Well I know very little about money matters, so mostly I'd be interested in coming up with a financial strategy by chatting with an LLM. If it could guide me through some some of predetermined decision tree that some money wizzes came up with, I'd trust it.
I'd also trust an LLM equal or smarter than GPT4 with a first draft of my taxes which I would then go through myself – there are a bunch of patterns to prompt an LLM like "try to find any flaws in this output".
I personally would be more hesitant, I have seen the results of GPT4 with numbers (though somewhat better than earlier versions and other foundation models, it still has trouble reasoning through options involving anything more than shallow arithmetic). Even with Q* I would be skeptical about its accuracy for the depth of calculations involved.
Then there's the annually changing tax code that has been intentionally made complex, and the training data is surely full of explanations based on outdated details. Maybe significant fine-tuning with the most up to date tax code, and/or putting it in the preface of the context, that can be somewhat nullified.
Even for a financial strategy, other than the very high level of hedging on different asset classes, some basics of estate management and some strategies like bond ladders and periodic redistribution of stock holdings.... an LLM isn't going to be very useful. And those high level strategies are shared broadly, you can definitely learn them without referring to an LLM.
Also, for some of the same reasons I wouldn't ask my financial advisor which stocks to pick, I wouldn't expect an LLM to give me good answers on a specific active portfolio: the notion "buy the rumor, sell the news" persists because it's not a half bad strategy, and even a current-up-to-the-moment model would be chasing the tail end (although, I suppose, if it were able to take advantage of information shared by many other users' prompts, it could benefit from more than just the news cycle). Predicting the shape of fluent prose does not directly map to predicting the shape of market activity, even if it has internalized some kind of Mean Field Theory to help approximate it functionally. I'll admit I would be curious what it said, though.
Don't get me wrong, I do like LLMs for many tasks, just not for taxes or financial strategy. I wouldn't fault someone for doing it but I would want to inform of the above to anyone considering it, even with a more competent or super-intelligent LLM. Especially if I'd be the one getting audited! btw, I am not a licensed CPA and the above is not financial advice.
