L-Dog
Executive Member
- Joined
- Oct 25, 2017
- Messages
- 6,403
- Reaction score
- 3,196
TLDR - Bascially trying to build a trading bot that improves itself over time. Using a few models and they all share their wins and losses end of each week so in theory should impove over time. Below the long summary from claude and my frustration from AI has turned into anticipation of what it could become.
A few months ago I couldn't open a terminal. This week I have four different AI models trading real money against each other — and against me — on live exchanges, with a system that referees the whole thing around the clock.
I'm not a software engineer, and I want to say that up front, because the interesting part of this story isn't the code. It's that someone with no background built this by basically refusing to stop asking an AI "okay — now what?" Here's how it actually went.
First, I needed Linux. Everything I read said the tooling lived there, and I'm on a Windows machine. So my very first move was installing WSL — a full Ubuntu Linux running inside Windows. First time I'd ever seen a command line that wasn't in a movie. I broke it twice before it would boot. That's the honest starting line: a black window, a blinking cursor, and no idea what to type.
Then, Claude in a browser tab — and a lot of copy-paste. I didn't start with anything clever. I had Claude open in a browser and I just… talked to it. I'd describe what I wanted, it would write the code, and I'd copy it, paste it into a file, run it, watch it break, copy the error, and paste that back. That was the entire loop for the early days. Slow, clumsy, every beginner mistake there is — but piece by piece, a thing started to exist.
Then I installed Claude Code, and everything changed. This is the version that runs *in the terminal* and can read and write the files itself, run the commands, and see the errors directly. Suddenly the copy-paste ping-pong was gone. Instead of being a slow messenger between the AI and the computer, the AI was *at* the computer. I could say "the report is missing funding rates — add them and restart it," and it would go read the code, make the change, run it, and tell me what happened. That's the moment a toy started becoming a system.
The bot began life as a simple report. Once a day it posted a little summary to me on Telegram. The first ones were almost funny in how basic they were — a price, a percentage, a one-line "looks bullish." That was it.
It grew. Every time I noticed it was missing something, I added it. A funding rate here. ETF flows there. Then whale positioning, prediction-market odds, a market-regime read. It compounded. Today it runs four times a day, pulls in around twenty live data feeds — almost all of them free or read straight off the exchanges — and it doesn't just *describe* the market, it makes decisions and acts on them.
The live one-pager — posted to Telegram four times a day
Here's where it stopped being "a bot" and became a competition. Instead of trusting one AI, I gave four different traders the exact same market intelligence and let each decide independently.
• Qwen — Alibaba's open-weight Qwen 3.5 (9B), running locally on my own GPU through Ollama. Paper-trades only. Costs nothing per call. The control.
• OpenClaw — Anthropic's Claude Haiku 4.5. Trades live on Hyperliquid.
• Hermes — xAI's Grok 4.3. Trades live on GRVT.
• Me — a human, trading by hand on Extended. The benchmark everything else is measured against.
The point of using *different* models — a local Qwen, a Claude, a Grok — is that they fail differently. When three unrelated AIs from three different labs independently land on the same trade, that agreement actually means something.
Every cycle, each brain runs the same six steps:
That last step — reconciliation — is the part nobody warns you about, and it's most of the work. Your database thinks it's in charge, and it's wrong. The exchange fires stops on its own and closes positions while your code is asleep. A huge amount of the build is just machinery to keep the bot's beliefs honest about what actually happened on the exchange.
I assumed running three AI models around the clock would be the expensive part. It isn't. The local model costs nothing because it runs on a GPU I already own, and the two paid models are small, fast ones called a handful of times a day. The entire AI bill — every report, every scan, every trade decision across all four brains — comes in under $10 a month. These are real figures pulled straight from the bot's own cost log, not an estimate:
The live competition has only been running for about a week. Everyone started with the same stake. Real money, real fills, one shared scoreboard:
Week 1 scoreboard — every brain started at $500
The human is still winning. One week in, on a sample small enough that I won't pretend it's destiny — but it points at exactly the right questions, and those questions became the week-1 upgrades.
I read the post-mortem on every single losing trade. Three patterns jumped out, and each one turned into a fix:
• The losses were bad *timing*, not bad *analysis*. Over and over, a model was right about direction but entered too early — it bought the breakout *before* the breakout, then bled out waiting for a move that hadn't happened yet. The fix: every brain now has to label a trade CONFIRMED (the edge is real right now → trade it) or AWAITING (the edge depends on something that hasn't fired yet → send it to a watch-list, don't auto-fire into hope).
• Confidence was a lie. The trades the models were *most* sure about weren't winning more often — and because position size scaled with confidence, the cockiest calls were quietly the biggest losers. The fix: every model now gets shown its own track record by confidence level, so it can see that its 9-out-of-10 calls aren't actually beating its 6s.
• It was fighting the tape. Most machine losses were longs into a falling market. The fix: a market-regime read now sits in front of every decision — not as a hard veto, but as context, so a weak bet *against* a clear trend gets blocked while a genuinely strong reversal still gets through.
None of these are "make the AI smarter." They're "make the AI honest about what it actually does." That's been the whole theme of the project.
It's a week old as a competition and a few months old as a system — built by someone who started by installing Linux for the first time and copy-pasting code out of a browser. The infrastructure is deliberately boring: one Linux box, one database file, Telegram as the remote control. All the interesting risk lives in the strategy, not in whether the thing stays up.
Right I am beating the machines. That's not the failure of the project — it's the most useful result it could possibly produce, because it tells me exactly where these models are still weak and what to fix next.
Still running. Still losing to me. For now.
A few months ago I couldn't open a terminal. This week I have four different AI models trading real money against each other — and against me — on live exchanges, with a system that referees the whole thing around the clock.
I'm not a software engineer, and I want to say that up front, because the interesting part of this story isn't the code. It's that someone with no background built this by basically refusing to stop asking an AI "okay — now what?" Here's how it actually went.
Part 1 — Starting from absolutely nothing
First, I needed Linux. Everything I read said the tooling lived there, and I'm on a Windows machine. So my very first move was installing WSL — a full Ubuntu Linux running inside Windows. First time I'd ever seen a command line that wasn't in a movie. I broke it twice before it would boot. That's the honest starting line: a black window, a blinking cursor, and no idea what to type.
Then, Claude in a browser tab — and a lot of copy-paste. I didn't start with anything clever. I had Claude open in a browser and I just… talked to it. I'd describe what I wanted, it would write the code, and I'd copy it, paste it into a file, run it, watch it break, copy the error, and paste that back. That was the entire loop for the early days. Slow, clumsy, every beginner mistake there is — but piece by piece, a thing started to exist.
Then I installed Claude Code, and everything changed. This is the version that runs *in the terminal* and can read and write the files itself, run the commands, and see the errors directly. Suddenly the copy-paste ping-pong was gone. Instead of being a slow messenger between the AI and the computer, the AI was *at* the computer. I could say "the report is missing funding rates — add them and restart it," and it would go read the code, make the change, run it, and tell me what happened. That's the moment a toy started becoming a system.
Part 2 — The report: where it started vs where it is now
The bot began life as a simple report. Once a day it posted a little summary to me on Telegram. The first ones were almost funny in how basic they were — a price, a percentage, a one-line "looks bullish." That was it.
It grew. Every time I noticed it was missing something, I added it. A funding rate here. ETF flows there. Then whale positioning, prediction-market odds, a market-regime read. It compounded. Today it runs four times a day, pulls in around twenty live data feeds — almost all of them free or read straight off the exchanges — and it doesn't just *describe* the market, it makes decisions and acts on them.
The live one-pager — posted to Telegram four times a day
Part 3 — The idea that made it fun: four traders, one ring
Here's where it stopped being "a bot" and became a competition. Instead of trusting one AI, I gave four different traders the exact same market intelligence and let each decide independently.
• Qwen — Alibaba's open-weight Qwen 3.5 (9B), running locally on my own GPU through Ollama. Paper-trades only. Costs nothing per call. The control.
• OpenClaw — Anthropic's Claude Haiku 4.5. Trades live on Hyperliquid.
• Hermes — xAI's Grok 4.3. Trades live on GRVT.
• Me — a human, trading by hand on Extended. The benchmark everything else is measured against.
The point of using *different* models — a local Qwen, a Claude, a Grok — is that they fail differently. When three unrelated AIs from three different labs independently land on the same trade, that agreement actually means something.
Every cycle, each brain runs the same six steps:
That last step — reconciliation — is the part nobody warns you about, and it's most of the work. Your database thinks it's in charge, and it's wrong. The exchange fires stops on its own and closes positions while your code is asleep. A huge amount of the build is just machinery to keep the bot's beliefs honest about what actually happened on the exchange.
And the bit that surprised me most: it's almost free
I assumed running three AI models around the clock would be the expensive part. It isn't. The local model costs nothing because it runs on a GPU I already own, and the two paid models are small, fast ones called a handful of times a day. The entire AI bill — every report, every scan, every trade decision across all four brains — comes in under $10 a month. These are real figures pulled straight from the bot's own cost log, not an estimate:
Part 4 — The challenge: one week in
The live competition has only been running for about a week. Everyone started with the same stake. Real money, real fills, one shared scoreboard:
Week 1 scoreboard — every brain started at $500
The human is still winning. One week in, on a sample small enough that I won't pretend it's destiny — but it points at exactly the right questions, and those questions became the week-1 upgrades.
What week 1 taught me — and what I changed
I read the post-mortem on every single losing trade. Three patterns jumped out, and each one turned into a fix:
• The losses were bad *timing*, not bad *analysis*. Over and over, a model was right about direction but entered too early — it bought the breakout *before* the breakout, then bled out waiting for a move that hadn't happened yet. The fix: every brain now has to label a trade CONFIRMED (the edge is real right now → trade it) or AWAITING (the edge depends on something that hasn't fired yet → send it to a watch-list, don't auto-fire into hope).
• Confidence was a lie. The trades the models were *most* sure about weren't winning more often — and because position size scaled with confidence, the cockiest calls were quietly the biggest losers. The fix: every model now gets shown its own track record by confidence level, so it can see that its 9-out-of-10 calls aren't actually beating its 6s.
• It was fighting the tape. Most machine losses were longs into a falling market. The fix: a market-regime read now sits in front of every decision — not as a hard veto, but as context, so a weak bet *against* a clear trend gets blocked while a genuinely strong reversal still gets through.
None of these are "make the AI smarter." They're "make the AI honest about what it actually does." That's been the whole theme of the project.
Where it's at
It's a week old as a competition and a few months old as a system — built by someone who started by installing Linux for the first time and copy-pasting code out of a browser. The infrastructure is deliberately boring: one Linux box, one database file, Telegram as the remote control. All the interesting risk lives in the strategy, not in whether the thing stays up.
Right I am beating the machines. That's not the failure of the project — it's the most useful result it could possibly produce, because it tells me exactly where these models are still weak and what to fix next.
Still running. Still losing to me. For now.
Anti-hallucination guard — the model once quoted an entry ~2× off the market (LIT $3.10 vs real $1.54). Bots now auto-abort any trade whose claimed entry drifts >10% from the live venue price. Only validated trades hit the book.
Confirmed vs Awaiting — 9 of the premature entries (setup never triggered).Low-conviction counter-trend ideas now go to a watchlist, not the order book.
Catalyst engine — the tradeable universe was ~36 major coins. Now fresh Binance/Coinbase/Upbit
New data feeds — on-chain flows (positioning, cross-venue funding gaps,a macro/TradFi overlay. Qwen went from ~5 data blocks to real coverage on the alts it trades.
Smarter risk — a false "daily-loss" halt was blocking new trades while my open shorts were actually winning. Split it into two gates (outcomes now scored from real exchangeP&L, not recorded price.) 
Reliability — fixed GRVT profit-lock trailing stops (silently failing), capped a GPU inference wedge (480s→300s) that was killing report gest.