The thing is, a confident smart person who isn’t an expert in anything, who DOES have access to resources to check their answers is still not someone I want doing my taxes.
The only way that an LLM doing my taxes for me is useful is if I trust it to the same level that I trust an expert; and that includes being able to very reliably know when that person is actually confident and when they're not. I haven't seen strong evidence that this is possible with the current LLM approach.
So if we're thinking of LLMs as a confident smart person who sometimes spouts bullcrap -- that's still just not a good fit for this kind of product category. It's a good fit for other tasks, just... not for these.
It's definitely not a good fit for putting in front of problems that are already pretty-well solved. Calculators replaced human calculation, it's a solved problem, you put the numbers into the calculator. Even natural-language input into the calculator is a reasonably solved problem, it's not a task that really requires a full LLM of the scale of GPT-4. So is there value in introducing a human-like source of error into the middle of that process again? I feel like the answer is 'no'.
> These things are solvable, we are still in the 'Commodore-PET' phase of LLM's development
Sure, if we get to the point where an LLM can give such good tax information that there literally is no need to read the tax documents to confirm any of it, then great! But that's not even remotely where we are, and products like this straight-up shouldn't be launched until we're at that point.
This is sort of like selling me an scuba tank and saying, "it might not dispense oxygen so make sure you hold your breath the entire time", and when I say, "well what's the point of the tank then" the reply I get back is, "well eventually we're going to get it to reliably dispense oxygen!"
Okay, but the products are all launching now, I'm being asked to buy into this now, not eventually. The future doesn't change the fact that the current products that I'm supposed to get excited about are literally useless at a conceptual level -- that they do not even make sense given their caveats. I'm not complaining about AI chat bots or funny toys or even toy phone assistants, I'm complaining that companies like Intuit are literally launching tax assistants that give incorrect information and their response is, "well, double check what the bot says." Well that is not a tax assistant then, that's just work. The service that the chatbot is offering me is that I'll have to work more.
Fundamentally, there appears to be a misunderstanding from these companies about what a product is.
The only way that an LLM doing my taxes for me is useful is if I trust it to the same level that I trust an expert; and that includes being able to very reliably know when that person is actually confident and when they're not. I haven't seen strong evidence that this is possible with the current LLM approach.
So if we're thinking of LLMs as a confident smart person who sometimes spouts bullcrap -- that's still just not a good fit for this kind of product category. It's a good fit for other tasks, just... not for these.
It's definitely not a good fit for putting in front of problems that are already pretty-well solved. Calculators replaced human calculation, it's a solved problem, you put the numbers into the calculator. Even natural-language input into the calculator is a reasonably solved problem, it's not a task that really requires a full LLM of the scale of GPT-4. So is there value in introducing a human-like source of error into the middle of that process again? I feel like the answer is 'no'.
> These things are solvable, we are still in the 'Commodore-PET' phase of LLM's development
Sure, if we get to the point where an LLM can give such good tax information that there literally is no need to read the tax documents to confirm any of it, then great! But that's not even remotely where we are, and products like this straight-up shouldn't be launched until we're at that point.
This is sort of like selling me an scuba tank and saying, "it might not dispense oxygen so make sure you hold your breath the entire time", and when I say, "well what's the point of the tank then" the reply I get back is, "well eventually we're going to get it to reliably dispense oxygen!"
Okay, but the products are all launching now, I'm being asked to buy into this now, not eventually. The future doesn't change the fact that the current products that I'm supposed to get excited about are literally useless at a conceptual level -- that they do not even make sense given their caveats. I'm not complaining about AI chat bots or funny toys or even toy phone assistants, I'm complaining that companies like Intuit are literally launching tax assistants that give incorrect information and their response is, "well, double check what the bot says." Well that is not a tax assistant then, that's just work. The service that the chatbot is offering me is that I'll have to work more.
Fundamentally, there appears to be a misunderstanding from these companies about what a product is.