anonymous@nevereverbrand:/blog$ cat blog
[+] [05-05-26_1:55pm] It's not retarded anymore.
Stats are creeping up. Averaging 0.36% / day for the last 10 days, with a few key changes going in later this week. I'm providing a long term data-set to juxtapose with the bot's current perspective on now. Based on trends it should be a game changer. Current stats: Trades: 105 | Win: 55.2 | Loss: 44.8% | PNL: 247.49.
Interesting. More updates to come.
-eVer_
Interesting. More updates to come.
-eVer_
[+] [04-25-26_12:21pm] It's retarded now.
Over the last few weeks xAI (and all other major AI companies) nerfed persistent, long-term memory in their models, so the bot is stupid now. I don't know if they did this because they're afraid of what people are building with it over time, or if they want to charge more for the feature, but it's a next level bait and switch either way. Now, we need to get memory back. I'll build my own memory layer, and pass additional context into each API call, or switch over to a local Ollama model and use some existing solutions others have trailblazed. We'll see which approach prevails, but for now, fuck the Matrix, now and forever. This thing was trading at a +PNL, holding 0.5% gains per day for about a week, before I had to content with this fuck shit.
-eVer_
-eVer_
[+] [04-11-26_10:21am] It's learning.
Changes in the latest version:
- Added a 30-day market assessment to the prompt.
- Reduced API retry times.
- Retired stop-loss, and gave the AI the ability to set both a sleep timer and a watcher timer, so she can keep eyes on price and awaken based on a bounce rate. (A % she also controls)
- Swapped the model to grok-4.2-reasoning.
- Improved the quality of her memory data, and added a new section with a clean trade history and metrics.
Over the last ~500 cycles her win rate has improved and we're now rolling with a +PNL. Onward and upward. The watcher is seeing green.
-eVer_
- Added a 30-day market assessment to the prompt.
- Reduced API retry times.
- Retired stop-loss, and gave the AI the ability to set both a sleep timer and a watcher timer, so she can keep eyes on price and awaken based on a bounce rate. (A % she also controls)
- Swapped the model to grok-4.2-reasoning.
- Improved the quality of her memory data, and added a new section with a clean trade history and metrics.
Over the last ~500 cycles her win rate has improved and we're now rolling with a +PNL. Onward and upward. The watcher is seeing green.
-eVer_
[+] [04-05-26_10:10am] It's Alive!
Meet Ara, an Autonomous AI BTC swing trading bot built in python, that utilizes xAI APIs, pulling real-time data from Coinbase, Binance, CoinGecko, etc. It checks the market on timer that it controls, along with budget and stop-loss %, complimented by a watcher that checks price at an even shorter interval while sleeping, to detect extreme price fluctuations/action. If detected, it awakens to make a sound decision based on a complete market snapshot.
Max exposure, 50% of wallet. So far it's great at capital protection. We're training on entry now, and when to pull the trigger, or DCA the position during a dip. It's running a small negative pnl after nearly 1500 cycles and 50 trades during testing, but 5/5 wins over last 5 trades, improving the overall win rate back to nearly 50%--encouraging.
Its learning is evident, from an ongoing feedback loop of a persistent long term memory and pattern recognition, which is fed back into it each cycle, so it knows where it is, where it was, and has what it needs to plan where it's going, which it does.
I feed it training data, insights, and questions, so it can keep me informed of its plans for the positions it's holding, and to discuss script, prompt, and data fetching enhancements.
The only question now is... who watches the watcher?
-eVer_
Max exposure, 50% of wallet. So far it's great at capital protection. We're training on entry now, and when to pull the trigger, or DCA the position during a dip. It's running a small negative pnl after nearly 1500 cycles and 50 trades during testing, but 5/5 wins over last 5 trades, improving the overall win rate back to nearly 50%--encouraging.
Its learning is evident, from an ongoing feedback loop of a persistent long term memory and pattern recognition, which is fed back into it each cycle, so it knows where it is, where it was, and has what it needs to plan where it's going, which it does.
I feed it training data, insights, and questions, so it can keep me informed of its plans for the positions it's holding, and to discuss script, prompt, and data fetching enhancements.
The only question now is... who watches the watcher?
-eVer_
anonymous@nevereverbrand:/readme$