Claude Opus 4.5’s Breakout Moment & Investing in 2026 with Qiao Wang

Okay, here's what I realized over the holiday. It's exactly what you described. The line between general chatbot and coding is blurring.

Now when you code, you don't actually write code, you write plain English. So the coding assistant and general chatbot are merging. The software mode is diminishing very rapidly, and for early stage startups, there's basically no modes.

But if you look at the likes of Facebook or Google or Microsoft, Apple, the modes still exist. The AI coding Agents won't kill them.

Nothing said on Empire is a recommendation to buy or sell any investments or products. This podcast is for informational purposes only and the views expressed by anyone on the show are solely their opinions, not financial advice or necessarily the views of Blockworks. Our hosts, guests, and the Blockworks team may hold positions in the companies, funds, or projects discussed.

All right, everyone. Welcome back to Empire. Very excited about this one. Got good friend, friend of the pod, friend of me, Ciao from Alliance. Welcome to the show, Ciao.

Great to be here. Have I been? You come on. I've been on other Blockworks pods. Yeah.

I have an embarrassing thing to admit to you, Chad, which is I've changed my scroll. I still scroll social just like any normal human being, but I only use list now. I don't do the algorithm. And I've got a list and it's just called brains.

And I was like, "Oh, this is really enjoyable." Like I really enjoy my consumption behavior now. And I was like, "Who did I put on this list?" It's like you, this guy Will Manitis, and Chris Berniski. There like three people on the list. I was like oh that's why I enjoy it.

So I'm just basically telling you I've consumed a lot of your thoughts and tweets recently.

Appreciate it. I started doing the same about half a year ago just to get rid of all the I don't want to see all the political stuff, all that. Like now it's pure either markets or tech. I have two lists, markets and tech. Yeah. I have health and brains.

I mean it's interesting. I had a crypto top 200 list that Blockworks made and I curated that whole list. I looked at the top 200 list. I made it probably a year ago. It's all outdated. All the people who are big names, the thought leaders, I'd say like 150 of the 200 have essentially washed out some version of washed out. It's very interesting. Yeah.

I want to talk about Claude and Opus 4.5 to start. You've been just talking about it a lot. I think maybe the best place to start is you there's a tweet, speaking of tweets, I'll mention a lot of tweets in this episode. You said the holy it's over moments for me in the last few years.

One was the release of chatbt. Two was the first reasoning model M01. Three was the Tesla FSD V13. By the way, I just used that. I think it's now the V14. I just tried that for the first time.

And I live in a parking garage with very tight parking. I was like, "Oh my god, it got me out of that. Drove me all the way to New Jersey." So that was crazy, too. And then you said Opus 4.5. Why is Opus 4.5 such an holy moment here?

I can't explain to you what's happening under the hood because I didn't build Opus, but I can tell you what I felt as someone who used it. The last time I wrote any sort of code was maybe a year ago, but I really seriously only coded. It's been like three years I haven't really seriously coded.

Give me your background before we go too deep to just contextualize your coding ability.

I would say between 2010 and 2017, right before I got into crypto, I was in quant trading and I spent all day coding both the super low-level C++ the actual trading engine as well as the higher level Python data science stuff and that's my work for like for seven years all day is coding and I wrote some front-end code a little bit of front end code and backend.

Ever since I left Masari, I haven't coded for seriously coded for three years. I did some personal projects here and there like maybe at the end of the year for every year the last three years but like Opus 4.5 like everyone has been talking about VIP coding but the big difference between Opus 4.5 and before was that before you can sort of put together a demo very quickly but the last 5% of the project always like it's really hard for the AI to do it for you like completely it still requires a really good engineer here like a human in the loop to get that last mile done.

Like the last mile meaning like the bugs and the corner cases, all that stuff. And with Opus 4.5, what I felt was I just tell it what I do in plain English. I obviously I have to give it a very clear spec. But as long as my spec is clear and comprehensive, it can just do it in it canes one-shot it. And that's how I felt the biggest difference was.

Of course, I think Claude and Opus might be going through this hype cycle on Twitter because my engineer friends told me that for example, OpenAI, the latest OpenAI GPT 5 Pro is as good or almost as good as OP. But I've used both and I actually opus is the first time I've switched out of chat GBT because for me chat GBT the memory of chat GBT had locked me in so intensely that I was like I guess I'm a chat guy for my life like it's it knows everything about me and now Opus 4.5 is so to me it's so much better.

I know the models I looked at the model ranking but to me Opus 4.5 is so much better as a chatbot or as a coding assistant both actually. Yeah. Okay. I haven't tried Claude as a general chatbot, but just on the pure on the coding part, it's a huge jump.

There's a level of work that I don't know if it's coding or chatbot, but it's somewhere in between. For example, let me tell you a thing that I just did with Opus 4.5, which is we have a TAM of Blockworks, right? And it's always a headache. We have an eight person sales team at Blockworks.

And it's always a headache to be like, who's which account should we sell to? Is that a tier one or a tier two? Each seller has a determined like, you know, you get 10 tier ones, 30 tier twos, 100 tier 3s, something like that. It's always a total headache.

I just put the problem into Claude. It pulled in and I was like, look, if they have an FTV, if the token has an FTV over a billion, that's a, you know, that that gets extra points. If they're a public company, that gets it pulled in everything. It pulled in info from Coin Gecko, Coin Market Cap, all this different data and it ranked the list and then it looked at every deal the sell each seller had closed over the last eight years at Blockworks.

It looked at the products that we sell and it then assigned basically accounts to each seller based on the likelihood that they would close a deal with that account. That was that is co that is that's not just chatbot work. It's not coding.

Okay, here's what I realized over the holiday, it's exactly what you described. The line between general chatbot and coding is blurring, right? Because when you talk to a general chatbot, oftentimes it writes some code on the fly for you to do the thing you want to do.

Whereas now when you code, you don't actually write code, you write plain English. So the coding assistant and general chatbot is they're merging. It's one. Yeah, it's merging. Yeah, exactly.

Why don't you use Replet or Lovable for your vibe coding? Why do you use Claude?

I feel like I think they're targeting different persona. I haven't tried the latest versions, but I feel like lovable is generally targeting people who want to build a really good demo or maybe a simple app. I think they're really good for front end. That's my impression. I haven't used it myself. I think they're really good for front end.

But the tool I wanted to build for myself was mainly a backend thing and so I used I started with cursor plus opus 4.5. Then I tried cloud versus with opus 4.5. I actually couldn't tell that the like in terms of productivity boost I think they're about the same for me. But that's just my personal experience. Yeah.

Talk to me about how you so for people who don't know you've been running alliance where you were I mean helped build Msari then launch Alliance you guys are on cohort 17 I think or 16 we're going to be on 16 16 you've seen thousands of crypto startups now at this point how does this impact companies? Yeah.

In a huge way and this has been happening for the last three years which is that basically almost every cohort I ask people especially the technical co-founders of every startup what you know how much productivity improvement have you experienced since Chad GPT in 2022 and basically every cohort I get a number that is bigger than the previous cohort like It's it's really consistent and the latest number I got from the last cohort was like some somewhere between four to three to 4x product improvement.

Of course I think this probably impacts early stage startups even more so than later stage like than bigger companies in terms of productivity improvement because I think fundamentally the limitations with coding assistance AI coding assistance is the context window.

So for example for argument sake say like let's say cloud can take in a million tokens then what that basically means is it can write a it can start a project very easily and do it really really well but if you ask cloud to change the entire Google codebase it's it's not possible right so it probably affects early stage startups even more so than the bigger companies I think what what's happening within the bigger companies is the way they can best leverage coding assistants is they they have clear layers of abstraction between different different divisions between different different like departments of of of the company so that they can apply so they can shrink the context window and feed it into the AI system if that makes sense. Yeah, it does.

I've seen that talking about AI in terms of productivity is the wrong lens to view it from with the earliest stage startups. So I think if you're a company 100 employees, 200 employees, a thousand employees, what's your productivity gain? 3x, 4x, sure.

I think that the earliest stage startups don't even understand what they don't think about it in terms of productivity. They just think about it as like why would we ever go hire more people? That's what I've seen in the like, you know, companies I've angel invested in and like friends are running like new startups and stuff like that. They're just like we we don't have to hire. It's really bizarre.

Or we you know even this thing of this target accounts example that I was doing with sales the other two things that I just built are a sales commission calculator sales team's always like how much am I making like then you got to ping the finance team and then finance team's like oh my god like I got to calculate the sellers's commission like look how much money they're making like that's a pain in the ass and then the other thing is how much does a for us we sell these data dashboards right so it's like how much does a dashboard cost then they got to bog down the data team and the data team's like h well like are they indexed by Dune or gold sky it's like no okay well But we got to add a little extra money there. I just built a dashboard calculator based on all the.

So so it's like these little And so what what does that actually mean for us? We don't have to go hire a sales ops person. Yep. Right. So it really does feel like a one or two person unicorn is possible in 2026. Yeah. And such a startup probably has already started. It's not at a billion dollar valuation yet, but it has started 100%. Yeah. Yeah.

I mean, I know multiple people who are running $10 million ARR businesses by themselves right now. Yeah. Yeah. They're often like exineers from Meta and Uber who got sick of the bureaucracy and they're just like Yeah. And they're just cooking by themselves.

There's an interesting thing though I've noticed with AI some of the best AI AI start not AI startups like a chaty like an open AI but like these little businesses that are leveraging AI is they don't want to tell people what they're doing. It's one of the first times I've seen an usually if you have a startup and it's ripping right ar through the roof you want to shout to the world you want to go raise the money you want to go tell everybody you want to tell post on Twitter people are very secretive about what they're now it's it's too easy to copy your software it's too easy yeah exactly how do you think about moes then for companies and do does that definition of a moat change it doesn't change I mean the the software mode is diminishing very rapidly and for early stage startups there's basically no modes but if you look at the the likes of Facebook or Google or Microsoft Apple the modes still exist the the AI coding agents won't kill them.

And so the modes for these companies for example it's developer ecosystem for Apple and and maybe PC like Microsoft it's switching cost for cloud AWS Azure Google GCP it's proprietary data right? Like YouTube, YouTube owns a lot of video data. That's how they're able to build a really good video model in the first place. Like enterprise integration and like super mission critical software like that's Microsoft.

Sure you can copy office but you know are you are enterprises really going to switch from office to to a new you know software? It's it's too mission critical for them.

Actually I'll give you one example I've been looking at recently. it might be one of the biggest inefficiencies or mispricing in the public market right now it's Adobe you know it's Photoshop and and you know the creative suite of Adobe the market currently thinks that the latest video models and imaging image models are going to kill Adobe I think that's completely wrong because Adobe's mode is the enterprise integration like all the enterprise users of Adobe creative suite software they're they store their images and their videos on Adobe's cloud and their creative professional like the switching cost of moving from Adobe cloud to somewhere else is very is very high and the fact that they've been using Photoshop for so many years like the it's it's muscle memory right it's very hard for these people to to switch and and so Adobe right now is trading at 124 PE is crazy for for for such a for such a high quality of a company.

Is Scott Bellski next in line to run that company? Have you seen because that would be that would he he's a phenomenal entrepreneur. I'm not sure if he's still in line, but that would be very interesting.

How do you think companies should like for what's the let's say you're you know so there's a lot of like you know very maybe young entrepreneurs who are like this is kind of second nature to them already. What if you are at a company where it's you're listening to this and it's not second nature? Like maybe you work at a I don't know 500 person startup been around for like a decade and you're just like yeah no one's really maybe people are like putting things into chat GBT and sending like automated emails but like how do you how do you push a company onto this like next wave of what obviously needs to happen?

I don't think you can force them to adopt a new technology for the sake of it. I think they have to just figure out where a techn like AI can help them immediately and make a huge impact. By the way, that's that's what's has been happening within alliance for the last three years. We didn't, you know, we knew AI three years ago. We knew AI was going to be a big thing. We said we're going to become AI first, but we didn't like try to force AI into every part of our organization, but instead we used AI, for example, to automate some of our admission process.

Like we get like thousands of applications. Like three years ago, I used to read all every single one of them myself. 5,000 applications per year. It's brutal. It's soul crushing. Now I'm able to automate about 50% of it using AI.

So how do you how did you actually do that? Did you So is that an Opus 4 point? Like how did you how do you how did you build that? How did you put in the inputs? How good is it actually on the outputs? Why is it only 50%? Why isn't it automating 99%?

So I didn't write it myself. The software was written by one of our engineers. I think the coding part is no longer the bottleneck. the prompt is is where the I wouldn't necessarily say the secret sauce because everyone knows what a or for the most part what a great founder might look like on paper, right? So I just translate my heristics of what a good founder or a good startup might look like and put that into the prompt and I use that to to filter and what and right now the what what that piece of software tries to do is filter out the obviously bad ones rather than tell me which ones are obviously good because I think that that part like it still requires some taste taste from the human interviewer.

I think the AI isn't quite there yet, but I think you could get there like by the end of the year. Like, yeah, I by this end of this year, I think AI and AI VC is going to do a better job than than I do. Yeah. Yeah. That's interesting. That's I think the Revolute founder launched a venture firm and that's what they're doing.

I mean, this was Chimas's idea with Social Capital if you remember. Right. It was and then it didn't didn't work. I think maybe AI was too early. It was too early. The tech wasn't ready. Yeah.

Very interesting. How much how much value do you think you get out of the like how much would you pay? How much? So to me it's shocking how cheap these are and it's you know it's kind of like taking a $5 Uber in New York from like you know Fey to the Upper East Side when Uber launched. You're like okay obviously this is mispriced. But they're just throwing venture dollars at it to onboard users.

How much more what do you think that mispricing is? Like how much more would you pay for Opus 4.5 right now?

Opus. Uh, I'm not quite sure yet. Depends on what I build with it, but I'm not quite sure yet. But for Gemini, I think it's underpriced by two orders of magnitude because right now I'm just paying $20 for a Pro. I'm not even using the the Pro Plus or You would pay $2,000 a month for Gemini.

Yeah. Be because the the the the it's a research like I actually tweeted about this. It's a research assistant. It's a junior research assistant. junior coding assistant. A a pretty good health adviser, like medical adviser to double check whatever my my doctor tells me, a pretty good parillegal, and so on. So, and a therapist. Like, if you combine all these things into one, $2,000, I think, is underpriced.

I think Google knows this. Open eye knows this. Yeah, they know it's being mispriced. Yeah. How does your wife feel about you putting every single thing into into Gemini? We both put a lot of our things into junk in separate accounts of course. Yeah.

How much time do you spend a day on how much time do you spend a day on some combo of Gemini and Claude these kind of things? And has it where has that time taken and been taken out of? Is it out of Zoom calls? Is it out of social? Is it out of Slack?

Yeah, I actually analyzed this maybe half a year ago and which led me to to to to the Google trade. Basically at the bottom u what I was unbelievable call by the way like you were the only one longing Google at the time and now it's you know I I I looked at my my my iPhone usage the top three apps I was using Chrome, YouTube, Gemini, they're all owned by Google. Yeah.

And of course I I I was worried that chat GPT was going to kill the the search business. So I talked to my wife and my wife gave me an insight where she said she does all her Google the mo the biggest use case of Google for her is shopping and that part was at least in the short term was not going to get killed by chatbt at least at that point in time. So then I did some research. I realized actually more than half of Google's search revenue comes from shopping. I was like this this castle is is intact. Yeah.

No one's going to kill it anytime soon. And then on top of that there's the GCP, there's TPU, all these things that people realize later. I thought that was a no-brainer. So Google was my my biggest trade. It was like the only big trade I took last year. Did you size it? Well, I put like maybe 10% of my work into it. That's great. 50%.

So, yeah, that's great. Yeah, I got dinner on New Year's Eve with someone from Google and we're, you know, we started talking about these things and people talking about what they use and the Google person's like, "What all you guys misunderstand is we've got the shopping data. We have the shopping data and no, the market is not still does not understand that yet."

So I think Adobe might be the the Google of this year because it's so cheap. Like it's experiencing the same thud that Google experience. I don't I just don't. But like when's the last time you've used Adobe? Never. But no, no new people will use Adobe. That's my problem with it. Like Google I am the same. My top apps are like you I'm on YouTube like YouTube. I use YouTube so much. I use Gemini. All these I don't use Adobe. So all the new people don't use Adobe but all the new people use Google.

I think what the market misunderstands is that Adobe is not a consumer product. It's enterprise product and but my enterp like quoteunquote my enterprise would never adopt Adobe. So what what would they use? Figma or Figma? Yeah.

So Figma and Adobe they're actually targeting two different market segments. The the war is actually over for that segment of developing like new new demos and and you know website interesting like that that war is over. Adobe lost that war. Figma won. Uh Adobe is more then what does Adobe win? Like the the marketing people um but Canva Canva is for lower end segment of the market. It's it's more amateur versus like true high-end enterprise.

So I agree with you. Adobe's net new users has not grown for a very long time. But but they're able to char their pricing power is so insane that they're able to charge more and more every year. Did you put on the trade yet? Yep. Small one because I I don't use it personally. I I don't have the same amount as Google as Yeah.

Well, I want to get into investments in a little bit. But I want to talk about this idea of like where you should be spending time. Yeah. How do you think about how AI changes what you should spend time on?

I think everyone can and should write code and if they don't, they're going to be left behind. And by writing code, I mean specifically, automating some part of their own life or their own work. I think what's going to happen is or what what might I have no idea but what might happen is there's still going to be these like B2B SAS software that everyone purchases and every company uses right so for example we use I don't know Gmail or or whatever or Zoom and we pay for these things but there's always this long tail of workflow that is unique to yourself and to myself that it doesn't make sense for third party software providers to build it because there's not enough money for them to make. But you and I can build a commission calculator at Block Works. That's a good examp There's not a SAS business to go calculate commissions. But yeah. Yeah.

And when I say like everyone should be coding like it shouldn't sound scary anymore because you no one's going to write any code anymore. You just talk to the thing pure natural language. I would highly recommend anyone listening to this don't even use claude opus because there's it can give it it's too much. just use in 10 or or something. Use use Replet for the love of God. Go on Replet. It is that was to add to your you know FSD V13 opus 4. For me it was Replet. Replet was my moment where I was like oh my god AI is not just a chatbot where I talk to chat GBT and it gives me some nice answers. I was like you can build anything. The whole world is going to completely ch I would really recommend people try Replet. It's amazing. It's very cool.

Why how do you think like the people who use AI in the workplace like so I'm kind of envisioning it like this people are watching on YouTube so they can't see but it's like you know people who use AI people who don't use AI and then it's just going to go like this. Yeah. Yeah. How will that show up? There will be a K shape of the workforce. Can maybe just talk to me about how you see that happening?

I feel like it's going to be similar to to the internet. Like these things just make the high agency smart people smarter and the low agency people dumber. It's just an incredible tool that's out there for available for you to use and it makes you more productive. So, it's really a matter of how much you want to use it.

This episode of Empire is brought to you by VanX onchain economy ETF Node. If you feel like you are constantly missing alpha, never have time to scroll through crypto Twitter, Node is for you. Node is actively managed by an experienced team with years in both trady and crypto. That team at Node looks to capture the most powerful trends across all of crypto while actively managing volatility as markets evolve. Super important in today's environment. Check out Node's track record so far at van.com/mpire/node. That's van.com/mpire/node. You can also little easier just click the link in the show notes. Thanks to Van and Vanex onchain economy ETF Node for making today's episode possible.

Mantle is launching the Global Hackathon 2025 to accelerate the future of real world assets. With a $150,000 prize pool, backing from a $4 billion treasury and direct access to Bybit's 7 million plus users, this is the ultimate ecosystem for builders. Mantle has entered a new phase as the distribution layer connecting Tradfi and onchain liquidity. To accelerate this vision, the Mantle Global Hackathon 2025 is inviting developers to build scalable realorld assets and DeFi products. By building on Mantle, you get direct access to Bybit's 7 million plus users for potential listing exposure, support from the Mantle Treasury, and mentorship from top VCs like Spartan and Anamoka brands. With six tracks prioritizing realworld assets and real fi and a $150,000 prize pool plus grants. This is your chance to deploy on a high-performance modular layer 2. To register, click the link in today's show notes. Thanks to Mantle for making today's episode possible. As always, investments in blockchain technology involve risk terms and conditions apply.

Is there some sense that it's a to play devil's advocate, is there any sense that like you built your Warren Buffett stock tracker? Y took took you a lot of time probably. Was that a good use of time?

Yes. The the the end result was okay, but the the process of building this tool helped me understand this this technology better.

What did you build? Tell us about it.

So it's a it's a digital clone of Buffett and Munger and Howard Mars and Draen Miller and all these people that scans on a regular basis thousands of tickers and it it conducts a a deep research on every single one of them and it tries to emulate the mental models of Warren Buffett and and Mar and so on. And so the prompt the pro the coding itself was very easy but the prompt is took me several months of iterations. So I I actually have like very detailed prompt to emulate how buff and monger think about evaluating a a potential investment. So it's a six-step thing.

And and so it uses deep research to gather information for every single one of these six things that they would potentially care about. And then I run another API call that uses the reason model which is separate from the deep research. Deep research is really good. And what by the way this one thing I learned is deep research is really good for gathering facts and data and reasoning models is it can hallucinate a lot but once you give it the facts and data you can reason much better than than deep research.

Then my second round is a is a reasoning model call to make a decision on whether or not we or whether or not the investment committee consist consisting of the digital buff and a monger would actually like what investments they would make and so and it gives me the output. That's the whole.

Have you heard the theory that you know Renaissance Rent like have you heard the theory that Renaissance basically discovered LLMs before everyone else did and they just kept it a secret and that's why their returns have been so high. It's possible. And the the reason why I did Buff and Margaret is precisely to not compete with Renaissance because those guys are really good at short-term trading, right? like with a horizon of maybe daily or or weekly horizon time horizon. Yeah.

Um in that at that time horizon there's no zero chance the latest AI models are going to be you're not competing against Jump or or Suscoana or say like ST or Yeah. Yeah. But at a multi-year time horizon no one no one has the patience to to invest and and hold it for like more than five minutes these days. So yeah, you can leverage LLM to to.

So then whose model do you pick? Do you go with Howard what Howard Marx is saying to do? Because what Howard Mark is saying to do is very different than what Duck is saying to do, which is very different than what Buffett is saying to do. Exactly. So, who do you go with?

I average them out. Ah, they all say Adobe and you long Adobe. Exactly. And here's the other thing. If you run the same prompt multiple times, they'll it'll give you different answers. Yeah. So, if you run the same prompt multiple times, you average average them out. If all of if every single one of five runs tell you tells you that you should buy it, then you have pretty good conviction that it's it's a good idea.

What's another stock that it says to buy? The I think as it output about 10 stocks and four of them are actually in the current portfolio of of Berkshire Hathway. So it has but isn't that obvious because it's pulling in specifically tell the prompt do not use in any circumstance the current or past portfolio information of Bergkshire Hathway. Ah okay. So what's what's a stock?

Chub which is an insurance company. Chub. Interesting. Yeah. Um Google. Yeah. And there's one more I forget.

How are you uploading the data? How do you like how do you because I I built I built something. So I built I have these things called cluster headaches and there are these like gnarly Yeah, it's a head thing and there's all this new stud there's always new studies and research that come out about it but I don't know how to read like studies that well and so I always upload my studies and then I was like all right I'm just building a site.

So I built a site it's like cluster headaches.co I think it is. And it pulls every single study I I put the study in and any article I want in. And then it uploads it and starts to build a database and it and it starts to actually build a like a really nice user interface where it's like, oh, it'll pull in like, oh, this medication. Oh, this medication was mentioned seven years ago here. Oh, that's tied to like this thing in the brain. Oh, this thing in the brain is tied to because like I was trying to do all that basically like that meme of like, you know, the wires crossing on the whiteboard type of thing.

I was trying to do that in my head and like on all these notebooks and now this website does it, but I have to manually upload the studies. How do you have it pull in automatically like the studies and research and papers and things like that?

I I I tell the the deep research agent in my prompt that you need you can only use the following sources and it'll do the search it'll do a Google search on on these sources. Okay. However, I I think that the the output isn't perfect. I think most of the time it'll follow my prompt, but occasionally it'll go to some random websites like seeking alpha like which you know that kind of stuff.

Do you have you found a way to pull in YouTube videos? YouTube videos I think there is a setting. It doesn't work for all of YouTube videos, but it works. It could work. I've been manually So YouTube now gives you the transcript for everything and you can So I just there's a button that says like copy this transcript. So then I manually upload the trip, but I'm like there's got to be a better way.

So, how does this change? The most important thing, the the insane thing is one of the latest Gemini models can actually watch