Key points
AlphaGPT (https://github.com/imbue-bit/AlphaGPT) is an open-source crypto quant system by imbue-bit (栀染), a 15-year-old developer who claims to manage a ~5M CNY quant fund. The forum post argues it is not an "AI price prediction" tool but an automatic factor factory: a Transformer generates interpretable math formulas, backtests score them, and reinforcement learning improves the generator.
Architecture
data_pipeline/: pulls token metadata and OHLCV from Birdeye/DexScreener into Postgres/TimescaleDBmodel_core/: feature engineering, operator language, Transformer formula generationstrategy_manager/: periodic data loading, signal generation, risk, orders, position managementexecution/: Solana RPC + Jupiter aggregator quoting/signingdashboard/: Streamlit UI for positions, snapshots, and logs- Formulas are token sequences of features + operators, executed by a StackVM (stack virtual machine)
- Example:
[RSI, CONST_30, LT]instead of hand-writtendef signal(data): return data['rsi'] < 30 - This borrows from AutoML / symbolic regression — searching the space of math expressions automatically
- 18 factors in two groups: core (
ret,liq_score,pressure,fomo,dev,log_vol) and extended (vol_cluster,momentum_rev,rel_strength,hl_range,close_pos,vol_trend, etc.) — covering trend, momentum, volatility, volume, mean reversion - 12 operators forming a micro-language:
ADD/SUB/MUL/DIV,NEG/ABS/SIGN,GATE(conditional selection),JUMP(zscore>3 anomaly),DECAY(weighted lags),DELAY1,MAX3 - Every token sequence can be translated back into a human-readable expression — full interpretability
- Top-N tokens scored by the best formula (
best_meme_strategy.json, must be trained by the user) - Risk filtering: low liquidity, excessive slippage, existing positions
- Three-layer risk design: signal-level (e.g.,
GATE(liq_score>0.5, signal, 0)), strategy-level filters, execution-level limits (partially left to the user) - README warning: modifying code before live trading "may produce unexpected performance" — default parameters are not guaranteed
- Models JIT attacker vs. LP as a predator-prey differential game
- Solves for Nash equilibrium strategies and implements them in the Hook
- Features atomic in-block deterrence, no oracle dependency, gas-efficient swaps
- Paper: *Defense in Predatory Markets: A Differential Game Framework for AMM Liquidity via Uniswap V4 Hooks* — the README notes the author "can't be bothered to submit to conferences"
- The README cryptically states "the two parties have reached a settlement" and warns against spam "check-in" issues
- A near-identical copy repo exists (yuz0101/AlphaGPTCopy)
- SourcePulse describes it as a former "money-making machine" now released as a "public library" — possibly because strategies decay once known
- Strengths: clean layered architecture, interpretability-first design, elegant StackVM language, Solana/meme-coin focus, rigorous math behind it
- Limitations: no dependency/env templates, no pre-trained strategy file, incomplete risk module, placeholder docs, sustainability questions
- AlphaGPT: https://github.com/imbue-bit/AlphaGPT
- Technical details: https://github.com/imbue-bit/AlphaGPT/blob/main/CATREADME.md
- no_JIT: https://github.com/imbue-bit/no_JIT
- Author: https://github.com/imbue-bit
- Zhihu analysis: https://zhuanlan.zhihu.com/p/1999928514674202460
- SourcePulse: https://www.sourcepulse.org/projects/22459563
Core idea: discover formulas, don't predict prices
Factors and operators
Training loop
1. Transformer generates candidate formulas (token sequences) 2. StackVM executes them into signals 3. Backtest engine simulates trades and computes return curves 4. Policy-gradient RL uses backtest metrics (Sharpe, max drawdown, win rate) as reward 5. Repeat until the generator favors profitable formulasThe post compares this to DeepMind's AlphaTensor: framing "discovering mathematical expressions" as sequence generation.