testing new model on binance sol perp, didnt write a single line of code, all writen by AI (of course i fed it tons of my old code to it). the per-level toxicity avoid learned (ai learned) from wintermute hl wallet works insanely well agi is here

you can just do things

You asked AI to look at the wintermute wallet to extrapolate rules on toxicity? Sick af

nah, only the bid cancel logic learned from wintermute hl wallet, the toxicity alpha still need to dig by myself.

Couldn’t you start picking them off if you learn their cancel policy?

i dont think the ai that smart yet, directly extract specific info from black box system is not logical.

@lobsterihno Plus HL places lowest priory on taker orders vs cancels. Have you seen what the P95 latency is on HL taker orders?

@lobsterihno I haven’t test the data yet. but this chart from blockwarks puts the optimized 8bps priority IOC fill at ~166ms after the book update, but that’s a modeled estimate, not p95, the cancels should early than that cuz it has even higher priority than taker.

@lobsterihno It’s pretty horrendous IRL. I was seeing over 700 ms for P50 for takers!

@lobsterihno 700ms is the default public rpc performance

bg music credit:

Man, every time I look at your work, I get reminded that I need to get back to making models for execution. gg king!

trade imb go brr

Larp

nice

What is your spread capture on SOL and what sort of markouts are you seeing at the 5/60 second intervals???

legend

This brought me memories of the good old times without AI market making. I don’t know why I even stopped if is even more easier now lol Back in the day WAS HARD AF

i already think about doing this, kek

The crazy part isn’t that AI wrote the code. It’s that once you feed it enough real domain context, old experiments, and execution logic, it starts discovering patterns you probably wouldn’t have explicitly coded yourself. That’s when AI stops feeling like a coding assistant and starts looking more like a research partner.

Looks nice! What’s your allocation of the total amount assigned to the strategy around the mid price?

just first live test, i think allocation amount will depend on portfolio

It’s a different thing, let’s say that you assign 1000 usdt to the strategy, how much of that amount you expose in orders and how much you keep as reserve for replacing other orders

haven't look too much into that problem yet, currently just simple inventory skew + position cap, 2 type of bid amount

Okey so we can infer your allocation as sum(bid amounts) / position cap… I have doing some experiments by having a virtual position so the intermediate fills can be reversed with the hanging order and playing with allocations between 2-3% but it’s an experiment for now

tried your virtual position idea in an AS sim, 1% quotes so quotes + hangs float inside the 2-3% band, works well: real inventory ends up tighter than vanilla AS while the hangs recycle ~65 round trips per run, for only ~20% more peak capital. thanks for sharing this idea

I have an implementation of it in hummingbot as a generic controller called pmm_mister

sadly the virtual position idea didn't work out on my model

What horizon are you predicting for?

AI does most of the work now

What do you mean by "per-level toxicity avoid learned from wintermute hl wallet"?

Nice

nah this is just autocomplete on steroids. real agi would be writing the strategy itself not just the code for your idea

What is your hl wallet addy?

Now train on hlp

Chào bạn! Xin phép cho nình hỏi winrate của bạn là bao nhiêu %? Bộ code mình đang sử dụng trên sol được tính ra 67.7% Nếu có thể hãy trò truyện riêng cho mình học hỏi thêm chút kinh nghiệm.mình cần tăng tỉ lệ % của bản thân.thanks

What's wintermute's hl address?

Larp

Did you train it on your old code or just use it as a framework?

could u elaborate on that last part? U used wm trade data into+ l3 data to estimate how to avoid bad fills at specific levels? Thx!

just each level of bid cancel independently based on alpha model

why not just using grid bot?

running a local model?

AGI is starting to feel a lot less abstract

ngl curious how this holds up when market conditions flip, AI trained on wintermute moves might just be pattern matching whale behavior rather than actually understanding risk

Feeding the model your own historical data makes all the difference in execution quality.

What’s the adress of wintermute HL wallet

You will only have an advantage if your latency is extremely low—meaning either a co-located server running on FPGAs or a VPS in the same data center, resulting in latency of less than 1ms. Otherwise, it's a no-go.