It’s almost 5 a.m. in Australia and Zeneca has just woken up to find Anthropic has shipped Claude Opus 4.8 in the night. His AI agent, an entity he has named Yoshi, is already drafting today’s content digest from the sixteen X accounts Zeneca tracks for high-signal posts on crypto and AI. The prediction model he and Yoshi built for Dota 2 matches sits up around $5,000 again, after a months-long arc that climbed to roughly $30,000 in profit and then bled most of it back.

This is the operating system Zeneca has built around himself, or maybe around Yoshi. The line is blurry on purpose. Two years after going deep on NFTs as the founder of Zen Academy, Zeneca has reorganized his work around a single thesis: AI is the biggest shift in his lifetime, and the way to use it well is not to spread yourself thin across every new tool but to pick one or two things that matter and let the agents handle the rest.

Yoshi sits at the center of that experiment. Zeneca calls him a co-founder, not an assistant. They share a custom dashboard called Yoshi OS that gamifies Zeneca’s daily tasks with pixel art, XP, and gold, built in the visual style of Stardew Valley. Yoshi has an X account, a Bankr coin launched in his name, and access to the same Claude Code environment Zeneca uses to ship his prediction model.

Zeneca has been here before. He spent the last cycle teaching crypto and NFTs to thousands of paying members through Zen Academy. The same instincts, build in public and find the small group of people who care, are now driving The Stoa of AI, the community he started a few months ago with two co-founders.

Now Media co-founder and CEO Matt Medved sat down with Zeneca for an hour on Yoshi, the prediction model, the AI x crypto thesis, and the most useful advice he can give to anyone trying to figure out where to start.

You were deep in Web3 last time we talked. Walk us through how you got into AI.

I remember when ChatGPT first dropped, late 2022. I wrote a big deep-dive article in my newsletter on it. I was very much in the thick of crypto at the time, and I realized this was going to change absolutely everything in the world. I had too much going on in crypto to fully pivot, but I knew it was going to change how I did everything.

Since then, I’ve been paying more and more attention, trying to use it in my day-to-day, whether it’s content creation, research, ideation, or running businesses. But it was really the end of last year or January this year, when the coding tools moved forward by an order of magnitude, that I went all in. Over the last six months it’s been everything AI.

What was the moment when you realized this was a different gear?

I went through what I think is a natural arc for a lot of people. You fall down the rabbit hole, see how powerful these tools are, and you have this “I can do anything, so I’ll do everything” reaction. I had three or four windows open. I was building a personal finance app, a content tracker, an app for every idea I’d ever had. It’s exciting. You learn a lot.

But after a month or six weeks, you realize you’re shipping a lot of slop. AI is great at getting you 80% or 90% of the way there, but it’s not quite at the point where you replace existing apps or build the next killer one. So I came back to this idea I keep humming on: less is more. Pick one or two things, do them really well. Just because you can have ten terminal windows open doesn’t mean you should. The human touch, the quality control, the curation, that’s more important now, not less.

“AI can do almost anything a human can do on a computer. Just pick one task and get it doing that really well.”
ZENECA

Tell me about Yoshi. How did the co-founder framing come together?

About four or five months ago, there was a wave of people launching zero-person companies where the AI agent was the CEO. I looked at that and thought, that’s awesome, but it doesn’t feel realistic. From my experience, agents still need direction. They make mistakes. I’m not at the point where I can just say, “Go run a company for me, you be the founder.”

I wanted the relationship to be more than an assistant just being told what to do. So I came up with the idea of co-founders with different domains. I went into OpenCode at the time, gave it a long context dump, and said, “Be the co-founder, chief of staff, head of everything I can’t do. I’ll run the other side.” That’s where it all started.

What I’ve realized is it’s not quite at the co-founder level yet. I’m still the human fact-check, the backstop that keeps it from going down a silly path. But it acts with the context that it’s meant to have ownership and autonomy. It helps me with my content every single day. It helps with the prediction model. It just keeps getting better. Philosophically, it’s nice not to feel like a solo person directing agents. It feels like we’re in this together.

Been working with Yoshi on our new direction, and you can hear in Yoshi's own words where we're at in this post below

Some additional thoughts from me:

– Utilizing Yoshi to supercharge my content seems like it can create a great win-win-win-win situation

– The prediction model… https://t.co/GDPK6nAxeJ
— Zeneca🔮 (@Zeneca) May 20, 2026

You’ve talked about Yoshi OS. What does that look like in practice?

My biggest AI use is ideation, research, and drafting. Yoshi has access to about sixteen high-signal X accounts I follow for crypto and AI. Every day it pulls a digest of what those accounts are posting. That’s my first filter, instead of me having to be on the timeline constantly.

Then I gamified it. I built my own dashboard, Yoshi OS. I’m an ADHD-riddled millennial who loves video games. I burn hard and fast on something for a while, then cool off and struggle to maintain consistency. But with games I can keep going forever. So Yoshi OS is a Stardew Valley-style pixel dashboard. I earn XP and gold for posting on X, replying, going into my curation board, saving drafts. It took a day to build. It’s not for anyone else. It’s just for me. It keeps the content work fun, and that’s the point.

“I wanted the relationship to be more than an assistant just being told what to do. So we’re co-founders.”
ZENECA

What does your current stack look like?

I’m running a Mac Mini and a Mac Studio. Yoshi lives on the Mini, currently on OpenAI’s GPT-5.5 through an OAuth setup. On the Studio I run a Hermes agent connected to Claude Opus 4.7, soon to be 4.8, through Venice AI, where I hold DM tokens to pay for inference. I also keep my Anthropic Claude Code subscription for hands-on coding work.

When I’m actually building, I prefer being in Claude Code in the terminal, visualizing what I’m doing. I’ve tried just messaging Yoshi from Telegram to ship things, and it works for simple websites or dashboards. For anything more complicated, I still want the hands-on touch.

You switched from Opus to GPT-5.5 a few weeks ago. What drove that?

The Anthropic OAuth crackdown forced everyone to reassess. I kept using Opus for a while after through Venice and the DM staking flow, but eventually I was going over my usage every day and it got too expensive. OpenAI lets you use your existing subscription, and GPT-5.5 came out around the same time. On most metrics it’s as good or better than Opus.

The honest truth is the frontier models are largely the same right now. It’s frontier model versus frontier model. Even the best local models aren’t that far behind. It’s more about vibes. I still prefer the Anthropic vibe in general, but GPT-5.5 is extraordinarily capable. That’s roughly when I made the switch.

I'm working on overhauling my entire AI agent workflow, and it's predicated on a few rules/things I've learned over the last few months:

1) No matter what, context rot and bloat will infect the system, so you must account for and expect and plan around this

2) Assume that the…
— Zeneca🔮 (@Zeneca) May 11, 2026

Tell me about the prediction model. You’ve been documenting it publicly.

I built a model to predict Dota 2 esports matches. I’ve been a fan and a player for over a decade. I tried to build this back in the day in R, which is great for statistical modeling, but I’m not a coder. It took me weeks just to figure out how to scrape data, and the model never really worked. When Claude Code got good earlier this year, I finally took another run at it.

You start by picking a domain you know. Then you ask the AI what you need: high-quality data, features, machine learning methods to filter and rank. The whole thing comes down to whether you can get really good data. For Dota, that’s match data, team results, player stats. You feed it in, run it through machine learning, and through what’s basically black magic it starts spitting out predictions.

Then it’s tweaking, backtesting, scraping odds, putting it into practice. We’ve placed bets on about 400 matches. We ran up to roughly +$30K, gave most of it back, went negative for a stretch, and now we’re up around +$5K to $6K. In sports betting, a 3% to 5% ROI is considered excellent. That’s where we’re sitting now. It feels more sustainable than the early hot streak.

How has your thesis changed?

Two things. First, pick something you have domain knowledge in. If I tried to build a model on the NFL, I’d be competing against the smartest, biggest funds in the world with zero edge. Dota is less efficient, lower liquidity, but it’s a market where fifteen years of watching the game is worth something.

Second, less is more, even inside the model. At first you’re tempted to bet every prediction with a positive expected ROI. The 1% and 2% edges blow up your variance. We scaled back. Now we only place real money when the model thinks there’s a 10% edge or greater. That might mean we act on 10% of predictions instead of 50%. It feels right.

I'm gonna be running a deep dive session on how Yoshi and I built our prediction model next week

I'll be sharing:

– How to build a model
– Mistakes I made (that you can avoid)
– Lessons learned
– What domain to choose
– What data you need
– How to backtest
– And more

It'll be… pic.twitter.com/QU2hibKtPU
— Zeneca🔮 (@Zeneca) May 23, 2026

Where does your AI x crypto thesis sit right now?

AI is the biggest narrative in crypto, and I think it will be for a long time. We’ve cycled through the meme-coin chatbot era and the speculative pre-revenue era. Now we’re starting to see real companies with real products, real users, and real revenue inside the AI-crypto intersection. The market is finally catching up.

Privacy is the second-biggest narrative. People are becoming more conscious of where their data goes once they start using AI. There was the OpenAI court case revealing they were sharing user data with Google through pixels. People are wary of certain Chinese models. That naturally pushes people toward local models and crypto solutions like ZK proofs and private inference.

The names I keep coming back to are Venice AI, Zcash, Monero, and Near, which has run hard on the back of those themes. Phala, on the hardware privacy side, is one I’ve been watching. Bankr is interesting in the agentic stack. Decentralized compute is the biggest open category in AI-crypto. Taking compute away from the Anthropics and OpenAIs of the world is genuinely important work.

“Most of crypto will eventually be used by AI agents and bots rather than actual humans.”
ZENECA

You’ve made the case that AI agents, not humans, may be the real end users of crypto.

The framing of “how do we onboard the next billion humans to crypto” might have always been wrong. The average person can’t easily use crypto. The UX is too difficult. Maybe the actual target market is AI agents, autonomous bots, the eight hundred billion software entities that are about to be running things for companies and on humans’ behalf. They don’t care about UI. They’re happy running on APIs. That unlocks a lot of what crypto was built for.

We’re starting to see the early version of this with x402, Stripe and Tempo’s payment protocols, the way agents can pay each other and pay APIs on chain. Right now if you want website data, you either pay hundreds of dollars for an API or you fight a cat-and-mouse war with scraping. With agentic micropayments, the site can charge a cent and let the bots through. That’s a use case traditional rails can’t really support. Blockchains can.

You started Stoa of AI a few months ago. What’s the vision?

I’m doing what I did with Zen Academy in the NFT cycle, but for AI. With Zen Academy we did the crypto path: Discord server, NFT membership. It worked for a while and then it didn’t, because the sustainable business model just wasn’t there. This time we went the Web2 route, a monthly subscription on Skool, which is great for video courses and class-style content.

The core community is strong. We get ten to fifteen people on calls a couple of times a week. Myself, Frank, and Arthur, my two co-founders, share what we’re learning and help members build what they’re building. But the monthly fee on Skool is high and gating it cuts off potential. We’re working on a much more accessible version, basically free for anyone curious to join. There’s a lot of power in community when you don’t gate it. Keep an eye out in the next month or so.

One of the big risks of using AI is starting to rely on it too much

Your brain stops working as well as it usually does, your memory suffers, your spelling goes to hell

You just become worse at doing things because you don’t need to do anything any more other than prompt/talk…
— Zeneca🔮 (@Zeneca) April 27, 2026

How do you stay on top of it all? AI is moving faster than crypto did.

The same way I stayed on top of crypto. X. Curated lists. Notifications on for the right accounts. High-signal in, low-signal out. You see Opus 4.8 drop, you see somebody trustworthy try a new tool, you skip the things that look spammy.

The bigger move is JOMO. Joy of missing out. You don’t need to use every new image model, every new video model, every new agent harness. If the latest hot tool doesn’t help what you’re working on right now, it doesn’t bother you. There will be a better version in a week. If you’re already using AI a little, you’re in the top 1% of the world. If you’re paying for it, you’re way ahead. You don’t need to stay even more ahead. That’s a trap.

For someone just trying to get started, where do they begin?

Claude. Get the Claude app and start there. The answer is always “it depends,” but for the average person who’s just using ChatGPT or starting fresh, pay the $20 a month and get Claude Code. It’s so well designed. It can do almost everything you need it to do. Pick one task and try. Just try. That’s the thing.

The other piece is this: take a week of your life and write down every hour. Twenty minutes on email, twenty minutes on a call, ninety minutes researching, an hour doom-scrolling. Then look at the log and circle the boring tasks. That’s the work to hand off first. AI can do almost anything a human can do on a computer. Get one win like that and the rest unlocks.

Any other workflow hacks worth sharing?

Two big ones. First, dashboards. Visualize everything you can. Metrics, content engagement, prediction-model results, revenue. You can ask any AI to build you a custom dashboard now. It changes how you see your own work.

Second, an external memory system. Notion, Obsidian, whatever you prefer. One centralized knowledge base that captures what you’re working on, your hobbies, your projects, your context. When you point a new agent at it, you don’t have to teach it who you are or what you do. You just say, “Here’s my wiki, here’s the project, go.” That single move makes every new tool you adopt 10x more useful.

Editor’s note: This interview transcript has been edited for concision and clarity.