andy. Like, um, glad to be here. Like, uh, yeah, I I think like this year 2026 has been just like, um, very exciting on the AI front. So many new things are happening. Um, very excited for the year ahead as well.

So yeah >> definitely, man. I mean, it's been the year of AI. I actually tried to um predict on on Khi that um uh this was going to be AI was going to win the time person of the year award, but [laughter] I think something just leaked yesterday and it's not. Uh but we're on AI supercycle. Welcome to AI Supercycle Tang presented by Near the Leading AI blockchain. They just released their Near AI chat product as well as their Near Cloud. Uh so you can actually have an LLM that you use and speak to uh that is not storing your data, using it for training or giving it to the local jurisdiction of choice to put you behind bars. We're not going to do that today.

Tang, we're going to have a fun time. Let's start at the top. Tang, who is Tangen? How did you get started in um uh kind of researching and creating content about AI? I'm a big fan. Um so thank you for joining us today. Yeah, you know, thanks for having me. So, you know, before and not too many people know this, but like actually before getting into crypto full-time, I actually spent quite a bit of time in the AI space and and that was of course um pre-Chat GPT. Um I you know co-founded a healthcare AI company back in 2015 and we were doing very traditional machine learning on very unstructured data which was mostly like

clinical notes, doctor's notes and from there we sort of like built platforms that help um pharma companies run some of the studies in in a more efficient um sort of way. Um so that was really where I I spent a lot of time getting to know like the know what was happening in the AI space. Um and then yeah went full-time into crypto. Um worked with Deli Digital doing a lot of like consulting research um in the space. Um my focus was a lot on the consumer side of things like consumer applications uh gaming NFDs and stuff. Um but I would say like sometime around like two years ago I started to notice that hey you know what was happening in the AI space was really starting to take off. like if you use some of these products at the

time it was probably like chat GPT maybe you know GPT3.5 or something like that um and you could see that actually it was able to you know start to do some very amazing things like things I never expected that we were able we would be able to do um so that made me realize that hey you know there's something here that's going on I need to dedicate as much of my time as possible um to understand what was actually happening in the space um so yeah that's where I I left Deli um started like researching into what was going on. I would say um the thing that really brought me into this space was Bit Tensor. I think it was as I was looking like what people were trying to do with this um it was just like so fascinating. Um and when I

started researching about it, started writing about it. Um you know, I kind of realized that it was kind of like difficult to get out of this rabbit hole. And you know, it's been a year and a half. I'm still spending a lot of time looking at what's going on and still learning new things every single day. So that's >> yeahing that's the beautiful thing about AI is that it just it it's never ending rabbit hole. Um it's just a never ending rabbit hole. And I mean I've scratched the surface but man is there a lot more to uncover. Um so let's kind of start on a on a on a more broad here broad note here w with regards to the evolution of the last year or so. Um I is is there a particular sector of the AI uh category that you feel like has accelerated

far faster and far um you know uh more effectively and efficiently than others this year? Like you know has the has the model uh improvements this year just been so much better than anything else in AI? Has the uh ability to uh have actual like memory in certain LLMs been so much better? like is there a specific part of the stack that has really really improved this year that you can notably point to and be like this is why uh your tangible experience has become a lot better this year. Kind of give us the uh lay of the land as to where the major improvements across the general LMS happened this year.

>> Yeah. So, you know, I think uh it is very obvious that um especially been you've been using some of the AI tools that things have improved a lot like over the last year um >> compared to just about a year ago and a lot of it has been driven by better and smarter models like the recently released like Gemini tree I think is kind of like blown a lot of people's minds away in terms of just how fast and how good it is at doing um more complex um task. Um and I think a lot of it is driven by just a relentless drive of improvement throughout the industry. If you look at what goes into an AI model, right? There are probably three main components. There is the data, um there

is the compute and then there's the actual algorithm itself. And all three of these like frontiers are seeing like constant improvements. We're getting better data. People are producing more and more like relevant data, the synthetic data even to train models around like very certain task. um we're seeing like bigger and bigger data centers being built out to train these large um and larger AI models. Um and then of course like all the researchers are also doing um um looking into how we can improve like the algorithmic efficiency of like training these models, right? So all in all like it means that the scaling laws of of AI are still pretty much intact which means like the more data we put in the more compute we put in the more um efficiency

the more improvements in the algorithm we make like the smarter and better the the models and the LMS will be and we've kind of seen that with some of the latest models out there. So I I think we're going to continue to see a lot of these like very steady improvements. Um I'm going I'm expecting um newer better models like throughout like 2026. Um many of the major labs like OpenAI have already committed to releasing like um their new models on a more regular basis. Um so I think it's just going to be a good time to kind of like see. So I I think like things are just like fully um [clears throat] fully charged. Um and that has really driven I would say like uh a lot of interest and growth in uh

particularly the AI agents kind of like space because like um ultimately um AI agents are really constrained by what the the models that power them can do and with these new and better models I think we're going to see them be able to do a lot more task um and maintain that focus over a longer s of time horizon. >> Yeah, I love that. I I haven't tried to use the um Google Gemini uh Pro 3 yet. I'm sure you have. Is it as good as everyone is saying it is? >> Yeah, I I think it's it's to me like um I've actually always been using like many of the Gemini models. I kind of like split my time between like chat GBT and the Gemini stuff. I use chat GBT for like easy, you know, easy queries and I use Gemini for, you know, selling more

complex stuff. Um, and I think like definitely what we've seen is that the the new Gemini 3 models have been able to are much better at following your instructions. If you give them a couple of things, you ask them to do a few things in a sequence is actually able to to follow the sequence pretty pretty well compared to before. So that was very personally um surprising. >> Holiday is unified crypto payments. Give your users more wherever they are. Go to holiday.xyz to never write a smart contract again. Looking for yield that stands out?

Infinify gives you just [music] 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. Yeet 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. Tang um when it comes to actually creating positive outcomes for your sequences of instructions when when uh prompting um a given agent, right? And and you mentioned how these agents are going to be uh much more sturdy and uh have better memory and have stronger performance next year because of the uh consistent model improvement. You know,

you alluded to Open AI committing to scaling their models quickly, which obviously means that everyone else is going to try to do the same thing. When it comes to actually prompting models for the best sort of outcome given a sequence of instructions or a sequence of tasks, how have you found to do it the most effectively? uh either in specific instances or more broadly like what are some of your you know mustd do prompting instruction hacks or tactics that the audience has to know about?

>> Yeah, I I think that's a great question. um you know maybe about like say a year back I think the the the idea and the rage that was um on everyone's mind was around like prompt engineering right and the idea was that hey if you're able to prompt the model well you're going to get even better answers I think that has kind of um been proven somewhat wrong in the sense that if you realize that today like many of the models are able to understand what you're trying to do even without very long specific sort of like instructions and I think that's kind of like speaks to some of the progress that's been happening in the model development. I think for me personally in terms of like using in terms of like

prompting models I think what's important is to understand what you want to produce ultimately what is the outcome define the outcome very clearly for the AI to follow and then all the intermediate tasks I think that is really um up to you like if you know very well what the model needs to do you just give it a very specific set of instructions um tell exactly what it needs to do and then it'll probably be able to do and follow your steps and give the the instructions as you sort of like determine if you are not an expert in that particular domain for that query, you can give it a fairly general sort of like instruction. For example, hey, find me um um like a particular like research on on this particular

sector and topic which you're not particularly familiar with. It will actually go through its own sort of like thinking and reasoning. It will search the different relevant sort of like websites. You don't have to specify exactly where you want the model to be looking at, what kind of like research papers you want it to find. It is kind of able to sort of like curate a lot of that. I think what I find very helpful in terms of like giving me the best sort of like outputs is giving the model the right context. It needs to know like what you what do you want to use it for. Um the good thing is that like try GBT. >> Could you give >> Go ahead.

>> Sorry. Could you just give a quick example of like um a a giving a model context versus not giving the model context and the type of like output you could expect by not doing so. >> Yeah. So for example um let me see like one of the things I do um I use some of the AI tools like for example um I have some of these like Twitter lists of uh very interesting like accounts that I follow but there are two 300 of them and then I I I basically cut the tweets I paste it in the model I ask it to sort of like find me the most important tweets um that are relevant for me that I need to know so that I don't need to filter through all the noise and all the um random stuff and that. So I I give

the model a set of examples of like what are the kind of tweets that I think are important and interesting to me. I want you to tell me um um tweets where um the startup has raised at least 5 million in funding or they're launching a new product mainet is coming up very important events um that they are sort of like running and then basically with all of that context then it sort of like is able to sort of like help me >> um rank and surface out the more important things of that I kind of like want and need right so basically um if I would to just pass it the whole bunch of tweets and ask it to just you know, aggregate all of that. Um, it's going to just try and guess what I might want to see. But if I know exactly what I'm

looking for, if I just give that additional context to the agent, it's going to get be able to give me a much better sort of output output. >> Yep. Yep. I appreciate that. And and I've experienced that quite a bit as well in my in my uh, you know, usage of these LM. So, so you know, you were kind of building on how context helps a lot in this in this realm of of of trying to get better results there before I interrupted and you were you were kind of rounding out your entire thought with regards to the best ways to kind of uh do prompt prompting or prompt engineering even though it's kind of um maybe not as important or was proven to be not as important as people had had thought. So, just if you want to kind of

close that loop there. Yeah. Uh I guess I'll just add a couple of comments like on my own sort of like experiences in terms of using these AI models. Um I think one is that there isn't really any right answer right now. It just comes down to a lot of experimentation. Um for example recently I saw a research paper that said that you know one of the common tricks that people used to use in prompting models was to say um imagine that you are a worldclass um marketer or imagine that you are a world class basically you ask the you you ask the AI model to sort of like role play for you in terms of what you want it to do um and that the soft like research showed that actually it didn't actually improve

the outcome. For example, if you ask the AI to role play as a physicist and then you ask it very hard mathematical questions, it would be the same kind of result as if you ask the AI to be like a lawyer and then answer some of these like questions, right? So, I think we're just only starting to scratch the surface of understanding what's the best way to prompt these models. Um, so I think it's a lot of experimentation, figuring out what works best for you. I think the other thing I've realized is that the different models actually have quite different personalities. So the same prompt whether you use it in chat GBT or Gemini or Grock you could get quite different answers. So it comes down to understanding um which model kind of like works best

for which particular situation. Um, for example, I think Grock has u I I find Grock the most creative in terms of like coming up with his outputs, but sometimes it feels a little bit too unhinged to use in a more sort of like professional um sort of like setting or or answers that I'm kind of >> That's so funny. That is so funny. Um uh do do you do you often check the results of what you do across multiple LLMs? you know I I think I I would definitely love it love to do something like that especially for for the queries that um are more important. I think basic queries I think most models even chat GPT is actually doing pretty well and I trust actually most of the answers that

it gets but um for some of the things some of the tasks they ask it to do >> um sometimes I use tools like there's there's there's something called like yep.ai AI which basically allows you to run the same query across like you know a couple of different models and you can kind of like see the answers evaluate these things and figure out which model sort of like works best. So that's great. I've even seen some systems where you send in a query and then it has like three or four different models um run the same query and then they have another model that evaluates the answers from each of these uh models to give you the final best answer from there which I think is pretty cool. It just takes a bit of time.

>> Have you um have you tried Near's product? >> Not yet. Which which product specifically on Near? >> It's so it's So if you go to Near.ai on your on your computer there, you can you can click in the top right uh uh try it and then you can click uh private chat. Ah yes >> and then click on it and basically it uses deepsek openai a couple other models but it's completely encrypted with TEES and with basically uh some ZK and and it is I think it is the one and only current AI tool that is private today like I I think it's the only one

um and so you get all the benefits of you know using these models the same models that you would use on a normal day, but your data is not leaked to who you know whomever it is. And um you know your information literally could not be given to the authorities if they wanted to because it just it's encrypted in such a way that is so you get all the capabilities of memory of model training data training etc without actually sharing your data. So anyway, they're a they're a partner of the show of the of the experience and you know, I just wanted to show it to you. But I I guess it brings a question, you know, are we

so obviously you've got this like orwellian viewpoint of AI from some people where it's like um you know uh we're entering the endgame here. Uh what's that book? Is it is it um what's the book uh 19 there there's a book uh that people often refer to as a result of AI anyway it's gotten so infiltrated into our lives where people are so comfortable giving away their personal information their financial information their social information etc you know it begs the question of like who owns AI who owns your AI do you own your AI who you know who is in control of the of the responses that you you're

getting. There's a lot of questions that we can kind of like go down the rabbit hole on, but I guess like the broad one here is like Tang, do you think that we have an AI privacy crisis brewing on our hands or are we in a fine spot right now? Is it too late? Where are we at? Like are how how problematic is the current state of the public uh AI platforms that everyone knows, loves, and uses? um and and how does that kind of relate to privacy and some of the I guess blockchain ideals that you know you I'm sure are uh aware of and have been a part of in the past.

Yeah, you've outlined a very great point in terms of like where crypto and blockchain actually plays a part in sort of improving AI and that's really around um privacy um today for sure like if you look at if you're just using like Chipd and Gemini basically um you know it's collecting all of this information and data um from you um the thing is that most people don't see it and because the data right now is just being used to train their next version of the model so it's something that's happening on the back end so you kind of like feel it. But I think like what's going to happen um pretty soon and we're already seeing some of this with the chat GBT sort of like shopping features that they're going to start to with all the data and the queries that you've been putting

into the model. The model is going to start to have a very good idea of who you are, what your interests are, what your likes are. And naturally, as we've seen with every kind of like business model on the internet, >> we're going to see them start to sell sell stuff to you, right? advertising is has always been the model um that most of the internet has sort of like thrived on. So that is where I think the consumer is going to start to feel a lot about where this privacy starts to become very important because it becomes very vis visceral to um them and their lives out there. So yeah, no I think anything that that can be done to sort of like protect that um is great. I think um people don't kind of like care too much about privacy mostly right now

because the value that the AI gives to them is far greater than I guess like how they value their privacy. So that's kind of where it is. But I think like for many of us I think it's definitely a concern. I think in this day of age with so many um in a very chaotic kind of like world I I'm starting to feel that privacy is a very important thing and you know very cool stuff that NIA has been doing on that. >> Yeah. Yeah. For sure. Um yeah interesting. Yeah. I mean I think um the whole like legal obligation that OpenAI has to give your information to the authorities no matter what is is pretty insane.

Um I mean that's kind of what any company has in the world or at least in the states. Um >> yeah I I think just onto that as well is that >> you know we only going to figure out what happens when this really gets tested right like when a government actually makes that request that comes then we'll know whether these companies are going to succumb whether they're going to actually share all of the data. >> Oh they definitely will. Oh they for sure will 100%. You don't think they will? I >> I I think they I don't know man. I think some would maybe would try to fight a little bit but uh maybe it's just for show but ultimately right if the government says they want it I think it's really hard especially if your whole company is is in the country.

>> Yeah. Yeah. Yeah. Um cool man. Uh so let's talk about some of the agents that you've seen um excel like uh is there um so look let's just lay the context here as if you were an LLM. the there there's all these agents that we hear about. Salesforce has got their commercial for an agent and everyone's got these agents and I'm just I'm never going to book a flight again because the agent's going to do it for me and like why would I ever book a flight? I'm I'm way too cool for that. I'm superior. You know, why am I going to ever make a reservation for dinner because, you know, I'm going to have an agent do it and they're never going to make a mistake. They're not going to book the wrong restaurant in the wrong city

because they hallucinated. That would never happen, right? And and so we've heard all about this. Uh but to this day, I have no agents that are doing anything for me outside of like prompting back and forth, creating images, videos, um or just generally like giving me like basically what I'm getting at is I'm missing the uh here's the instruction. I want to get information. Okay, now I got the information. Now go do this for me. and like that go do this for me task outside of like whatever their capabilities are from like generating an Excel sheet, generating code, generating an image video, editing a video or c you know giving me like the right clipping etc.

like the actual action of going to do something in the real world um if you will like not within the its own capabilities I I haven't really seen um you know give us an update on the agent landscape where we at um it will 2026 be the year of the agent actually having capabilities as I mentioned to go outside of the internal kind of like uh LLM to human interaction and actually go do something and kind of report back whether that be onchain Whether that be book a flight, whether that be uh email a friend, where when do we get the actual uh fulfillment of this promise of agents doing all of our work and these types of things?

So you know I think u I guess a couple of thoughts on that is that firstly um you know we already seeing like AI agents you know be good at certain things and for example things like research like being able to um synthesize um digest a huge amount of information and present the things that are important um you know agents are getting good at coding I think that that's because like you know code is very clear. We know whether it works or it doesn't work. It's kind of easy to tell what's good and bad. Um so we going to see a continual like improvement um over the next you know couple of months, one to two years of agents being able to

do more and more tasks in a better way. I think one of the main um hindrance I think for AI agents to be able to do stuff like you mentioned being able to book the flight for you and do everything in autonomous autonomous way um is that there is still a bit of that reliability issue where let's say if the AI hallucinates maybe like 1% of the time or maybe 2% of the time which is really pretty good which means like 99% of the time it still gets the task done but then if you chain that over like multiple like different tasks and steps that AI agent needs to be able to do. He needs to search the different websites.

He needs to pull all the information. He needs to go to the booking page. He needs to key in the payment details, right? If you chain all these together, then it comes you can see that over a series of actions, right, the the failure rate um becomes higher and higher. So, I think that has still been the problem. Um and nobody wants to use a product where you know that you're only going to get the the outcome you want like um 50% of the time or 60% of the time. um we will get there. Things are going to start to improve um faster and quicker. Um and I'm I think the other thing that also we've lacked a little bit is around the data side of things.

Um for example, if you've tried out some of the newer sort of like browsers that uh or or computer use agents that like OpenAI has kind of like you know used, you find that um it's very interesting. It's able to sort of like understand what's on your screen. is able to click buttons for you. But uh it's still pretty bad, right? It kind of like gets stuck. It's very laggy. It doesn't work. And a lot of it is because there isn't that data set out there that shows a human sort of like doing his work and clicking all these different buttons on the screen. That's an entirely new data set that has to be sort of like gathered in order to train the models to do some of these things. So the big research labs are still continuing gathering more

and more of this data. And I think at some point in time when there's enough of [clears throat] that data, all of these things will actually start to improve uh pretty fast. >> Trade on anything, [music] economics, weather, sports, crypto, and more. All on couch.com 24/7 for the world's [music] biggest prediction market. >> Havi is fast and private per trading secured by ZK with Celestia underneath. [music] Go to habachi.xyz to start trading today. Not your keys, not your [music] 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. [music] Get a treasure at treasure.io. >> So, the agent problem is a data problem is what you're saying. >> It's a data >> it's a training data problem. >> Correct. Yeah. I think data is always context probably context right if you need to if you can feed the agent with the right type of data um then I think it's going to be able to do things that you want in a much better way. So, so do you anticipate like middle of next year to see this agent booking you a flight idea h come to life?

>> I think we'll probably see something like that. >> No, I think 2026 will be that year. I I kind of like think of 2026 as the year of the AI agents like we're going to see some of these um use cases really come to come to life at least given the pace of progress so far. >> Yeah. Beautiful. Okay, cool. Um kind of completely different topic more on the on the more on the morality and economic side of these uh AI massive massive Goliath companies like OpenAI who's trying to IPO at a trillion dollar valuation Nvidia at five trillion or something or two trillion or maybe it's two and a half two um you know where where do we draw the line on forprofit AI companies and for the betterment of

society and how these things kind of come together. We saw this we saw this all kind of happening with the internet and you know Elon Musk buys Twitter to try to uh create a place for free speech because he thought all the left all the left wings people were taking it o over Twitter and you know all these different forms of social media now Rumble and these other different kind of like interesting little pockets of media. It's also going to be happening with AI. Uh Disney has also just announced they're they're going to invest a billion into OpenAI. So looks like OpenAI is just taking checks from whatever rich friends that they want to have a part of this this pie and integrate their stack into their world.

You know, I could see an AI oriented Disney World soon sometime soon. um think you know I don't know if you consider yourself uh you know a a capitalist a socialist a narco capitalist a neoc capitalist a communist I don't know what you are but I'm curious where do you draw the line between profit seeking for AI companies and um basically human morals morality and not trying to just farm everybody as if they are statistics on a on a spreadsheet Chat GBT is talking about uh introducing data uh that will allow them to uh actually advertise to you in the chatbt

interface. Um this is how a lot of companies have monetized in the internet. There's a lot of push back on that. How do we where where we at? What is your mental model for the state of the for-profit corporations that exist uh which own a lot of the AI superpowers? No, I I think that's a great question and I think it's something that um you know a lot of people are also trying to figure out what's the right balance around some of these things. Um my my general viewpoint is that um free markets are good for innovation. I think we've seen that sort of happen, right?

Um and a lot of the great like um tech innovations be it like self-driving cars um have come up of like um startups and sort of like private capital and market funding a lot of that innovation um that's come up. Um AI I think everyone can kind of like uh agree that this is probably one of the most important technological developments for humanity as a species as a race. Um the thing about AI that makes it a little bit different from some of the previous tech innovations we've had so far is that AI is extremely um expensive, right? We need to build a lot of data centers which requires a huge amount of like power, compute, memory and chips in

order to continue to progress. Um and and so that money kind of like needs to come from somewhere. And the way that we going to continue to be able to fund these bigger and bigger training runs and data centers that will produce these better models, it's through, you know, both a combination of private investors like putting money into a company to fund it, but they put the money it never money never comes without like strings attached, right? So there's always expectation of a future profit, future return um for the monies out there. So I I think that model will will continue. I I think there's nothing overtly wrong with that. I think what we do need is some form of like check and balances to

make sure that no one monopolizes this hugely important technology for humanity. Um the good news is that um because the entire world has recognized that this is such an important technology like everyone's eyes are on this. So, so far we haven't seen like uh a single company that seems to be that dominant player that's going to rule over the world because it has the most powerful AI that nobody else can kind of catches. We have seen like quite a number of different players have models and AI that are at almost like similar levels of like capabilities and progress and I think that will sort of like continue to sort of like um happen. So at least today I think um I'm not too

worried about the ethics of these um for-profit like companies. U at least they seem to be you know I I think we're in the sort of like early days of this whole buildout where you know companies are making these AI models available for free. They're not even charging consumers to use some of these things. ChatgBT's like subscription is pretty cheap relative to amount of value that you get. Yeah. So I I think um it's still in a good place for consumers, but if we get to a point where you know these companies start to charge very pricey fees for you to access this AI or they limit it to only very select group of people, I think that's a time that I would start to get very worried. So SpaceX plans to go public at a $ 1.5 trillion valuation in

2026. The Google CEO is talking about building data centers in space. You made a really good point with regards to the way that the structures of these VC Silicon Valley led rounds create this need to kind of build out efficient technologies, create innovation, create efficient markets. In addition to that, uh investors are looking for returns. Founders are looking for exits. Um as long as we keep a baseline of democratized access, I do agree that I think that uh we're still in a good spot as as as far as um kind of outlook goes.

Um but talk to us about this data center uh hype like this data center rush. It's like a space race happening right now. Chimamoth Palipulia just announced he was doing building one out in Nevada I think uh like about a year ago or so or six months ago. Talking about building space centers or data centers in space. Is this a bubble? Tang. Is this a big freaking bubble? Are we in way over our heads right now? Yeah. So, you know, I I think everyone kind of has the eye on the big prize out there, right? Which is some form of like super intelligence. And the way that everyone believes that we're going to

get to this super intelligence is by throwing more and more compute at it. So, and that's why the data centers are getting bigger. you know this is this is probably the biggest experiment that humanity is sort of like conducting today because we don't know when this um when this sort of like scaling law will stop right is there a point where we've hit a certain limit and even if you build a better and bigger data centers and you throw more chips you throw more compute at the AI we are not going to improve it any more significantly right we hit that sort of like point in time point we don't know we only will know once we actually get there and we see it and if we actually see something like that where we actually not able to scale the level of improvements of AI

alongside scaling the data and compute then I think there's a good risk that the bubble might sort of like burst at that point in time because people realize that actually we won't be able to you know all the additional spending on data centers may not work but until that point in time um everyone just has to play the game it's it's it's like musical chess right you just have to keep walking you have to keep building you have to keep racing to build the biggest data center and accumulate as many chips as you can so that you have a seat in this game and you're not sort of out of it especially if you're one of these like big tech companies that want to maintain it sort of like market share and dominance in this uh world. Yeah man that's uh that is uh an interesting

picture that you paint here Tang um you know you mentioned some about uh about some of these uh opensource kind of decentralized inference networks that you like like bit tensor I'm curious like what signals do you are you looking for that indicate real traction in this market right like is it cost efficiency model diversity uh specialization techniques like when you're looking at this whole kind of blockchain AI sector, what are you looking at right now? Where are you seeing actual traction and where should our you know investors in our community actually look look to uh you know do more research into and you know what should we avoid?

>> Yeah. So I think the way I would evaluate like companies who are building in the intersection between like crypto, blockchain and AI is is pretty much similar to like how you evaluate a startup like a tech startup. You look at uh users, you look is there any sort of like product market fit? Are people using the product? Um or if you're building like infrastructure protocol, are people building very cool applications on that particular protocol? Um overall I would say that we are still in the pretty early phase I would say of u crypto and AI. Um a lot of teams have been focused on building infrastructure whether you're building an infrastructure for agentic payments

you're building infrastructure to support decentralized like compute or decentralized training of AI models um you know and once that infrastructure is in place and some of these teams are starting to get these like um protocols launched and on mainet we're going to see hopefully we'll see people start to build on this. So I'm looking for that. I'm looking for developer activity. I'm looking for users on the platform. Um I'm looking for some of these companies to be able to build a sort of a sustainable business model and actually be able to generate revenue. U because AI is not cheap. Like if you are running inference, it costs quite a bit of money to run um an application that uses u or provides the user with a lot of queries.

Um I think Bit Tensor is interesting. >> Sense. Yeah, Bit Tensor is interesting because to me I kind of see it almost like an incubator or an accelerator. Um, so like by buying into Bit Tensor, you are sort of like um buying into say 60 to 100 different AI startups that are trying to do all different kinds of things. U most of these subnets and startups will probably not make it. They'll probably fail at some point in time. Um but hopefully there'll be a few of these that will actually get to sort of like very sizable like revenue and business models and we actually starting to see like some of these subnets on Bit Tensor um like shoots are actually like you know producing quite a good amount

of like revenue uh which is subsidized by the TOA token of course. So um yeah I think the economic sustainability still needs to sort of like be thought about. >> Yeah. Yeah. Such as the way with all the I would say 99% of these crypto networks the economics sustainability still needs to be continuously improved on and iterated on. Tang Yan appreciate your time man. Absolutely. Episode absol absolutely epic AI supercycle episode today. Thanks for your time. Thank you for coming in answering some hot questions and having some fun, man.

would love to have you on in 2026 as we continuously explore the proliferation of this technology and the acceleration and improvement. So, thanks for your time man. >> Yeah. No, thanks Andy for having me here. Like, um, this was fun and yeah, looking forward to chat more. 2066 is going to be a great year, I'm sure, for AI. >> And Tang, where can our audience find you? Chain of thought. Um, how can we how can we support you? >> Yeah. Um, you know, I write a weekly newsletter. You can find it at chainofthought.x. XYZ and I'm pretty active on X and Twitter as well. So you can find me at Tengen AI, I believe.

Awesome Tangen.