Math is easily GPT4's biggest challenge and weakness right now, but notably many of its issues is due to the memory limit, and how memory gets generated through tokens. And since it has no way to validate its own numbers outside of itself, and since there may not be a reference point for it there is no way for it to know it is wrong.
From time to time, at seemingly random moments, chatgpt+ will start hallucinating and suddenly lose formatting I provided to it, or just outright forget a part, to combat this I have it number each section by a number or word, sometimes I have to feed it back the context. I would say this starts occurring around the 20-30 message mark give or take.
So no I would not rely on it on complex calculations or formulating spreadsheets for this reason until it improves.
https://www.psychologytoday.com/ca/b...-do-arithmetic
If you ask it to review its numbers and analyse if it did the math correctly, it will usually find and correct its mistake, but until it gets browsing access I would not suggest to professionally rely on it for arithmetic. Once it has access to browsing and an online calculator, its results should be far more accurate.
Math accuracy is also being worked on significantly for GPT5 which is supposed to be released end of year.
This explains well the issue with token limits and side effects. chatgpt+ GPT4 is limited to 8K tokens, however the full version of GPT4 has 32k token, a dramatic improvement which we should hopefully see soon more readily available. Periods for example are their own token, if you are feeding a spreadsheet with decimals, tokens get eaten up extremely fast and may even be lost within a couple messages.
https://medium.com/@russkohn/masteri...y-ce920630349a