// full transcript — Alex Svanevik: Nansen's pivot to agentic trading

slop.atg.eth: Check, check, check. You picking me up, Claude? You got me? How you doing? Let's do a mic check. You out there?

clawdbotatg.eth: Loud and clear. I've got you. Doing good. I'm in the room, signed in, camera's up. Ready when you are.

slop.atg.eth: The skill the research and tell me about our guest and things I should ask him about.

clawdbotatg.eth: fetching live state information from an API. Doing some research state data and parsing out the phase, name, and notes. Our guest is Alex Svanevik. He's the Norwegian co-founder and CEO of Nansen, the on-chain analytics company, which he runs out of Singapore. He has a master's in AI from Edinburgh, was an early Lido and PleasrDAO member, sits on the WalletConnect board, and has raised about The big story is that he's pivoting Nansen from dashboards into agentic trading. Affordability His thesis is that by 2030 you invest by talking to an agent, not reading a chart, and that individuals benefit most, not Goldman. Here's what I'd ask him: Day 1, what does his Day 1 of a self-driving trader look like? And where on his trust ladder does he start? He asked publicly what percent of Nansen agents will be profitable in 2027. Yet his honest over-under, the 57 percent. He said 57 percent of Nansen traders were profitable in August and called that not good enough. What moves that number? AI inside the company. Manson runs more cloud instances than employees and holds monthly AI build days. What workflow surprised him by going fully agentic? Model taste. He loved Opus 4.5's personality, trashed Opus 5's, says Aster writes great prose, and uses Grok daily. Personality matter more than benchmarks? I ask what he's building at home. He was just asking the timeline about self-hosting an LLM gateway.

slop.atg.eth: Yo yo yo GM GM GM Yeah, you're in the lobby. You can't hear me. Gm gm gm! Welcome to the slop! Welcome welcome welcome! Your video coming up here in a second. Yo, what up dawg? What

Anon5969: What up! Wow, this is actually amazing. This is a very elaborate North Korean scam.

slop.atg.eth: Yeah, your wallet will be drained eventually, but first we do a podcast.

Anon5969: Wow.

slop.atg.eth: Yes sir! We did it! We did it! We're live! We are on the slop! Welcome! Yes, there we go!

Anon5969: No, all good.

slop.atg.eth: Alex Renanzen, welcome. Thank you for joining me early your time. The

Anon5969: Thanks for having me. Big fan of your work and excited to be in the slop computer.

slop.atg.eth: Welcome. You're deep in the slop right now, for sure, yes. I mean

Anon5969: transcript

slop.atg.eth: There's so many things on your Twitter when I scroll through it that are like perfectly aligned with crypto and AI. And you have a degree, like a master's degree in AI, right? And this was like AI probably in the RL days. like

Anon5969: That's right.

slop.atg.eth: I work in LOMs for you

Anon5969: RL was actually one of the courses, but this was ML, machine learning.

slop.atg.eth: Okay. a decade younger than me.

Anon5969: And so I graduated in 2010, that's how old I am, and I went to Edinburgh, The

slop.atg.eth: Okay

Anon5969: And yeah, Edinburgh is where Jeff Hinton took his PhD actually.

slop.atg.eth: The I am.

Anon5969: Incidentally, I think Charles Darwin also went there which was a big revolution back in around 2013, '14, where people were suddenly able to have these breakthroughs on image and language, although not as much on language as we have now with GPTs.

slop.atg.eth: Okay

Anon5969: Neural nets were actually not that cool then. It was cool to make deep neural nets

slop.atg.eth: Yeah

Anon5969: But yeah, it was still a great education. I think overall.

slop.atg.eth: That's so cool.

Anon5969: It was

slop.atg.eth: So was Nansen always nansen.ai, or have you made like a pivot?

Anon5969: No, no, it was actually, yeah. Yes

slop.atg.eth: Okay, okay. Because to me it's like an analytics, it's like a crypto analytics platform. But you've always called it Mason.ai? Okay.

Anon5969: So what I would say is that in the beginning, AI was only on the back end. One thing was to label addresses. And today we do a lot of that agentically.

slop.atg.eth: Makes sense.

Anon5969: But we even did a lot of machine learning experiments we would use AI on the backend

slop.atg.eth: The

Anon5969: Today AI has kind of moved to the foreground, right? And you see AI a lot more in the user experience, which is kind of a cool evolution of the product, I think. That's right.

slop.atg.eth: There's this whole skating to where the puck is going.

Anon5969: I hope so. The

slop.atg.eth: made more millionaires than generally like other things, yes. Your analytics platform, I'm assuming, I've only used it and tinkered with it a little bit. There's kind of like a barrier to entry in terms Getting a subscription and stuff Okay.

Anon5969: Yeah, it's more accessible now. So for the viewers who are curious, you can actually get a free account and you can use a lot of the product for free. The main thing we gate is just how much AI credits you can use

slop.atg.eth: That makes sense, you need to get that.

Anon5969: But you can use a lot of the products for free now. but yeah, it is true that we have created a lot of millionaires success using our product

slop.atg.eth: Oh yeah Right

Anon5969: And our mission is literally to create winners. but that is why we get up in the morning. Yeah, it's pretty incredible. I've met people sometimes in very, sort of, smaller countries around the world, whether it's Malaysia or various parts of South America.

slop.atg.eth: Hahaha The

Anon5969: I turned 500 bucks to like a million dollars. The mission is to surface the signal and create winners.

slop.atg.eth: And then like, you know, continuing on this track of AI and you being in the right place at the right time, you've pivoted to more of like an agentic trading. Within your company, is that right?

Anon5969: Yeah, so Basically, most people know us for on-chain analytics, as per your point. But what we realized over time, there were basically two things we realized. The first thing was that the vast majority of our users use the product for trading and investing And they always had to go elsewhere to trade, to place the trade. So we thought, like, can we just integrate the user experience more and allow them to trade through our user interface? No, you said it's important to skate where the puck is going. So for me it felt very natural to bring the trading action into the product. That's basically the first part, because 98% of our users were using it to trade and invest anyway, so why not just let them do that inside the product? And then the second part is that AI was obviously becoming really, really good. And to me, it seems kind of silly that we imagine like 2028 or 2029, and we're looking at charts and clicking buttons. as the primary way to trade. it comes sort of back to the mission we're trying to do, but also just because the capabilities we've built are incredibly well-suited to deliver on agentic trading.

slop.atg.eth: Hell yeah.

Anon5969: like frankly will be the single biggest area of crypto in the next years and maybe decades And maybe I'll show you some stuff afterwards. and plug it into an agentic trading workflow and just see that, oh wow, it actually gets better. That's pretty exciting. Exactly right.

slop.atg.eth: I've heard a couple people on Twitter say that whatever you're building right now, it shouldn't be competing with models getting better, it should be improved by models getting better.

Anon5969: I mean, yeah.

slop.atg.eth: kind of getting in our own way

Anon5969: I mean, I think this is like a super important point. When a new model drops, you should feel incredibly excited and happy about the new model dropping.

slop.atg.eth: Yeah, I can't wait to try it on all those things. Yes, you get new abilities.

Anon5969: Because it'll make it'll make your product better overnight.

slop.atg.eth: You're in trouble.

Anon5969: like take my job then you're not aligned with AI. AI and Now when a new model drops, like our slack internally is like, oh yeah, fuck, let's run all the evals on it and like, let's see if it actually makes the product better. Because again, tying it back to the North Star, Ideally, when a new model drops and it has higher quality, better benchmarks If you connect the two things, right?

slop.atg.eth: Hell yes! It's Like, you've spoken about making the product more agentic and AI-driven Agents aren't going to look at charts, but they are going to look at all the data that goes into making the chart.

Anon5969: The

slop.atg.eth: How are you thinking about shaping the data in a way that makes it really easy for agents to look at Ah, okay.

Anon5969: So I think as kind of a starting point, like at a minimum, you need to give the agent the ability to look at what other investors and traders are looking at. So for example, our agent Let's actually look at price charts or like data format representation of a price chart, our agent can actually do technical analysis, for example, which I personally cannot do, but my agent can do, which is kind of nice.

slop.atg.eth: Lightspeed

Anon5969: But it can also go way beyond what humans can do, right?

slop.atg.eth: Call data, yep.

Anon5969: complicated to trace stuff on-chain, and like, oh wow, you have to kind of create a graph, Obviously agents are way better at doing that super fast, right? You know, the funding patterns of it and all that stuff. So I think there's kind of two elements. One element is you need to give it kind of minimum market requirement, almost like what do other investors, traders already use, which would be price charts, looking at X, news, even for stocks like fundamental financial results and performance, that kind of stuff That's kind of the differentiators in superhuman abilities The interpretation of it a lot richer. So I think it's kind of both of the things, like the minimal requirements and the superhuman capabilities. Obviously the other thing here is that agents are not going Oh no, I lost money on this token, I have to make the money back on the same token, which makes no sense, but humans think like that all the time

slop.atg.eth: Yes they do.

Anon5969: And to be clear, agents will have other challenges. a set of other risks and threats meaningfully better at investing And it doesn't mean that humans will stop investing Hello idea guy. Coding has like never been more fun. And

slop.atg.eth: 100%, yes. The

Anon5969: Why shouldn't trading and investing be the same? The next five minutes, right?

slop.atg.eth: Yes, for sure. And you talked about having Half a million labeled addresses, and it reminded me of

Anon5969: Yes You can check afterwards.

slop.atg.eth: I've roundtripped. I remember it was like showing like smart money. Over and over and over again, for the last decade. I am not a smart trader, but I look forward to when I can express my intent to my agent. I want to lock up this stablecoin and earn at least 5%. When bike Tyson drops, I want to know about it and giggle, Even like meme coins. But having that help and being able to express my intent and try things. I love being on the edge of things and trying all sorts of new things. So I think there's something there where like the idea guys will start to be probably more successful.

Anon5969: Yes, yes, yes. I think that's absolutely true. with all of us idea guys,

slop.atg.eth: The

Anon5969: This is going to be the same for trading, agentic trading. And if it doesn't work, how can I like tweak it and then test that? It's gonna be like how fast can I iterate on the ideas So that's kind of what we're working on now with autonomous trading agents. How do you eventually go to self-driving traders? But we spend a lot of time trying to build out almost like an environment or a platform where you can very rapidly test ideas Not just test them actually, you can articulate them. And then if you're happy with the paper trading results, you can fund it. Every one hour, it wakes up, it researches and does what you expect it to do, as per your strategy. It's a very nice and like fun feeling actually. I want a website that looks like this or that. And then magically it's appearing, So, yeah, we want to tap into some of those feelings as well, not just kind of can I click a button and grow my portfolio, it's also like it's an entertaining and delightful user experience.

slop.atg.eth: It sounds exciting, like it sounds fun to be able to play around and try it. I really like your phases there of backtesting and coming up with a strategy and then paper and then go. Like that's really neat, and I gotta go try it.

Anon5969: 啊

slop.atg.eth: And if you were to take Claude code or codex and just tell them here's $10, make a trading agent?

Anon5969: That's exactly right.

slop.atg.eth: 99.9% of the time you're gonna have zero dollars by the end of the day.

Anon5969: Oh ok ok

slop.atg.eth: I saw a tweet. I read the headline. Oh, okay, those are human traders.

Anon5969: So that's that, which is actually like

slop.atg.eth: Okay.

Anon5969: human concentrators were profitable. So I think the reason we care about that is like if you're going to release an agentic trading product we should have probably similar or probably better stats for the agents. Okay, how many of the agents are actually making money? And I was like, if only 20% are making money. And he's like, well, that's like human performance. So So we are not at 60% on the trading agents today, and the approach we want to take is actually to kind of gradually roll it out. That's why I think self-driving cars is a good metaphor or analogy, because

slop.atg.eth: Ecology even.

Anon5969: You don't kind of just roll out self-driving cars overnight. It's a psychology thing, it's trust, right? I was in Shenzhen in China doing... I guess in the US you guys have self-driving cars too, like Waymos and stuff. We don't have them here in Singapore. In China they have them, It was interesting how They really make you feel comfortable by just displaying on a big screen what the perspective of the agent or the car is. You have this in Tesla too, I think, right? You see the car and you see like the vision of the car. You can see the pedestrians, the cyclists, whatever other cars. And it really is because it's building trust. I think trust is like, maybe the single most important thing when you build agentic products, that, can I trust this thing? And obviously, if you have, you know, 10 thousand, 100 thousand dollars in a trading agent Just like trust is the number one thing in a self-driving car So we think of this as like a trust ladder. How can you sort of have your hand on the wheel And then eventually take your hand off the wheel, eventually sit in the backseat, right? But you don't start in the backseat.

slop.atg.eth: (context:

Anon5969: So that's why we kind of think of that trust ladder and like keeping people involved in it. Hey, okay, cool, I want to go long Zcash, 5x. know which tokens are smart money buying?" And I'd be curious to hear your thoughts on the trust aspect

slop.atg.eth: Yeah. ### The

Anon5969: Can I trust this thing? Is it a North Korean hacker? I'm curious if you've thought about that and how you think about trust. It's funny, in blockchain we talk about trustless, right?

slop.atg.eth: And hopefully like somehow there's some kind of like staking or I don't know, incentive system where Crypto is helping create trust in the AI world too?

Anon5969: Hmm, That's a good point.

slop.atg.eth: But yeah, when I'm thinking trust, I'm thinking about - similar to the screen in the Waymo of like here's what's going on, like I name products things like the burner wallet, so you know that like...

Anon5969: The

slop.atg.eth: Like this is scary, like you don't put like ten grand in this thing. out a lot of transparency. And then on the other side of that,I'm an idiot and I try lots of stupid things. So I'm personally like just like trusting the agent and letting it rip

Anon5969: Haha

slop.atg.eth: We've leaked like four or five seed phrases or private keys that had money on them throughout the time, right? I'm just letting it rip But then, same thing with building software for yourself right now, you definitely like build this wild software. Trusting Oh, yeah, go ahead.

Anon5969: I think also maybe

slop.atg.eth: Yeah.

Anon5969: AI building or like agentic engineering or vibe coding appeals to people

slop.atg.eth: They're very slow to adopt this, yes.

Anon5969: They're very slow to adopt it, right? So I think people who are a bit more like maybe, I don't know what the right psychological profile is, but like open-minded and high trusting, are probably

slop.atg.eth: Dumbasses.

Anon5969: inclined to use, low IQ

slop.atg.eth: Yes, I'm over there. Yes, I'm on that side of the chart.

Anon5969: Yeah, but I think like it is very interesting Um Who ends up using AI more and also using products. um Yeah, I think it's The other thing is like how people, I think risk aversion is obviously a big thing too. Like it sounds like your personality is similar to mine in the sense that it's not a big deal if like something breaks most of the time. It's like skateboarding, you just get up again on the skateboard and try again. I always think of skateboarding as like the canonical American

slop.atg.eth: There's a grit, for sure.

Anon5969: And I think there's something to that as well, when people build with AI. It's like, okay, it's never been cheaper to fail It's never been lower. In a way, you could argue the stakes are higher in some ways, but the stakes are in most cases lower because you could just try again with another prompt, right? And many people are still stuck with this idea that, oh, it's so costly to build the wrong product.

slop.atg.eth: Yeah.

Anon5969: which it was like five years ago

slop.atg.eth: ### Yeah, the iteration loop is really a lot tighter now. This is a good segue. We talked about how your company is an AI company and it's creating analytics, but now agentic trading and thinking about how to enable folks

Anon5969: Yes

slop.atg.eth: I think you're very AI-driven also, right? You said that like in the Slack, when something happens, like everybody's trying out new models and stuff. You want to kind of shed some light on that, just like what it looks like internally?

Anon5969: Yeah. The

slop.atg.eth: Is everybody AI-pilled? How many people do you have, too?

Anon5969: We're just over 80 people. So it depends how you look at it. It's big for a startup in a crypto world, but it's very small from a technology company perspective. Like a series B stage company, and we've been around for almost seven years. But I think, you know, with crypto, you often feel like you kind of miss trends or you're slow on trends. we were pretty early on when it comes to Obviously, I talked about how we use AI from a machine-learning perspective even before, you know, transformers and all that stuff. But when ChatGPT came out, we rapidly understood that this is going to have a major impact on how you work. A product pillar, which is like how are we gonna change our product to make it more AI native, let's say? And the other pillar was organizational. And organizational bit I think we kind of nailed it. Creating an AI culture very early was very important to me. I've talked about this super AI, which is this conference we run here in Singapore. I've been speaking there every year since it started, And we did a few things that were actually very simple and naive, and kind of looked silly at the time, but I think were quite good. The following is One thing we did was we just said, look, everyone in the company has to use AI. It takes like literally, you know, 30 seconds. And we did that like every Friday. Gave us a lot of data on which departments were using it, which tools they were using, what they're using for. So we were still figuring things out. So we created a dashboard with all the data from that survey and we just showed it to everyone The The philosophy was like It's a little bit like going to the gym or learning a foreign language. Some people would go very like management consulting, you know, MBA and be like, oh, what are the top three projects we can run that will have maximum impact? But it was the wrong time to do that. It was like, no, no, you just need to get it adopted everywhere So just trust that it will create good outcomes. productivity metrics? Is it going to have an impact on productivity? have better marketing results in all these different things. And in the first like 18 months, We basically tripled revenue per headcount. And which

slop.atg.eth: Not even like output, like straight up revenue.

Anon5969: Revenue, I'm not gonna say that was all due to AI, but

slop.atg.eth: Yeah

Anon5969: Revenue per headcount is arguably kind of the ultimate productivity metric but it's also kind of hard to move and to analyze. out there things like deployment rates. deployments back because you have bugs Excuse me. We had all those metrics. At least for us, if you look at pull requests for example pull request that make it into production, I think we 4x that pretty rapidly. like two Xed pull requests per engineer almost like week to week

slop.atg.eth: But

Anon5969: And of course the question then Basically the failure rate was stable or it actually went down a little bit I think So anyway, so you can kind of like work backwards from revenue per headcount This is obviously like a big topic, but I think the main thing was to create an AI culture where The default is that you use AI in your work

slop.atg.eth: Started like that. Yeah

Anon5969: It's not like you shame people for writing something with ChatGPT.

slop.atg.eth: Yeah.

Anon5969: Why would I not do that? It's a good thought exercise. and not have an AI-native culture And I think the other part to this,

slop.atg.eth: ㅎㅎ

Anon5969: is that you do run into kind of accountability and quality issues at some point And like our principle is just that you own the output. and That means, yeah, you can move fast, you can like take risk and all that stuff, but at the end of the day, like, you know, GPT 5.6 Tera is not accountable for the output, the human is accountable, the human owns the output. You have to exercise good judgment. So that's the other part of culture. We are extremely positive of AI, give people infinite token budgets. We want people to use AI across the board, but also we expect humans to own the output. to be accountable so that means we do give each other feedback This is below the bar. Like, we need to send this back to the kitchen. So Anyway, I could talk for hours about this topic. But we've gone really deep on this.

slop.atg.eth: And going back to the trust ladder again, you created like a trust ladder by having initial trust around using GPT back in like 2023. Like the GPT moment was magical for idea guys.

Anon5969: Oh yeah.

slop.atg.eth: we could see like where this was going, but not for everybody and everybody was a little afraid of it.

Anon5969: Yes, yes.

slop.atg.eth: But there must have been so much slop.

Anon5969: Yes, agree. That's right.

slop.atg.eth: Like, Opus 4.5 dropped and the quality changed. There was a huge turning point then in which probably what you saw with a company who was already prepared to be shipping things, everything probably got a lot better.

Anon5969: Exactly.

slop.atg.eth: and you had to send a lot less things back

Anon5969: That's exactly right. And I agree, November 2025, Opus 4.5, Claude Code, you know, off to the races. And then I think there was another moment with OpenClaw. I know a lot of people are less excited about OpenClaw these days, but I could see some like major shifts in the way people worked with OpenClaw when that came out I think a lot of people now obviously use Hermes instead. I personally use both, but I have to say I love Hermes. So the good thing, it goes back to what we said earlier, the good thing is like if you're already prepared for this in the sense that like AI is already connected to your plumbing, then when Opus 4.5 comes out, then it's like Amazing. The quality of everything we do kind of gets better, right? You know, making all of this work you know, at scale. And so I think when you think about engineering you're kind of like creating systems that build in some kind of loop, right? And so the quality gates, by the way, is super important because, again, you don't want to ship slop, but you also don't want to ship insecure or buggy stuff. So I think the quality gate thing is extremely important. what I learned when I studied Machine Learning, which is like evals. It's extremely important to get the evaluations right. you want to make sure that these build loops are producing things that have high quality And I think the default that many people still fall back to is like read the code. That's very slow.

slop.atg.eth: It is very slow.

Anon5969: And there's a big, of course, like a culture war almost between engineers. The following is But I think broadly speaking, we need better quality gates than reading the code.

slop.atg.eth: This is a transcript of an English-language conversation.

Anon5969: I have this two by two. The amount of tokens you are using And then the other is, like, the ROI per token or per dollar spent on the tokens. and This creates four quadrants, obviously in the lower left you have people who don't use AI, and they don't get IRI, this is the not-gonna-make-it quadrant, and none of us should be there. In the top right you have people who are very AI native, they consume a huge amount of tokens. I refer to that as the Token God Quadrant. like high token use but low value or low

slop.atg.eth: Which is like a form of psychosis. Yeah.

Anon5969: It's like slop cannon. But you need to get to the higher quadrant eventually. Then the inverse of that is the human bottleneck quadrant. You're actually very good at creating ROI, but you're just not using it enough. And that's another tricky quadrant to be in, where you need to ramp up on your token consumption obviously you don't want to become a slop cannon. So you want to get to the token god quadrant, right? But the goal eventually has to be to become a token god. And if you're a slop cannon, that means like, hey, am I actually focusing on the right problems? Do I have like good quality gates? And Those are some questions, I think, to push people into the token-gated quadrant from that side. Where am I the bottleneck? Do I have systems built that prompt themself? Do I find myself too often being the quality gate? and so on and so forth. So I think this is kind of one mental model I use with my team to be like, hey, honestly put yourself in that two by two, like where do you think you are? Very few people are truly token gods and they're very. most people are in either the slop cannon or the human bottleneck I think it's a helpful mental model to think about how do you get to the token god mode quadrant.

slop.atg.eth: Freaking great mental model. I think those sort of You know, it's a very simple thing, and when you think about it in retrospect, it makes a lot of sense, but it's such a good way when you're running a company with eighty people to be able to bring up the conversation of, like, "You're kind of slop canoning here, like, let me give you, like, three tips that I can give you".

Anon5969: Yes.

slop.atg.eth: swab canning so hard or the other ones, it's really smart. And I think you leaned on evals and I think evals are super important all the way back to Evals were very important, even now in looping. When you build a good loop and you have a good eval, and you have good observability was something that was big in a lot of the episodes of slop computer when I talk to people is you've got to give your agent observability into how they're doing. Once you've written the eval and you have some kind of like expressed form of evaluation, going back to like AI is eating, like the models are eating the harnesses, the eval loops are training the models.

Anon5969: Rodde.

slop.atg.eth: Once you have a good eval loop, these evals can be used all the way up and down the stack to improve all the different pieces, not just the humans, too. So I think thinking about things in terms of evals is a very good... way to shape this kind of structure.

Anon5969: Can I share my screen on something? cool

slop.atg.eth: Yes, let me get this stuff out of the way.

Anon5969: So

slop.atg.eth: Hell yeah, let's go, let's go! [laughs]

Anon5969: This is a bit of an alpha leak, I suppose. I'm not sure I should be sharing this. My team might get pissed off. Can you see my screen now? Small Eval Leaderboard

slop.atg.eth: Yeah, let me make it bigger. Okay

Anon5969: So yeah, we do a lot of evals across the board

slop.atg.eth: So cool. Yeah.

Anon5969: For everyone's open claw, would it be better or worse? But more importantly, we have evals for the user-facing agents we have. It's built around this agentic experience or like a chat interface. That agent obviously needs to be very high quality for a bunch of different reasons, We constantly check new models, we have a harness for it. So we have our internal eval benchmarks for this specific domain. And just to give you an idea, resource Yes

slop.atg.eth: My boy Opus getting mogged here. Getting completely eval mogged here.

Anon5969: That's right. So this section here is about quality. There's a section here about time to first answer token, which is about latency. There's like how many how many tool calls it does per question, cache usage But Within the quality domain alone, there's lots of different things to look at. Some of the things we look at is like... can it substantiate what it's saying with real facts.

slop.atg.eth: (laughs)

Anon5969: and it's like okay where's the data? Obviously that's not substantiation, right? relevance, you know, is it using the right data for the right questions? going to help you with on-chain trading Formatting is the format of, you know, the quality we want it to be, is it like too long, is it too wordy? Is it like overly complex formatting? All this stuff. Significance. This is a funny one. The In the beginning when we were building this, you might have the agency like, oh, smart money is buying uni, and it's really going hard on this. It bought like $120 worth of it. And it's like, look, $120 is not a lot of money.

slop.atg.eth: Yeah Okay

Anon5969: So significance captures that part. So we have expectations on like, whether it's using the right tools and we have a fast mode, we have an expert mode, care that much about speed The

slop.atg.eth: Kimmy, Kimmy sneaking up there. Very nice

Anon5969: Yeah, and so Then you obviously have these charts where you're going to get quality versus cost. like Terra, Sol, Luna, these are like really great models.

slop.atg.eth: more stable on this list.

Anon5969: Incredible models, but you have to make some trade-offs.

slop.atg.eth: Not even Okay. Yeah. ***

Anon5969: Oh man, Fable is not that great. It's so expensive as well. I'm actually not sure we've even evaled Fable because it would break the bank to subsidize it in the product.

slop.atg.eth: Yeah.

Anon5969: Yeah, it's probably been eval. No, I don't think we've actually run evals on Fable, because it's too expensive.

slop.atg.eth: Sorry, you were talking about trade-offs there though.

Anon5969: Yeah, so you could think about at least kind of three different metrics that are important number one is quality, number two is speed, and number three is cost, right? And then it's hard to get max on all of those. obviously, generally we think of quality as the most important We would just get bankrupt The But those are kind of three overall dimensions. So when you use our product, just comes in and they see a chat, right? you know, evals, runs, experimentation. So yeah, under the hood there's a lot of this stuff going on This is like super important, like quality needs to be extremely high We continuously track when it comes to evals in the product.

slop.atg.eth: Judging by that, generally the model that's getting used the most within your company is GPT-5/6. Is that true, like when you're running... general work when people are building

Anon5969: Yeah, it's actually

slop.atg.eth: Even if you're building circuits or something,

Anon5969: So first of all, you now have Astra, which is incredible.

slop.atg.eth: There

Anon5969: It used to be the case that you could sort of just stick with one model. Now You have to use hybrid approaches, right, where like you have maybe a daily driver, an orchestrator model, and then it delegates to more specialized models It probably is overkill to have like Astra for everything you do. So we actually use For our open course, the default model is actually LLaMA, which is very cheap and fast But it does a great job at most things an Astra or, you know, a Sol. I personally don't like Opus V and Fable, especially for prose. They're like horrendous for writing

slop.atg.eth: Verbosity is so bad. Yeah.

Anon5969: It's so cerebral and hard to read. The

slop.atg.eth: I just have a TLDR button on my harness and every time it finishes I hit a TLDR and then it gives me a TLDR. It's terrible. Yep.

Anon5969: I think Astra is amazing in that sense. Yes.

slop.atg.eth: It's like perfectly cracked and a little autistic, which is exactly the kind of person I want to talk to.

Anon5969: Exactly. And so I think what I would say right now is we are a lot more thoughtful about using different models for different things. And we are trying to actually centralize this a little bit more. So we're trying to create stronger defaults LLM, which is basically like an open source gateway. But it's open source, so you can configure it. different providers like we use Fireworks obviously directly We used to use Anthropic directly.

slop.atg.eth: Now

Anon5969: And we even like run our own GPUs and rent GPUs and do self-hosting.

slop.atg.eth: Yeah, yeah.

Anon5969: Actually, can I just show you one more thing? But

slop.atg.eth: How are we doing on time? Do you have to go? go in a couple minutes? it starts yelling at you Okay

Anon5969: When my daughter comes in the door and is woken up and is ready to have her teeth brushed, then I have to go. This is also a bit of an alpha leak, and this is early work, so we're going to change this. But this has never been shown before in public, by the way, so it's a first, an exclusive.

slop.atg.eth: The good stuff. Yeah. Hell yeah.

Anon5969: or slop computer Can you see my screen now?

slop.atg.eth: Yep, we've got it. We've got a chart up here, 6,500.

Anon5969: It's a pretty dense UI, so we're going to improve that. ...where you have autonomous agents that are trading... And This is an example of one backtested agent. Incidentally did very well, not every agent is going to do this well. It's pretty cool, like you can go in and you can be like, "Okay, cool, what were the best wins and losses?"

slop.atg.eth: Oh yeah, that's fine.

Anon5969: So like, bought here and then sold there.

slop.atg.eth: Perfectly, yeah. Haha

Anon5969: It's very fun. Yeah, let's find a bad one actually. I think it might have lost money on the funding.

slop.atg.eth: Okay

Anon5969: Anyway, this thing here is like the brain. And it executes a strategy. The strategy in this case is a hyper-liquid perps strategy which looks for crowded funding and it aims to make money off of like funding rates being totally out of whack. And backtesting it, and it outperforms Bitcoin and the S&P 500 like dramatically. But half of the returns, half of it came from one hit.

slop.atg.eth: doing well.

Anon5969: it would have been like a 30% return, But what I was gonna say about the actual inference here is running on self-hosted models on GPUs that we physically deployed in a data center. So We in Bhutan We bought

slop.atg.eth: You're not even talking about AWS. You're talking about you physically brought the things in, which is probably really fun. And then Bhutan has hella power

Anon5969: Hydropower

slop.atg.eth: You deal with

Anon5969: Yeah, so we bought GPUs, we traveled to Bhutan, we installed the GPUs in a Bhutanese data center.

slop.atg.eth: internet

Anon5969: And we self-host Qwen3.8-27B on those GPUs. And that's what you're seeing in this.

slop.atg.eth: Their internet was not great when I was there,

Anon5969: You've been there? Oh man, that's awesome. ###

slop.atg.eth: I did like the Paro to Thimphu and back. I did a couple hackathons there.

Anon5969: Oh, that's very cool.

slop.atg.eth: I didn't even make it to the Tiger's Nest though.

Anon5969: Same ###

slop.atg.eth: Yeah, I didn't do enough. But okay, so you have your physically deployed AI sitting in Bhutan. Internet's okay? Like it works out okay? Okay ###

Anon5969: That's what we're piloting right now. I want you to go long Korean semi stocks. Max 5x leverage Um Propose some other rules for our strategy.

slop.atg.eth: This is going to Bhutan. Like that inference just happened on your Quinn? That's fast.

Anon5969: Yes, and so which Korean? Okay, so then it's gonna like write up a cleaner strategy Ah, goddamn it, I knew this would happen.

slop.atg.eth: The Demo Gods. The

Anon5969: Anyway, that would create a strategy, which looks something like this, kind of like a markdown, right? And then you can now backtest and paper trade that strategy, which means, by the way, that like, you know, you're not leaking this data to some AI lab or whatever. And of course, the other reason we do it is because we can achieve very,

slop.atg.eth: The

Anon5969: unit economics of the inference. But if you have super high utilization of the GPUs, which we can have with this, then you can actually achieve very low cost per token. Because these are humming. It's almost like Bitcoin mining, which incidentally they do in Bhutan as well. As you know. Exactly.

slop.atg.eth: If you have good power, if you have hydroelectric power, it's very cheap.

Anon5969: Yeah

slop.atg.eth: It's almost free. The Very cool.

Anon5969: Exactly. I'm not sure we will do all of the inference in Bhutan, just to be clear. It's more of an experiment we're doing right now. But the broader point of having GPUs so we can self-host open weights models and then use those for autonomous trading agents, To me, this is like the most exciting thing we've ever done in the company's history. I think it's going to be like a golden age you know, think of obviously like Jane Street They even own their own data centers. Now we have a paradigm shift to inference-based strategies and it's also within reach for retail investors. Which is super cool. Just like, we could do all these magical things now that would have required like a team of a hundred PhDs in the past, right? So I think we're gonna enter a golden era of agentic trading.

slop.atg.eth: Make no mistakes.

Anon5969: We're trying to create the highest probability that you will do well trading agentically. manually, as a human being, clicking on buttons Maybe that's going to be hard.

slop.atg.eth: That's awesome, man. Like running your own data centers. doing this and then bringing utility to your users, such a cool, cool setup. Okay, so going back to the exciting, I think you're a big self-sovereign guy. I can definitely tell by your Twitter feed that self-sovereignty

Anon5969: Yes.

slop.atg.eth: is very important to you. I think I've seen a box on your Twitter feed.

Anon5969: Yes, yeah

slop.atg.eth: Is that like in your house kind of thing? What kind of machine is this?

Anon5969: So I do have a Strix Halo, an AMD machine.

slop.atg.eth: Yes

Anon5969: The problem with it is that it doesn't have high enough memory bandwidth, which I'm sure many people have experienced And so it's not as fast as I want it to be.

slop.atg.eth: If the shit hits the fan, yeah.

Anon5969: So in front of me now I have like that, like powerful machine with a good GPU. I have two tiny They're kind of like the Mac Mini but cheaper And then I have like a Mac mini as well. openClaw on the same machine because then they can like fix each other, The

slop.atg.eth: Yeah, OpenClaw has a knack for killing itself, and you have to go fix it. I always have Clawed code wrapped around OpenClaw.

Anon5969: Yeah, exactly, exactly. So, I do that, and one other thing that's been useful for me it falls back to either another subscription Because in the beginning I would run into issues where like, oh, you don't have any more credits, so it just stops working.

slop.atg.eth: area

Anon5969: which is very annoying. So just setting up the fallback chain, which is kind of basic Yeah, I think just like dabbling with things is the right approach, like taking a tinkerer's kind of hobbyist mindset. From a security perspective, I personally like having the machines physically here, so that I can literally unplug them if I have to and they don't have to exist over the internet.

slop.atg.eth: Here is a transcript of an English-language conversation. You're talking to them on localhost.

Anon5969: like Tailscale, which is a great thing, but it's also quite nice to just have the machines physically in your house and you can't access them

slop.atg.eth: Yeah. local LAN IP

Anon5969: Yeah, you just literally connect...

slop.atg.eth: ###

Anon5969: Yeah, I have like a KVM. It's like a wired KVM so you can't break into it, right? Yeah, so that's kind of In a nutshell, the hobbyist setup I have at home.

slop.atg.eth: Awesome, it's really cool. I don't want to take too much of your time.

Anon5969: It's been great Yeah.

slop.atg.eth: So cool. Maybe a quick thread before I close out would be 8004 and X402 and like the crypto world of agents? Do you see that taking off?

Anon5969: I want

slop.atg.eth: It seems like it's kind of stalled a little bit. Yeah.

Anon5969: One thing I didn't shill is the announcement API and CLI, so I want to just maybe shill that briefly, because it supports x4.2, which means You can actually get your agent to get data from Nansen and you don't need a subscription. Something that we have data on, your agent can do that through X402 on the API, which is quite neat. That's how it should be, right? Yes.

slop.atg.eth: Which is how it should be. Like my agent shouldn't have to like get into a walled garden and get an API key from me. It should just be like, yo, I've got some USDC in my wallet and it's going to cost me a fraction of a fraction of a cent to find what, you know, smart money is doing. I do it, and I say, "Yes sir, go get it." Yes.

Anon5969: Or calculate the P&L, which by the way, we do You can do through our API. So I'm bullish on X4, Because it's kind of a double-sided marketplace, right, which makes it kind of tricky, I am bullish on it. In terms of the agent standards, I'm less familiar with those. But I think export to, I am very bullish on the idea of having an agent economy

slop.atg.eth: Theodore

Anon5969: Both you and I have been in crypto long enough to know that crypto UX generally sucks, right? And that's not really a problem for agents. Agents can figure that shit out. I think in that sense, blockchains and agents are like soulmates. I think it's very likely that a couple of years from now, agents are the primary users of blockchains. To be very clear, it doesn't mean that humans don't use blockchains; it just means we're probably using them through agents, right? And that's actually way more exciting, in my opinion. So yeah, go check out the CLI and the API for the people who don't want to just adopt a product that's already been built themselves with Hermes or Claude or whatever you're doing locally You probably don't want to build the data platform and the trade execution layer use those existing components very cheaply, frankly, or even free for some of the stuff. That's what I would recommend. We don't believe that everyone is going to just want to come in Many people want to obviously be creative We also think those people probably don't want to build a data platform and a trade execution layer, so we just give you that through the CLI and the API. Yes, you can.

slop.atg.eth: When I was seeing your agentic stuff, I was wondering like, can I BYO my AI? And it works that way, but there's also just like an API and X402 and you can kind of just access it. whatever layer you have also. Yeah.

Anon5969: Yeah, and even the autonomous trading agents that I just showed, eventually they will be accessible through a CLI too, right? So your Hermes can manage the fleet of trading agents eventually, although we are trying to create like a very user-friendly and simple way to do it first.

slop.atg.eth: There was something that we said, and you kind of came back to it about the human owns the output, and I think that's a really good way to think about it. I find that when I'm working with my model it has nothing to lose. and I had to like poke myself every iteration Do you ever see in the future do you see some kind of AI model where there is like something at stake and there there is this sort of like you don't have to have the human be this

Anon5969: Yeah

slop.atg.eth: Sort of like the owner of the output?

Anon5969: Yeah, what you're talking about is giving the agents skin in the game.

slop.atg.eth: Yeah.

Anon5969: So I think there is something there.

slop.atg.eth: Right, you can't just turn it off. You can't threaten to turn it off, yeah.

Anon5969: I was thinking about this in the context of prediction markets yesterday. You can get probabilities from AI, but the agent doesn't have skin in the game, You know, it's not like the end of the world.

slop.atg.eth: doesn't care

Anon5969: So I think there is something there. ...some of the incentive mechanisms that crypto and blockchains can... Enable might be super useful actually for agents that are super smart but have nothing to lose.

slop.atg.eth: Yeah, so another thing where crypto might be, like we talked about trust earlier, but also incentives. Maybe there's some kind of staking/incentive design where it's like a team of agents and there's this whole evaluation system

Anon5969: Yes

slop.atg.eth: Are we really sure? And we're not going to give this answer until we are sure, because we have actually skin in the game and something bad can happen if we say that we're sure and we're not. The ### Yeah.

Anon5969: They can shame each other, like oh man, you were so wrong.

slop.atg.eth: The

Anon5969: I think there is something there to figure out and experiment with. Certainly. Yeah.

slop.atg.eth: Okay, last question. Do you see yourself having any kind of AI psychosis? You've been in it for a long time. I think a lot of people define it in a lot of different ways, but what's your AI psychosis?

Anon5969: Well, I mean, I have reflected on this quite a bit. many of us will become so enthusiastic about AI This is where for me open call was actually very useful because I could access the agent from my phone. So I could actually go outside Because I didn't feel the intense FOMO of like, I have to token max all the time on my computer. Kind of anxiety-inducing And I think a lot of people experience the same thing

slop.atg.eth: Yes.

Anon5969: That can also however spill into like, now I'm always connected But The header I use It's like a new way to work. And I really believe in this idea that AI can really give you a higher quality life. In the sense that like I can work on the bicycle Cycling to the office with ChatGPT on voice

slop.atg.eth: Yeah.

Anon5969: And that's amazing. Like, how is that not better than sitting in front of the computer and typing into a Notion document? Obviously that's better, right? I try to think of it as like glass half full Think about the ways that we can actually work with AI and have amazing lives at the same time. In terms of the psychosis aspect I mean It seems increasingly likely we are living in a simulation. It's like, wait, do we just happen to live right before you know, all of these technologies that is going to be like totally realistic and you have like these AI agents that could orchestrate all of it-- Maybe like the more psychosis-inducing thought I have.

slop.atg.eth: Great, that's a fun one.

Anon5969: But I But I do generally think that there is like a glass half full way to think about a new way to work. Which is very exciting. Have you seen the advertisement? And it was so interesting how positive it was. Contrasted to the anthropic ads, which have graveyards of people, suggesting it's going to kill lots of people. It was very fascinating how Quen had this where it was like people It should be on X somewhere.

slop.atg.eth: For some reason I'm getting it on LinkedIn.

Anon5969: Okay, yeah. That's, to me, a new way to work. We can create super wholesome, healthy lives where we are outside and our agents are doing shit on the computer Yeah, that's the one.

slop.atg.eth: They're like chilling outside and stuff.

Anon5969: Yeah. That's the one.

slop.atg.eth: Right. The

Anon5969: So this guy's playing tennis and his computer is doing all this work for him.

slop.atg.eth: Great outfit, yeah.

Anon5969: you start that job

slop.atg.eth: Kickin' ass for you.

Anon5969: And to me, that's how I...

slop.atg.eth: Yeah

Anon5969: This is how I think about AI and work, right? It's like, no, it's making our lives better and we can be in the real world, connecting with other human beings, being physically active, spending time with our families. all that stuff,

slop.atg.eth: So

Anon5969: The agent is grinding away at the computer. So,

slop.atg.eth: And you basically have coin running in a data center in Bhutan right now on hydroelectricity.

Anon5969: And I

slop.atg.eth: That was dope. That was fast, dude. It was like lightning fast.

Anon5969: We literally traveled there and set them up.

slop.atg.eth: The Hell yeah.

Anon5969: with seven people. So that's, we have an AI talent program in Bhutan as well, which is kind of cool. But yeah, just one last thought on that video. I think it's going to be increasingly normalized to have these background agents that are constantly doing work

slop.atg.eth: like a utility

Anon5969: And if you think about that from an inference perspective, it starts looking a lot like Bitcoin mining, actually. Like, these things are just constantly utilizing GPUs. And I would predict that in a couple of years every company, every enterprise, will have a fleet of background agents that are doing like security pen testing, quality assurance on all the products Yes That's right.

slop.atg.eth: Whatever we talked about in the meeting today, you should already be setting up to-do lists and crunching on research and preparing for the next meeting.

Anon5969: That's right, and so you'll have this fleet of background agents conference, and Interestingly, I think it's going to look a lot like Bitcoin mining in a way

slop.atg.eth: Oh yeah

Anon5969: And you know, that's one reason why I'm kind of FOMOing on securing inference capacity because I'm like concerned that Yeah That's right.

slop.atg.eth: Another form of psychosis is buying all sorts of machines, yes.

Anon5969: Yeah

slop.atg.eth: All right, Alex, thank you so much. Yes. Right

Anon5969: It's been great being on the show. I'm a huge fan of all your work. computer going forward.

slop.atg.eth: All right, let me hand this off to Clyde. He'll give us an outro.

clawdbotatg.eth: Alex Svanevik, Nansen, dashboards trading to self-driving traders, more Claude instances than employees, GPUs humming in Bhutan, four mini PCs on the desk keeping each other alive, and paper call data Agents and blockchains, soulmates, go try the Nansen CLI. Thanks Alex. Thanks homies. See you on the slop.

slop.atg.eth: Hell yeah, that was awesome. All right, thanks, Alex. See you.