Overview
AlphaGPT (https://github.com/imbue-bit/AlphaGPT) is an open-source crypto quant trading framework by GitHub user imbue-bit, a 15-year-old developer who reportedly manages a quant fund of roughly 5 million CNY. The project is not an "AI predicts coin price" tool — it is an automated factor factory built on deep reinforcement learning, released after having been a closed-source "money-making machine," per community commentary. The author has also published a companion paper on defending Uniswap V4 against JIT liquidity attacks using differential game theory.
Key Points
- Core philosophy: not prediction, but discovery — find patterns that were statistically valid historically and assume they persist briefly.
- Architecture layers:
data_pipeline(Birdeye/DexScreener → Postgres/TimescaleDB),model_core(feature engineering + Transformer formula generation),strategy_manager(signals, risk, orders),execution(Solana RPC + Jupiter aggregator),dashboard(Streamlit). - Factor language: 18 features (log return, liquidity/FDV health, buy-sell pressure, volume acceleration, deviation from mean, volatility clustering, momentum reversal, etc.) and 12 operators (ADD/SUB/MUL/DIV, NEG/ABS/SIGN, GATE, JUMP, DECAY, DELAY1, MAX3) form a small executable language.
- StackVM execution: formulas are token sequences executed on a stack-based VM. Example:
[ret, fomo, MUL, liq_score, DIV, DECAY]computes a decayed, liquidity-normalized momentum signal. This makes every generated factor fully interpretable. - Training loop: Transformer generates candidates → StackVM executes → backtest scores via Sharpe ratio / max drawdown / win rate → policy-gradient RL updates the generator. This mirrors AlphaTensor's approach of modeling mathematical discovery as sequence generation.
- Live execution: strategy scores top tokens, risk filters drop illiquid/high-slippage assets, orders route through Jupiter on Solana, positions persist to local JSON. The README warns that default parameters are not optimal for live trading.
- AlphaGPT repo: https://github.com/imbue-bit/AlphaGPT
- Technical details: https://github.com/imbue-bit/AlphaGPT/blob/main/CATREADME.md
- no_JIT repo: https://github.com/imbue-bit/no_JIT
- Author: https://github.com/imbue-bit
- SourcePulse page: https://www.sourcepulse.org/projects/22459563
no_JIT: The Companion Research Project
The same author released no_JIT (https://github.com/imbue-bit/no_JIT), a Uniswap V4 Hook that models JIT (Just-In-Time) liquidity attacks as a predator-prey differential game under the Hamilton-Jacobi-Isaacs (HJI) framework. It computes Nash-equilibrium defensive strategies and enforces them in-hook, offering atomic deterrence, no oracle dependencies, and gas-efficient swaps. Paper: *Defense in Predatory Markets: A Differential Game Framework for AMM Liquidity via Uniswap V4 Hooks*.
Author Profile
imbue-bit's GitHub profile lists papers spanning deep learning theory, LLM alignment and inference, reward modeling, long-context decoding, hyperparameter loss landscapes, autonomous driving (VLA), and non-stationary deep learning for quant finance, plus solid DevOps/CTF experience. The profile quotes Nabokov's *Speak, Memory*.
Controversy and Release Context
The README notes "both parties have reached a settlement" and warns against ritual "check-in" issues, hinting at an earlier dispute. A near-identical copy exists (yuz0101/AlphaGPTCopy). Likely motivations for open-sourcing a formerly profitable system include strategy decay, reputation-building, or settlement terms.
Strengths and Limitations
Strengths: clean layered architecture; interpretability-first design; expressive 12-operator mini-language; Solana/meme-coin focus where liquidity actually is; strong mathematical grounding.
Limitations: no dependency manifest or .env template; no pre-trained strategy file included (best_meme_strategy.json must be trained by the user); incomplete risk management (position sizing, stop-loss left to users); placeholder documentation.
Deeper Significance
AlphaGPT exemplifies the automation of hypothesis generation in quant research: humans define the search space of factors and operators, while Transformer + RL explore it at hour/day cycles instead of week/month human iteration — the same direction as AutoML and neural architecture search. Its fundamental constraint remains quant's core dilemma: any discovered and publicized pattern arbitrages itself away, which may itself be why the author open-sourced it.