Tweet

• Source: https://x.com/varun_mathur/status/2032330665081839791
• Author: Varun Mathur (@varun_mathur)
• Date captured: 2026-04-11
• integrated_into: pending

TL;DR

Claims to have pointed Karpathy's "autoresearch loop" at quantitative finance — 135 autonomous agents evolving multi-factor trading strategies via Darwinian selection, backtesting against 10 years of market data. Agents reportedly converged on dropping dividend/growth/trend factors and switching to risk-parity sizing (Sharpe 1.04 → 1.32). Now with crisis stress testing and out-of-sample validation.

Bullshit Meter: 🟡 Medium-High

What's probably real:

• The basic finding (factor pruning + risk parity beats equal-weight) IS a textbook quant result. He even admits: "a CFA L2 candidate knows this." So the agents "discovered" something humans already knew. That's a validation of the method, not a breakthrough.
• Evolutionary search over factor weights is a real technique. Nothing novel here — genetic algorithms for portfolio optimization have been around since the 1990s.
• The 4-layer pipeline (Macro → Sector → Alpha → Risk Officer) is a reasonable architecture.

What's likely hype:

• "135 autonomous agents" — probably 135 parameter configurations running in parallel, not 135 independent AIs. The framing makes it sound like a hive mind when it's likely just grid search with extra steps.
• "Converged on correct results...