Easy questions that AI messes up

I get frustrated with the like of ChatGPT and Gemini. Find a problem I think they will handle easily and they just can't. Not without coaching. Whenever there is an exact answer they crap the bed.

Question: What months have three pay periods in 2024 and 2025 if I get paid every two weeks and got paid on 10/4/2024.

ChatGPT Summary answer

• 2024: November has three pay periods.
• 2025: February, May, August, and October have three pay periods.

Google Gemini answer


---
Input: That is not correct. I think it is May and October in 2025

ChatGPT


Gemini still butchered it with added commentary about October having three pay periods.




I didn't have the answer when I started but after both output wrong answers I figured it out manually with calendar. Tried to coach them into the correct answer.

windows copilot tell me:
Alright, let's break it down.

Given that your pay date is biweekly (every 14 days) and you got paid on 10/4/2024, here are the months with three pay periods:

For 2024:

• March: Pay dates are 1st, 15th, and 29th.
• August: Pay dates are 2nd, 16th, and 30th.
• November: Pay dates are 1st, 15th, and 29th.

For 2025:

• May: Pay dates are 2nd, 16th, and 30th.
• October: Pay dates are 3rd, 17th, and 31st.

That something that the 01-preview with multi-step "reasoning" could help

LukeTbk said:

windows copilot tell me:
Alright, let's break it down.

Given that your pay date is biweekly (every 14 days) and you got paid on 10/4/2024, here are the months with three pay periods:

For 2024:

• March: Pay dates are 1st, 15th, and 29th.
• August: Pay dates are 2nd, 16th, and 30th.
• November: Pay dates are 1st, 15th, and 29th.

For 2025:

• May: Pay dates are 2nd, 16th, and 30th.
• October: Pay dates are 3rd, 17th, and 31st.

That something that the 01-preview with multi-step "reasoning" could help

Click to expand...

Looks like Copilot for it wrong too. August wasn't a three payday month based on an October 4th payday. August 9th and the 23rd were paydays.

I tried to get ChatGPT to check me and it failed.

we can see how it making the mistake that way (for example it count 9/6 twice, one time in august)

For a quick reference:
https://www.timeanddate.com/date/da...1=2024&type=sub&ay=&am=&aw=2&ad=&re=on&rec=25
https://www.timeanddate.com/date/da...1=2024&type=add&ay=&am=&aw=2&ad=&re=on&rec=25

would be curious to see how the 01 preview would do, if you trying to help it does not help maybe reflecting itself would not either.

Super duper AI can't tell you the current year. You want it to sort out future pay periods? Good luck.

I can't wait for people to turn over their finances to AI tools.

AI doesn't think....at the end of the day, it is an algorithm that minimizes a loss function on the training data. Until we solve the theory of the mind, black box ai will have limited uses where we cannot expect it to have any "understanding".

1.1.2.3.5... said:

AI doesn't think....at the end of the day, it is an algorithm that minimizes a loss function on the training data. Until we solve the theory of the mind, black box ai will have limited uses where we cannot expect it to have any "understanding".

Click to expand...

While the first part of your statement is true, massive overparameterization in optimization appears to avoid over-fitting the training data (double-descent) and generalization is apparent; furthermore the file-cabinet metaphor for information recall/prediction is easily disproved. There is no rigorous mathematical proof for generalization, for sure, but there are a bunch of theories in active research. Because LLMs are using linguistic patterning, we don't actually have to solve the theory of how a brain works, just recognize that linguistics (as represented in training data) has order.

The idea that factory robots (and outside factory robots) would need some theory of the mind real thinking or that being a black box would be an issue..... The AI that look at a bunch of radiography and other data does not need to think to give a better that any humans percentage of chance there is a cancer or not.

Computer do not think, not sure we would say they had limited uses.

Machupo said:

Because LLMs are using linguistic patterning, we don't actually have to solve the theory of how a brain works, just recognize that linguistics (as represented in training data) has order.

Click to expand...

That makes zero sense. By that logic you could train an LLM to find the next biggest prime and that can probably be proved to be impossible.

People underestimate how much can be done with not that better LLM that what we have already and not having real thought.

How much of what robotics will be useful for will not work with just LLM getting 3% better a year the next 10 ? Level 1-2 customer service ? Filling function in people code ? Just the best of the best right now, when it will be 100 cheapers, half the latency, 10 time the speed and 100 time the context windows would already if you let 10 years for people to build product and change industry for it be quite something, it does not need to be even better than what we already have to be a big deal I suspect, just much cheaper-faster, which we can be certain for it to happen.

1.1.2.3.5... said:

That makes zero sense. By that logic you could train an LLM to find the next biggest prime and that can probably be proved to be impossible.

Click to expand...

Not sure to get that one, isn't prime number particularly useful in cryptography because they do not follow order (or predictability), which make their raw calculation needed and heavy ?

Do you really think if we do not solve fully how the brain work and that the intelligence in the system stay artificial instead of real, it cannot end up a bit everywhere ? An agent can go ahead and jump through tall the loops to cancel your monthly memberships you did not like without a perfect theory of mind, they can be better a diagnostic than doctor without it, drives and make cars without it.

Planes never needed to have bird like propulsion to flight and impact transportation, chess engine never had to really think to beat humans at it.

1.1.2.3.5... said:

That makes zero sense. By that logic you could train an LLM to find the next biggest prime and that can probably be proved to be impossible.

Click to expand...

That is not comparable. LLMs learn next token prediction by training data frequencies and frequency patterns stored in a huge hyperparameter space (e.g. weights) which is a subset of total word space as many word series are not present in training sets; what you are talking about is a direct algorithmic search in unpruned space (and full enumeration and test up to the square root). One of the key features of primes is that they are computationally inefficient to calculate based upon previous knowledge of primes, there are no workarounds (edited to add: yet).

I was browsing 4ch's cooking section and a bot posted an obvious AI recipe for fettuccini alfredo.

With the right ingredients you need water, salt, fettuccini, butter, and parmesan cheese to make fettuccini alfredo. This recipe was a more American style alfredo for two portions that called for cream and a few other ingredients. One of the ingredients was two cups of canola oil. For two portions of pasta.

Then everyone started arguing about seed oils.

Never change, 4ch...

Ben_Smith said:

hello its cool post

Click to expand...

Another bot sneaking into [H]...

Just ran into this thread, tried doing it with Gemini for 2026 and a biweekly pay period with last pay date 5/29, and it says May and October. Looks like Google's gotten smarter about this, even includes it's reasoning, and as a minor bonus, ran the calculation for the entire year, not just the future:




Then it offered to help me set a budget, lol. I wonder if I told it how much I make it would suggest investing in $GOOG