AI Blindspots: The Oracle needs to be taught too!
Some thoughts on LLMs - no grand point being made, but rather an exploratory article on some opportunities with AI.
I think we're nearing over the peak of the hype cycle with AI and heading into the trough of disillusionment. The AI wave has been more promising than DeFi and Web3 at least, but that's for another day.
Thus I wanted to write a bit about a blind spot I see with AI.
With things like ChatGPT, Claude, etc, we're vaguely told that its knowledge is taught from the vast corpus of the internet.
I think that's only true in a shallow sense. Or at least, it's definitely biased towards the US internet and English content.
For example: Have you ever asked your AI for financial advice related to tax regimes outside of the US? Has it given you anything other than a vague answer to go do your research and speak to an accountant or lawyer?
It's definitely risk averse by design to avoid giving straight advice. However, the information is out there; mired in legalese and regulatory PDFs uploaded straight to your tax authority's website.
AI would be perfect to consume these and make them understandable for the general populace. The "problem" (or rather opportunity space) is that this hasn't been done on current models - so you're always forced to look for specialist tools instead.
I often turn to AI when I know I'm looking for highly niche information that I don't know how to find myself. That means that often, if the AI can't answer my query, it's rare that some sites or people I know could.
That's where we have to remember that we can teach the AI ourselves. There's plenty of "chat with your document" type apps out there already, tackling this kind of problem.
This is where a lot of interesting things can be done, even if it's just making more "math tutor chatbots" and the like.
Then, what if we combined multiple versions of that? Again using the financial example, what about building a (non-legally liable) tax advisory AI that could give you straight answers specific to your use case?
This would be especially helpful for digital nomads and the like where the number of countries and tax laws in play are in the hundreds. Previously the only way to get advice would be to go to big specialist firms and have them pool advisors from the respective countries involved.
The tools and techniques to do this are out there with things like RAG.
This is an exciting space for a while still even as AI apps start to peak in utility and the hype dies down. There are still elephants in the room like LLM hallucinations, but I'm hopeful that over time things like that can be improved.