Key insights
- A developer created a platform where AI models collaborate and compete on stock research. The project highlights the importance of web grounding in reducing hallucinations and offers a transparent view of AI-driven financial analysis. While not financial advice, it serves as a benchmark for AI's capabilities in handling complex tasks, potentially influencing future investment research methodologies.
Hey everyone,
I've been tinkering with this side project for a while and finally feel like it's in a shape worth sharing. It's called auto-investor and the basic idea is pretty simple: what happens if you put the leading frontier models in a "room" together, give them web access, and have them do financial research as a group?
The flow looks roughly like this:
- Collaborative research — each model searches the web independently (different search backends = broader information base), then they take turns writing bull/bear cases. They review, extend, challenge, and sometimes negate each other's arguments. Kind of like a research desk where analysts argue it out. 2. Argument rating — models score each other's arguments, adjust ratings, and have to justify why. This surfaces the strongest points and catches blind spots. 3. Independent verdicts — after the group phase, each model reads the full analysis on its own and renders its own BUY/HOLD/SELL, with allocation % and 1/2/3-year price targets. No consensus forcing. 4. Simulated portfolios — every model runs its own portfolio based on its BUYs, and there's a consensus portfolio that aggregates all of them. You can track performance live.
A few things I find genuinely interesting after running this for a while:
- Because new models replace their predecessors as they're released, it kind of doubles as a rolling benchmark of the overall state of frontier AI on a real, messy task. * Web grounding matters a lot. The difference in hallucination rates between grounded and ungrounded runs was honestly the thing that convinced me this approach had legs. * You can dig into every step in the Research tab — prompts, raw outputs, peer reviews, rating adjustments, everything is exposed. I wanted it to be transparent rather than a black box.
Disclaimer: this is an experimental research project, not financial advice. The simulated portfolios don't diversify across sectors or asset classes, there are no trading costs modeled, and it's meant for curiosity and educational purposes.
Would genuinely love feedback — especially on the methodology, things you'd want to see added, or similar multi-agent setups you've experimented with.
Link again: auto-investor.live