QWE AI Academy The Blow-Up That Made Me Stop Watching Demos
July 2026: an AI-infrastructure fund that had reportedly printed 439% after fees in H1 watched the book fall about 67% in a month. Heavy gearing on concentrated longs (and related shorts) forced a discounted sale of public holdings to Citadel. I closed the highlight reels and looked for a desk I could paper-trade with a real ledger – not another one-shot signal script.
The repo that kept sitting at the top of GitHub was virattt/ai-hedge-fund (PyPI: aihf). As of early October 2026 it sat at about 63.9k stars, MIT licensed, latest tag v2.5.0. Multi-agent LLMs plus a portfolio layer. Educational only. Below is what showed up when I treated it like a tiny research lab.
What You’re Actually Running
Paper mode is the point. v2.5.0 (2 October 2026) added a persistent fund object: fake capital, one completed market session at a time, an approval step before anything hits the book, a kill switch, and a hash-chained ledger under ~/.hedge-fund/paper/<name>/. Backtests land in research/. Agents still argue – valuation-style and investor-persona prompts on the same data – then a risk layer shapes targets. No broker path. The README says educational use only, and it means it.
Pro tip: spend week one on data-cost and log audits. Tiny universes. Multi-agent calls burn requests faster than equity curves imply.
From Zero to Paper Ledger
Clean laptop, keys ready: under an hour.
- Install:
pipx install aihf(uv or pip work). Runaihffor the TUI. - First launch asks for a Financial Datasets key (prices, fundamentals, earnings) plus one LLM key – Anthropic, OpenAI, DeepSeek, Google, xAI, Kimi, and similar. Keys go to
~/.hedge-fund/.env. - Build a fund in the TUI: strategies, starting capital, rebalance cadence, small ticker list.
- History path: backtest a date window; session files drop under
~/.hedge-fund/research/. - Live-market feel, zero cash risk: deploy paper, advance session-by-session, approve or reject, open the session viewer, kill switch if agents drift.
Mechanics done. The useful mess starts in the agent logs.
Pitfalls the Clone Tutorials Skip
First liquid-name run looked clean until the logs. News-sentiment often returned total_articles: 0, bullish: 0, bearish: 0 – then stamped NEUTRAL anyway. Missing feed ≠ balanced view. Same pattern shows up in project GitHub traffic (issue #624 and related threads).
Worse pattern: HOLD at 100.0% confidence with reasoning like “No valid trade available.” Reads as iron conviction. Usually means no actionable edge. You have to split “sure about a signal” from “sure that nothing fired” yourself.
Data cost sneaks up. Credits: $20 one-time for 1,000 requests – fine for a short test run. Premium endpoints multiply burn (8× on that tier per the pricing page). One multi-agent pass across fundamentals + news + prices on a modest list chews the pack. Personal is $200/month for 100k requests and longer core history. I emptied a Credits pack validating the paper loop sooner than the marketing math suggested.
Real Desks, Faded Alpha, and This Toy’s Lane
Large multi-strats bolt generative tools onto research and risk stacks; they do not hand final allocation to an unmonitored LLM committee. On the academic side, NBER Working Paper 35273 (Chen, Sialm, Xu, May 2026) tracked roughly 7,896 U.S. hedge funds from 2006-2024. AI-labeled strategies beat non-AI peers by about 6% annualized early on, benchmark-adjusted – then the gap went statistically indistinguishable from zero after 2017, even for early adopters. Lower return comovement was the useful leftover.
Situational Awareness showed the other failure mode: sharp thematic AI book, thin risk staffing, borrowed-money stack. Huge upside print, then a ~67% month and a distressed bid. aihf sits at the other end – zero real capital, every thesis visible, kill switch you own.
Want cleaner curves only? vectorbt-style stacks win. Want seminar-grade agent graphs? research TradingAgents-type codebases stay more academic. aihf’s edge is the fund object that still exists tomorrow morning. Its ceiling matches every LLM trading toy: pattern match on the text and tables you fed it. Durable alpha does not appear from a persona prompt.
Two weeks of paper ticks later I still read each decision before the ledger advances. Coherent valuation stories one session; odd silence the next. That inconsistency is what keeps the exercise honest.
FAQ
Does any of this execute real trades?
No. Paper and backtest only. No broker.
How much does a realistic weekend experiment cost?
Budget the $20 Credits pack plus a few dozen multi-agent LLM calls. Keep the ticker list tiny and total requests in the low hundreds and you can finish a full paper cycle under ~$30. Widen the universe or the history window and you meet the $200/month Personal tier fast. Watch the request dashboard before the surprise invoice.
Why run this if “AI fund” labels stopped guaranteeing edge?
You’re not trying to mint 2015 quant alpha. You’re forcing conflicting theses onto one screen, catching empty sentiment feeds, and feeling a daily approve/reject loop with no cash at risk. Process and limits still decide who keeps money. The simulator makes that concrete – without a Citadel bid on your leftovers.
Next: install aihf, buy the $20 Credits pack, stand up one three-ticker paper fund, advance two real closes, and read every agent rationale out loud. Keep the ledger only if it taught you something you did not already know.