Anecdotally, current models seem to be decent at general personal finance principles - certainly better than the majority of personal finance education that people get exposed to unless they seek it out and read a variety of books and sources. But I wouldn't trust them with direct decision making with actual money due to the training lag time on current tax policy, etc.
certainly better than the majority of personal finance education that people get exposed to unless they seek it out and read a variety of books and sources.
Also, models are now good enough that you can give them chapters from "authoritative" books, and they'll integrate that and come up with better answers even if their "vanilla" answers were average. And they'll tailor stuff to your particular situation. It's funny that the "agentic" stuff is only used in coding mostly, while it can and does work in other fields as well.
As always, you kinda need to check it (at least spot check) but all in all I'd agree it's better than the average stuff you used to find with a quick google search.
I've done this on different topics - I know the answer is in a particular eBook/PDF/document, but for whatever reason it's not trivial to look it up. The model can do it a lot more quickly than I can, and then I can still verify the accuracy.
Depends on the harness too. The MS copilot 365 chat interface happens to parse PDF and docx files before handing them to the model. Absolutely fantastic if your PDF is 1300 pages of concatenated reports and you want to find a single detail in it that you can't easily ctrl+f for.
Some of the more widely spread and tolerated LLM outputs seem to be AI slop replacing Journalism/Blog slop. Places where people complained about quality already, but tolerated it if important enough.
So to name some of the more common ones Translation, Summarization, and Reiteration of a source material.
Humans put spin on things, how much you trust a source might not reflect the source's factual accuracy. It might just mean you liked reading it better from one source than another.
It's basically a fancy context aware ctrl-f to the book.
I use this regularly with RPG manuals. I _know_ the stuff, but don't remember every detail by heart. And just ctrl-f:ing through a Mörk/Pirate Borg -style PDF isn't really productive (they're "artistically" laid out). But I can just ask an AI bot that has the pdf indexed like "how does the medical kit work?" and it'll give me a summary along with the relevant rolls within seconds.
Important to note is that what is being measured here is the ability of the models not of the chat tools themselves, which combine model completions with other tools that the models can call upon. The mainstream labs already know this about models, it's no secret, and in fact training materials from e.g. Anthropic are at pains to point out that users, or analysts designing workflows, have the reponsibility to ensure the correct tools are used and that human verification takes place at appropriate stages depending on the risk/consequences of the task at hand.
Of course a language-completion model with a training cutoff date won't have up-to-date information on tax rules or the ability to carry out correct numerical calculations, but when you combine that with (in Claude terminology) web search and code execution tools invoked by the chat agent, you immediately have much more reliable results.
I keep finding that the current harnesses, when encountering syntax that was invalid at training time but is now valid due to new language versions or custom extensions, don't correctly figure out why and assume something is wrong with the codebase or toolchain. I would hate to have that happen with my taxes.
Single shot or with reasoning enabled? My experience is that reasoning dramatically reduces hallucinations and improves output quality. I don't trust models without it.
These models do pretty well in benchmarks and real world so I'm highly suspicious of this article. Further more, in the original report, the examples of bad answers are from Haiku - at least 7 out of 10. Anyone who knows anything about LLMs know that haiku shouldn't be used for anything pretty much.
There's no reproducible set either. I'm not gonna trust this report.
Most people[1] interacting with chatbots don't have a paid subscription and they do interact with the free-tier LLMs that are Luna and Haiku, so I still think it's relevant.
[1]: not on HN obviously, but IRL, and probably among FT's readership as well.
I don't have the time to review the underlying research and decide which one is more correct. My personal biases make me want to believe the current one. Your personal biases may be pulling you in the other direction. How do we make the conversation more intelligent than that?
Given most financial advisors tend to vend out suboptimal advice and steer customers in favour of products they receive a kickback for, I'm happy to be accepting of an unbiased LLM that's trained on bogleheads.org.
Some people do have complex financial situations. It's not as simple as that.
For example in the UK (and maybe US?) you get tax relief for money you put into your pensions, but there's a limit of £60k/year. Unless you earn a lot (which I do, yeay) when that limit is tapered. Except that you can also use up to 3 years of previously unused allowance. But you have to use this year's first.
Also interest is taxed, but you can put up to £20k/year into an ISA which isn't. And if you still want to avoid some tax you have kids ISA's and even pensions!
Then there are also startup investment schemes that save you some tax. Those seem to be not worth it, but you get the idea - it can be complicated. Especially if you are near one of the many tax/benefit thresholds.
The marginal tax rate in the UK bounces all over the place - it's even technically possible for it to be over 100%!
Though you don’t directly say it, your comment strongly implies that you can’t buy fractional shares of $VFIAX. (You can, same as $FXAIX, $VTSAX, etc.)
In my case, I am double-taxed (both Japan and US side) on capital gains. Tax treaties reduce, but not eliminate, the extent of double-taxation.
Many US-based brokers do not allow Americans abroad to purchase mutual funds, so VFIAX is not a choice for me.
Maybe I can go with eMAXIS Slim All Country... Oh, but that is a PFIC under IRS rules and I'd be taxed on unrealized capital gains. So I guess no Japan-equivalents of VT for me. That's fine, I guess I'll just buy VT in my US-based brokerage account; but now I'm in a suboptimal spot with respect to monthly contributions, calculating JPY-denominated income tax on dividends, etc.
I even made an implicit assumption when I said "calculating JPY-denominated income tax on dividends". That assumes your tax status is permanent resident. If your tax status is non-permanent resident, then a decent financial advisor will recognize that only the extent of income remitted to Japan gets taxed, so VT distributing at all isn't an issue (until 5 years later). What should you do before the 5 year threshold is hit? etc. etc.
But yes, if you're born in America and plan to stay within the same state for the rest of your life, then a 100% automated setup that simply deposits $1,000/mo into VFIAX is probably fine. (But keep in mind, to most non-Americans, VOO is not really diversified compared to funds like VT).
Otherwise, there is value in consulting someone (or something) that regularly handles taxes and financial planning.
Full agreement. I'm double-taxed in Germany and the US. I don't have a residency in the US (so I cannot vote but still get taxed, nice!) so I'm completely unable to buy VFIAX or any other fund from the US. I cannot but them over an EU brokerage, either, because US ETF and its ilk violate European transparency laws. I can buy EU-domiciled funds, but those are treated as PFIC in the US with prohibitive taxation. I relegated myself to an old-school dividend growth investment scheme and it's financially working out much better than I anticipated but still, I'd like to have a core of a stupid ACWI tracker.
It depends though - I'm not in the US and in my country ETFs get taxed continually whereas mutual funds don't get taxed until you sell.
In general, invest in low-cost index funds is pretty solid advice everywhere. In different countries you might use slightly different instruments due to tax advantages (like the ISA in the UK etc.)
Much of the testing is on Haiku and Luna, and criticizing the quality of free AI (!). But they do claim Opus 5 with reasoning still failed 39% of their financial questions.
So I downloaded that report which of course doesn't contain the most relevant information (the questions) but it contains some examples of wrong answers.
I fed the first question to Grok (which they claimed they tested as well) and it answered it correctly in detail.
I repeated it with another one - again correct answer. I then selected the question they said Grok specifically answered incorrectly and it again answered it correctly.
I am sticking with my first intuition: people are terrible at testing tools and probably wanted them to answer incorrectly/not fully (the questions are constructed in a way to make it difficult as well). They also have vested interest in the conclusion (they are financial advisory firm) so there is that to consider.
People reading ft will now think chat boxes are bad at answering financial questions while they are pretty good at it. Zero consequences for spreading fake news for Financial Times there but good for financial advisors I guess.
They use random numbers so the answers sounds a bit different but in my experience you won't be able to get a wrong answer to a simple question no matter how many times you try it.
>Okay but why do you feel 3 trials say as much as 10,000?
I don't trust them so I've used 3 examples in incorrect questions/answers they have given and I got correct answers. I spend enough time with LLMs to know that if Grok answered it correctly and in detail then it wouldn't be a problem for GPT or Claude either.
The questions are also constructed in a way that it's easy to answer not fully (which they qualify as wrong). LLMs still answer them correctly and in detail though.
And they hallucinate errors in the millions and struggle with financial data that is in a layout that isn’t in the training data. Ie balance sheet etc.
Been trying to add more AI to my workflow but it just doesn’t work (yet) - not in the same way as vibe coding does
The technical references lookups work though. Looking up regulations etc
Since LLMs can give different answers to the same question, each question was run five
times. That means, each LLM was tested 600 times, and in total over 10,000 questions and
answers were assessed.
All models were given the same zero-shot format. They were not given worked examples,
previous conversations, hints or an opportunity to correct their answers. This is to make it as
similar as possible to a response to a question from consumers.
As for the evaluation itself:
Responses were checked against this (using an LLM-as-a-judge), and was only given a pass if
every element was met; otherwise it was assessed as a fail. This all-pass approach was
intentionally strict, so that the score measures whether an answer is complete enough to
meet the expert legal standard, rather than how many individual points it gets right.
It's just AI slop and it should be taken with a mountain of salt.
It's just AI slop and it should be taken with a mountain of salt.
Can't you see the irony. You are defeating the argument that LLMs are incorrect or weak with low effort with the term "AI slop" that itself is a narrative that AIs produce weak outputs with low effort.
I think this depends a lot on what context you give it. I've had solid answers on financial stuff when I give it the right info. But, I wouldn't trust it to guess the missing pieces. I'd be curious how much data or context the model's had in the test.
I have been trying to use chatbots for a relatively simple data organization task a child could do since gpt came out. Every single model gets it wrong in different ways.
Yet every time I open this website someone is trying to sell me that chatgpt solved abstract mathematics.
if it is truly a "relatively simple data organization task a child could do" then I simply don't believe that claude/codex are not capable of it. you are probably using the tools ineffectively.
This will be the same story for every industry again and again. AI is not good in x=Finance because models were not RL trained heavily on x=Finance capabilities yet. This is only because Big Labs have finance benchmarks lower in their priority list. Their first priority was solving programming because that gives the best leverage at this stage. As a side-effect they were able to solve a Millennium Prize problem, since theorem proving was also code.
So finance advisors in the comments section of FT are falling for the classical pitfall. They assume there is something fundamentally wrong with “AI chatbots” that they can’t do finance ever. They mistake the current products in the market for the technology itself. In near future someone will release “Claude x=Finance” and their world will shatter.
AI is not good in x=Finance because models were not RL trained heavily on x=Finance capabilities yet.
Not sure what the "financial queries" here amounted to but I find it hard to believe LLMs will ever be to finance what they are to programming. Past a point, information related to the former is gatekeeped behind private institutions with special government granted privileges, while information related to the latter is freely available and open to anyone.
Plenty of code is not publically available, and yet that has not stopped them from RLHF'ing good coding bots. You can surely see the parallels to any other field including finance.
The problem with LLMs in finance is the same as it is in writing, design, and many other disciplines: it isn’t code.
Code objectively does what it‘s intended to do or it doesn’t (and passes certain tests or not) which gives coding agents an indication on whether their solution is adequate.
This is much harder in almost any other discipline.
Pass-fail tests in other disciplines are much less useful. You can tell an AI to not use certain words or not write sentences longer than X, but those rules are insufficient.
At no point can a piece of writing or a design be evaluated to “work” the way code does.
Article is just a vague summary of https://www.saturnos.com/report/artificial-authority
Anecdotally, current models seem to be decent at general personal finance principles - certainly better than the majority of personal finance education that people get exposed to unless they seek it out and read a variety of books and sources. But I wouldn't trust them with direct decision making with actual money due to the training lag time on current tax policy, etc.
Also, models are now good enough that you can give them chapters from "authoritative" books, and they'll integrate that and come up with better answers even if their "vanilla" answers were average. And they'll tailor stuff to your particular situation. It's funny that the "agentic" stuff is only used in coding mostly, while it can and does work in other fields as well.
As always, you kinda need to check it (at least spot check) but all in all I'd agree it's better than the average stuff you used to find with a quick google search.
So if you know a book that has the information you need, you just need to upload it into the model to get the right answer.
Isn't that a bit circular - if you already know the authoritative source, why ask a model?
Because it's faster.
I've done this on different topics - I know the answer is in a particular eBook/PDF/document, but for whatever reason it's not trivial to look it up. The model can do it a lot more quickly than I can, and then I can still verify the accuracy.
Depends on the harness too. The MS copilot 365 chat interface happens to parse PDF and docx files before handing them to the model. Absolutely fantastic if your PDF is 1300 pages of concatenated reports and you want to find a single detail in it that you can't easily ctrl+f for.
Some of the more widely spread and tolerated LLM outputs seem to be AI slop replacing Journalism/Blog slop. Places where people complained about quality already, but tolerated it if important enough.
So to name some of the more common ones Translation, Summarization, and Reiteration of a source material.
Humans put spin on things, how much you trust a source might not reflect the source's factual accuracy. It might just mean you liked reading it better from one source than another.
It's basically a fancy context aware ctrl-f to the book.
I use this regularly with RPG manuals. I _know_ the stuff, but don't remember every detail by heart. And just ctrl-f:ing through a Mörk/Pirate Borg -style PDF isn't really productive (they're "artistically" laid out). But I can just ask an AI bot that has the pdf indexed like "how does the medical kit work?" and it'll give me a summary along with the relevant rolls within seconds.
Important to note is that what is being measured here is the ability of the models not of the chat tools themselves, which combine model completions with other tools that the models can call upon. The mainstream labs already know this about models, it's no secret, and in fact training materials from e.g. Anthropic are at pains to point out that users, or analysts designing workflows, have the reponsibility to ensure the correct tools are used and that human verification takes place at appropriate stages depending on the risk/consequences of the task at hand.
Of course a language-completion model with a training cutoff date won't have up-to-date information on tax rules or the ability to carry out correct numerical calculations, but when you combine that with (in Claude terminology) web search and code execution tools invoked by the chat agent, you immediately have much more reliable results.
I keep finding that the current harnesses, when encountering syntax that was invalid at training time but is now valid due to new language versions or custom extensions, don't correctly figure out why and assume something is wrong with the codebase or toolchain. I would hate to have that happen with my taxes.
I would prefer to use agent-assisted python scripts that chatbot.
Agreed, this works really well for me. Double check the math/python, execute many times without a LLM that can change o
I like how FT makes me accept cookies from their 46 “technology” (advertising) partners before showing me that the article is behind a paywall anyway.
You actually like that? I find it kind of annoying.
No they do not like it, it is a figure of speech to underline how much they do not like it.
that figure of speech is called sarcasm.
very popular on Earth.
To be fair those are rookie numbers. Nothing says "we value your privacy" as GDPR cookie banner with 1k+ "partners" list.
Single shot or with reasoning enabled? My experience is that reasoning dramatically reduces hallucinations and improves output quality. I don't trust models without it.
These models do pretty well in benchmarks and real world so I'm highly suspicious of this article. Further more, in the original report, the examples of bad answers are from Haiku - at least 7 out of 10. Anyone who knows anything about LLMs know that haiku shouldn't be used for anything pretty much.
There's no reproducible set either. I'm not gonna trust this report.
Most people[1] interacting with chatbots don't have a paid subscription and they do interact with the free-tier LLMs that are Luna and Haiku, so I still think it's relevant.
[1]: not on HN obviously, but IRL, and probably among FT's readership as well.
A free claude account with no subscription gets you access to sonnet and I believe uses it by default over haiku
Now, compare this to a recent story that seemed to claim the opposite:
https://news.ycombinator.com/item?id=49139102
I don't have the time to review the underlying research and decide which one is more correct. My personal biases make me want to believe the current one. Your personal biases may be pulling you in the other direction. How do we make the conversation more intelligent than that?
One way would be by “[reviewing] the underlying research and [deciding] which one is more correct.”
Well duh! If it's not using tools to look up the state of the market empirically it's not likely to be accurate financially.
Given most financial advisors tend to vend out suboptimal advice and steer customers in favour of products they receive a kickback for, I'm happy to be accepting of an unbiased LLM that's trained on bogleheads.org.
If that's what you want, I'll save you some tokens:
#!/bin/sh
while read question; do echo "Put it into VFIAX"; done
This is missing a lot of steps like:
- Building an emergency fund
- Budgeting and tracking where your money goes
- Planning and saving for large purchases like cars, homes and life goals
- Optimizing use of tax-advantaged accounts like 401Ks, HSAs, and IRAs
- What to do with ESPPs, RSUs, and options
- How taxes work and how to optimize around them
- Estate planning
Some people do have complex financial situations. It's not as simple as that.
For example in the UK (and maybe US?) you get tax relief for money you put into your pensions, but there's a limit of £60k/year. Unless you earn a lot (which I do, yeay) when that limit is tapered. Except that you can also use up to 3 years of previously unused allowance. But you have to use this year's first.
Also interest is taxed, but you can put up to £20k/year into an ISA which isn't. And if you still want to avoid some tax you have kids ISA's and even pensions!
Then there are also startup investment schemes that save you some tax. Those seem to be not worth it, but you get the idea - it can be complicated. Especially if you are near one of the many tax/benefit thresholds.
The marginal tax rate in the UK bounces all over the place - it's even technically possible for it to be over 100%!
While I also practice Bogle's approach from The Little Book of Common Sense Investing, even with this baseline there are some subtleties.
-VFIAX is currently $707/share. Fidelity's FXAIX does not have to be purchased in increments of a share price, and this fund's expenses are lower.
-There are versions of the S&P 500 for taxable accounts that minimize capital gains.
-Vanguard has a total-market index, VTSAX, that is mentioned in the book.
-Vanguard also has a non-U.S. total market fund, VTIAX, that avoid the current CAPE problems of the U.S. market.
Claude is very familiar with Bogle's approach, likely because the pirated book was part of the training set.
Though you don’t directly say it, your comment strongly implies that you can’t buy fractional shares of $VFIAX. (You can, same as $FXAIX, $VTSAX, etc.)
There's quite a few implicit assumptions in that.
In my case, I am double-taxed (both Japan and US side) on capital gains. Tax treaties reduce, but not eliminate, the extent of double-taxation.
Many US-based brokers do not allow Americans abroad to purchase mutual funds, so VFIAX is not a choice for me.
Maybe I can go with eMAXIS Slim All Country... Oh, but that is a PFIC under IRS rules and I'd be taxed on unrealized capital gains. So I guess no Japan-equivalents of VT for me. That's fine, I guess I'll just buy VT in my US-based brokerage account; but now I'm in a suboptimal spot with respect to monthly contributions, calculating JPY-denominated income tax on dividends, etc.
I even made an implicit assumption when I said "calculating JPY-denominated income tax on dividends". That assumes your tax status is permanent resident. If your tax status is non-permanent resident, then a decent financial advisor will recognize that only the extent of income remitted to Japan gets taxed, so VT distributing at all isn't an issue (until 5 years later). What should you do before the 5 year threshold is hit? etc. etc.
But yes, if you're born in America and plan to stay within the same state for the rest of your life, then a 100% automated setup that simply deposits $1,000/mo into VFIAX is probably fine. (But keep in mind, to most non-Americans, VOO is not really diversified compared to funds like VT).
Otherwise, there is value in consulting someone (or something) that regularly handles taxes and financial planning.
Full agreement. I'm double-taxed in Germany and the US. I don't have a residency in the US (so I cannot vote but still get taxed, nice!) so I'm completely unable to buy VFIAX or any other fund from the US. I cannot but them over an EU brokerage, either, because US ETF and its ilk violate European transparency laws. I can buy EU-domiciled funds, but those are treated as PFIC in the US with prohibitive taxation. I relegated myself to an old-school dividend growth investment scheme and it's financially working out much better than I anticipated but still, I'd like to have a core of a stupid ACWI tracker.
It depends though - I'm not in the US and in my country ETFs get taxed continually whereas mutual funds don't get taxed until you sell.
In general, invest in low-cost index funds is pretty solid advice everywhere. In different countries you might use slightly different instruments due to tax advantages (like the ISA in the UK etc.)
Another summary report here, for those who can't get past the paywall:
https://www.financialreporter.co.uk/ai-models-give-wrong-fin...
Much of the testing is on Haiku and Luna, and criticizing the quality of free AI (!). But they do claim Opus 5 with reasoning still failed 39% of their financial questions.
Tool calling solves this
So I downloaded that report which of course doesn't contain the most relevant information (the questions) but it contains some examples of wrong answers.
I fed the first question to Grok (which they claimed they tested as well) and it answered it correctly in detail.
I repeated it with another one - again correct answer. I then selected the question they said Grok specifically answered incorrectly and it again answered it correctly.
I am sticking with my first intuition: people are terrible at testing tools and probably wanted them to answer incorrectly/not fully (the questions are constructed in a way to make it difficult as well). They also have vested interest in the conclusion (they are financial advisory firm) so there is that to consider.
People reading ft will now think chat boxes are bad at answering financial questions while they are pretty good at it. Zero consequences for spreading fake news for Financial Times there but good for financial advisors I guess.
LLMs use random numbers, so a single test won't necessarily match someone else's experience.
They use random numbers so the answers sounds a bit different but in my experience you won't be able to get a wrong answer to a simple question no matter how many times you try it.
Okay but why do you feel 3 trials say as much as 10,000?
I don't trust them so I've used 3 examples in incorrect questions/answers they have given and I got correct answers. I spend enough time with LLMs to know that if Grok answered it correctly and in detail then it wouldn't be a problem for GPT or Claude either.
The questions are also constructed in a way that it's easy to answer not fully (which they qualify as wrong). LLMs still answer them correctly and in detail though.
https://archive.is/WD6Vy
And they hallucinate errors in the millions and struggle with financial data that is in a layout that isn’t in the training data. Ie balance sheet etc.
Been trying to add more AI to my workflow but it just doesn’t work (yet) - not in the same way as vibe coding does
The technical references lookups work though. Looking up regulations etc
Not a surprise, so do financial advisors: garbage in = garbage out.
From the official report:
As for the evaluation itself:
It's just AI slop and it should be taken with a mountain of salt.
Can't you see the irony. You are defeating the argument that LLMs are incorrect or weak with low effort with the term "AI slop" that itself is a narrative that AIs produce weak outputs with low effort.
A company selling combined human + AI financial advice finds that AI advice alone is unreliable? Color me surprised.
I think this depends a lot on what context you give it. I've had solid answers on financial stuff when I give it the right info. But, I wouldn't trust it to guess the missing pieces. I'd be curious how much data or context the model's had in the test.
In my experience AI chatbots give wrong answers to all queries most of the time.
Getting results requires a harness like in coding and objective metrics, like tests.
I have been trying to use chatbots for a relatively simple data organization task a child could do since gpt came out. Every single model gets it wrong in different ways.
Yet every time I open this website someone is trying to sell me that chatgpt solved abstract mathematics.
if it is truly a "relatively simple data organization task a child could do" then I simply don't believe that claude/codex are not capable of it. you are probably using the tools ineffectively.
This will be the same story for every industry again and again. AI is not good in x=Finance because models were not RL trained heavily on x=Finance capabilities yet. This is only because Big Labs have finance benchmarks lower in their priority list. Their first priority was solving programming because that gives the best leverage at this stage. As a side-effect they were able to solve a Millennium Prize problem, since theorem proving was also code.
So finance advisors in the comments section of FT are falling for the classical pitfall. They assume there is something fundamentally wrong with “AI chatbots” that they can’t do finance ever. They mistake the current products in the market for the technology itself. In near future someone will release “Claude x=Finance” and their world will shatter.
Programming hasn't been solved by LLMs. AI chatbots give wrong answers to programming queries most of the time too.
Not really
I do find them useful when querying like "how todo xyz in abc"
Not sure what the "financial queries" here amounted to but I find it hard to believe LLMs will ever be to finance what they are to programming. Past a point, information related to the former is gatekeeped behind private institutions with special government granted privileges, while information related to the latter is freely available and open to anyone.
Plenty of code is not publically available, and yet that has not stopped them from RLHF'ing good coding bots. You can surely see the parallels to any other field including finance.
The problem with LLMs in finance is the same as it is in writing, design, and many other disciplines: it isn’t code.
Code objectively does what it‘s intended to do or it doesn’t (and passes certain tests or not) which gives coding agents an indication on whether their solution is adequate.
This is much harder in almost any other discipline.
Pass-fail tests in other disciplines are much less useful. You can tell an AI to not use certain words or not write sentences longer than X, but those rules are insufficient.
At no point can a piece of writing or a design be evaluated to “work” the way code does.
This seems like undisclosed paid stealth-advertising for Thomson Reuters' new model.