Claude Opus 5.5 Is About to DOMINATE Kalshi & Polymarket transcript

So, just to show you that title is not clickbait, you can kind of see if we scroll down here how high will the 30-year Treasury yield by January 1st. And you can see I entered here at 6 cents, 8 cents, 11 cents. Now it's at 85, 65, 69. And you can kind of see the gains here. So, 900%, 875, 455. And yeah, that is what we did in just one day using Anthropic's new Claude Opus 5.5. So, today I wanted to go over three ways I have been using Opus 5.5 over the last couple of days since it came out.

And I wanted to cover three different platforms. So, Kalshi is going to be what you saw now, how I kind of find this or how I use kind of Opus 5.5 to find these mispricings and how we actually capitalize on this. On Hyperliquid, I am building like this voice trader you can see here. So, I built this voice trader here that also has some kind of yeah, automated learning system. [music] So, it tries to give us like an AI assistant here.

But I'm going to show you we can actually control Hyperliquid and enter trades with our voice. I'm going to show you that later. And I also have been starting my Opus 5.5 auto ML researcher that is kind of based on Cartwheel's auto research on Polymarket 5-minute Bitcoin up and down. I guess I can show you quickly. So, here is kind of the dashboard, right? So far we just started, so we have run seven experiments so far.

We did find an improvement here. So, we did improve over the baseline. The baseline is of course the market at decision time. I'm going to explain this in a video Opus has created for me that's Yeah, you can see we have all the rejections here, the experiments. And the one Yeah, I'm going to explain this in a video because Opus 5.5 is also really good at creating explainer videos. So, I created an explainer video for the mispricing findings on Kalshi that made us like 700% on average.

And for the Opus 5.5 ML researcher. So, I think we're just going to start. I'm just going to play you how I I've been trying to do those auto ML researchers for a while, but now I kind of got back to it. So, I think I'm going to play you that video first to kind of explain how this works. So, I'm really excited to see where we end up when we have run this for like, yeah, a bunch of days and see how much we can improve our machine learning model here.

So, yeah, let me just play you the auto 5.5 ML researcher video and I can comment on it at the end. >> [music] >> Every 5 minutes, Polymarket asks the same [music] question. Will Bitcoin finish higher? 288 times a day. We built a machine that researches how to price it better than the crowd. On its own. The idea comes from Andrej Karpathy's auto research. Give an AI agent a training script, [music] a fixed 5-minute budget, and one number to improve.

It edits the code, trains, keeps what helps, and throws away what doesn't. And it never stops. [music] In markets, that loop has a flaw. Noise looks like progress. [music] An agent that keeps every tiny gain will end up fooling itself. >> [music] >> So, we split the machine in two. A researcher that proposes and a sealed evaluator. It can read, but never touch. The evaluator holds 42,000 markets, order [music] books, trades, and the Chainlink prices that settle them.

Every market settles on a 60-second average. In the final 90 seconds, that average starts to lock in. That is where we hunt for edge. Each batch, Claude writes its hypotheses before any results exist. It codes them, tests them, [music] and hands them over. The evaluator replays the future one row at a time. No look ahead, no labels, no internet, nothing to cheat with. A good score is not a promotion. A result must survive new seeds, a bootstrap across days, every test period, [music] and a sealed confirmation week.

Only then does it become the champion. The first champion was the market itself. >> [music] >> The first challenger to beat it was simple. As the average locks in, the market is under confident. Recalibrating it cut log loss from 0.1278 to 0.1262. It passed every test. The agent's next five ideas, four rejected. One looked better [music] until the bootstrap called it luck. Most ideas fail. That is the point. The machine's job is not to be clever.

It is to be honest and to keep going. Profit after fees is not proven yet. >> [music] >> Finding out is the next experiment. >> So, yeah, that's a pretty good video, right? And the sick thing is I just fed Opus some assets. Uh I pointed it at the the repo for um Auto Research and it made this explainer video for me. It's just insane good at doing this uh animations and putting everything together. So, let's take a look now.

You can see this we have run 13 experiments. I don't know really what to expect. You can kind of see we have had some close calls. This was like a falsified improvement. So, the idea behind this is it's going to be autonomous. So, it's just going to keep running and you can see we have this hypothesis. It is actually run by Opus 5.5. So, you can see experiment 14 is the markets under confidence depends on price level.

And we have some Yeah, I'm not going to read out this. Added to the champion, improved log loss slightly, mostly at 240. That's the time, right? So, this is just going to keep running all of these experiments and we're going to give it like a score and it's going to be rejected, falsified, or promoted if it's good enough. Yeah, it's a bit all over the place here, but you can basically see here is everything here is from the log, right?

All the experiments we have run. And lessons, this is important because we need to contain everything we have tried so that the next experiment is not duplicated. We also have a summary for kind of improving the memory of the research loop. But I'm not going to go into like insane details in this now. I maybe do a video on it again if I found that that we found something really interesting. And maybe I can do a more in-depth video.

Maybe we can start up the machine learning model on the 5-minute Bitcoin. Okay, so this looks promising, right? We scored here. Now we just need to be validated. So, maybe we get an improvement here. We can wait for this and see if it's improved. But while we wait for that, let me just show you the I guess I can show you the voice trader now because this is really cool. So, let me just go to HyperLiquid and I can kind of show you how this works now.

Okay, so you can see I'm logged in on HyperLiquid here now. So, if I bring up this now, you can see this is my bot now. I'm just going to fire up the demon here so we go live. Okay, so you can see now we are live, right? And here you can kind of see on HyperLiquid, this is kind of the Yeah, the daily here and it says lean short now. So, I just I can prepare now and I can just try to follow what it says here on the 60 second.

So, let's just try it and you can see the positions will pop up down here. Go long. >> Long working. >> And you can see we placed kind of a position here, right? And let's see if we enter because we do market, we place like an order. >> Long missed. >> Okay, so we missed that one. Let's try again. Go long. >> Long working. >> Okay, so we you can see we placed order here and now we enter, right? And you can see it doesn't really matter how it goes now, but basically now we entered here. >> Going long. >> And we we were got filled and you can see it now we can kind of control um this with our voice, right?

I don't know what the way it's going to go now. We just We just going to keep running it for a while and we're going to close it, but the the point is not how we do in the trading here. The point was to build this pretty cool actually um uh voice trader. So, now I can kind of We can just try to close this. Just We're going to lose money here, but it doesn't really matter. Close. >> Closed. >> Okay. Go short. >> Short working.

Going short. >> So, we can be pretty fast here, right? So, we try to place a position here. Let's see if we can fill. Now, we kind of turn the other way. You kind of get the point, right? It is pretty cool. I will We did enter. Uh yeah, it's going to go horrible now. I don't really care of the way it works. I'm I just want to show you how it this works now. So, I had a lot of fun with this and the speed, you can kind of see the speed here.

It's really fun to play around with it. Uh is it like very profitable? I don't think so, but I was really happy how Opus kind of solved creating this app here and the way it works. So, you can see we have like a Binance stream and we log all the the suggestions from the AI assistant and it tries to learn over time, right? That's why it has these suggestions here. So, we could turn it off if we wanted to. Uh but yeah, let's just exit here now.

It doesn't really matter. Close. >> Closed. >> So, it's pretty fast. That's not bad. So, yeah, that was the voice trader. I'm going to keep developing this a bit and you can see we also had like a $3,000 entry here, uh 20x levered, uh with a margin call at yeah, I couldn't see the margin here, but uh yeah, uh pretty fun and a lot of fun to just play around with if you kind of want to get more into this discretionary trading.

Is that what they call it? I think so. So, yeah, but uh now let's go over to kind of what happened the day Opus came out or was it yesterday? I can't really remember. Uh how I kind of found this mispricing, how I kind of used Opus to scan all markets to look for these mispricings and yeah, I made a video on that too. So, I guess I can just play you that first and we can talk a bit about it. >> On September 24th, a Treasury contract on Kalshi was selling for 5 cents.

Our research said it was worth closer to 60. The backdrop. The 30-year Treasury yield had just closed at 5.4%. It's highest since 2004. A week after the Fed raised rates. Kalshi lists a ladder of contracts on how high that yield will go by January 1st. The ladder broke at 5 and 1/2%. Above 5.49, traded near 88 cents. Above 5.51, almost the same question was offered at 5 cents. Two basis points apart. 80 cents apart in price.

The edge is simple. Fair probability minus price. [music] The 30-year sat just 14 basis points below the threshold >> [music] >> with more than 60 trading days to go. Historical daily moves put the odds of touching it at roughly 55 to 65%. Against a 5-cent price, that is about 50 cents of edge on every $1 contract. Then, we tried to prove ourselves wrong. The same contracts had traded at 46 to 72 cents the day before.

The yield itself had just jumped 11 basis points in a single [music] session. And the order book showed the source, one single 100-contract order. Not a market view, a stale quote. We read the settlement rules word for word, placed a small test order, confirmed the fill, then bought 444 contracts across three strikes at an average of about 10 cents. The next day, the 30-year fixed at 5.47% and the market caught up. Those contracts now trade between 65 and 83 cents.

The lesson, mispricings like this are rare, small, and gone within hours. Read the rules precisely, verify every fill, and let a model, not a feeling, decide what fair is. >> [music] >> Yeah, so that was a bit interesting. I guess it didn't mention too much how I actually did it. So, basically, what I did is I put up this It's some type of scanner that look kind of looks at all the markets, but you need to do it in a very efficient way with semantic search and stuff like that.

So, if you kind of look at let's say let's take a more simple position. Let's just take this one. If you kind of look at the market here, you can see we have actually some liquidity here. We have one at 86 and we have We have an ask at 86 for 100. And we have an 85. So, I'm thinking I'm just going to probably just going to hold this uh a bit more uh to see if we can kind of push it a bit more. And I just want to develop this and see where this goes.

Uh let's see about this. Is there liquidity? There is very low liquidity. So, we can't really get out for Yeah, maybe we can get out with everything here, right? You can see there are some market movements here, but not too much. But the point was that it not exactly my uh market. Um I wanted to kind of talk a bit about just how efficient Opus was to actually find this setup here. So, like I said, I basically just asked it to build a scanner to look for these opportunities.

And what is very important when you kind of do that is looking at kind of the semantics of the contract. And yeah, you you kind of get what you what I mean if you have seen these uh prediction markets. And you can kind of compare it, and you can also try it maybe try to do some cross arbitrage with this and see on maybe Polymarket or other places uh where there are some interesting stuff. So, there are opportunities out there still if you ask me.

So, let's just check in again. Now, you can see this was falsified. The one we were looking at here. So, we haven't really improved yet on our auto research, but it's very early yet. Uh but yeah, I just wanted to say how happy I've been with Opus 5.5, the speed. And I really enjoy it kind of looking for good opportunities both on Hyperliquid, Polymarket, and Cashi. And I've been also kind of looking at more standard options and stuff like that.

Also some sports book I've been using it for and with some pretty good luck, I think, actually. So, yeah. Um that's kind of what I wanted to share, my initial experiments with Opus. And yeah, I had a lot of fun. You definitely got to try out to do some animation videos if you haven't tried that. >> [snorts] >> So, yeah. If you like this kind of content, I'm going to try to bring you a few more examples, and I'm definitely going to come back with a video if we find something interesting here on the auto research loop.

I'm going to do like a dedicated video on it probably. So, yeah. Thank you for tuning in, and have a nice weekend, and hopefully I'll see you again probably on Sunday or something. So, yeah. See you.