Given that millions and millions of people will run AI queries, it seems fair to test it repeatedly to see if it can give a proper answer each time.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Study Reveals Critical Flaws in AI's Logical Reasoning Abilities
- Thread starter MacRumors
- Start date
- Sort by reaction score
4th paragraph of the article says they used o1.
Perhaps this is mostly a hint as to the direction they want to go in trying to get insightful and reliable responses in Siri etc - neurosymbolic AI
I came to the “AI” game a little late but have been using it the past month in my software development. It can be absolutely fantastic for simple things, but it often just flat out fails on complex questions. And, yes, the AI can give dramatically different answers with a slight change in wording. As an “assistant”, it’s pretty great most of the time, but I wouldn’t rely on an AI for anything critical.
I agree its consciousness level seems low, but these kinds of things are also things that kids get wrong too. When my kid was 4, she was super into math problems, and she’d ask me to give her problems like this all the time for fun. She’d make these exact kinds of mistakes. I’d add irrelevant information to see if she’d inadvertently incorporate it - for example, I’d ask, I went to the store and bought three apples, two sausages, and two bananas, how many fruits do I have? She'd often get questions like this wrong. So I’m not sure this necessarily proves anything other than it’s reasoning isn’t incredibly advanced.
I wonder how many here have actually taken the time to read, absorb, grasp, and understand the study?
From some of the responses, I'm guessing one or two. Maybe.
From some of the responses, I'm guessing one or two. Maybe.
I suspect these problems will become less and less relevant as the models improve. Also nobody really understand what goes on in our brains, so it could well be we are also very good pattern finders and that’s all our ‘reasoning’ is.
Yeah, this is exactly how significance testing works. I haven't read the paper, but it is assumed that they would have tested a large number of queries, a large number of times, since that's how you approximate to the true result (law of large numbers). A published, peer reviewed study should also find statistical significance, meaning that it is extremely unlikely that chance alone could explain the result. I'd imagine they also tested different contexts surrounding the question itself (though maybe not), since a model that gets is right in isolation, but fails when there is additional—tangentially related, or unrelated—content in the context window, is clearly not using reason (which would remain consistent regardless of long-term context).
One of a zillions reasons most of us see AI as "the next crypto"
Just being pushed on everyone as SV and the VCs are incredibly desperate for a "next big thing"
Just being pushed on everyone as SV and the VCs are incredibly desperate for a "next big thing"
Does this question exist elsewhere on the Internet? Yes it does. The conclusion is indexed.
LLMs absolutely cannot count. They can't do basic math because that's not what the algorithm is built for. They can do excellent pattern matching.
Can't wait to see people bring this as an example of AI lacking utility. They're Large Language Models for a reason.
This is more of an expectations issue than a tech issue.
This is more of an expectations issue than a tech issue.
LLMs have no reasoning ability. They're literally just next word guessing algorithms fed on the stolen works of others, and worse, public forum posts. There's no intelligence at work, at all.
I tried ChatGPT with the following question "why is 9 time 6 equals 42 a problem" and got the following result
"Saying that 9 times 6 equals 42 is a problem because it’s incorrect. The correct product of 9 and 6 is 54. Miscalculating can lead to errors in various contexts, such as math problems, financial calculations, or measurements. Accuracy is key in mathematics!:
When it says it's incorrect, that's not necessarily true, same for the answer being 54, it's not necessarily true. It's correct when it says it accuracy is key but so is completeness.
It made a fundamental assumption that I had fully specified the problem and was working in decimal, which was why it unequivocally said it was incorrect.
I wasn't I was using base 13 (where 13 is in decimal notation)
Decimal version: 9 (Decimal) X 6 (Decimal) = 54 (Decimal) , i.e. 5 X base + 4 -> 5X10 +4
Base 13 version: 9 (Base 13) X 6 (Base 13) = 42 (Base 13), i.e. 4 X base + 2 -> 4 X 13 +2
That the ultimate question is not expressed in decimal is why the universe is a strange and wondrous place.
"Saying that 9 times 6 equals 42 is a problem because it’s incorrect. The correct product of 9 and 6 is 54. Miscalculating can lead to errors in various contexts, such as math problems, financial calculations, or measurements. Accuracy is key in mathematics!:
When it says it's incorrect, that's not necessarily true, same for the answer being 54, it's not necessarily true. It's correct when it says it accuracy is key but so is completeness.
It made a fundamental assumption that I had fully specified the problem and was working in decimal, which was why it unequivocally said it was incorrect.
I wasn't I was using base 13 (where 13 is in decimal notation)
Decimal version: 9 (Decimal) X 6 (Decimal) = 54 (Decimal) , i.e. 5 X base + 4 -> 5X10 +4
Base 13 version: 9 (Base 13) X 6 (Base 13) = 42 (Base 13), i.e. 4 X base + 2 -> 4 X 13 +2
That the ultimate question is not expressed in decimal is why the universe is a strange and wondrous place.
AI has mainly been the always shady tech industry trying to cash in on a new big thing. It is not like it is actually hard for large software makers to convince media and politicians that something is huge and deserves massive investment. Add in the threat that other countries (China, etc...) may beat the US to the new technology and you have the makings of a true bonanza.
It is the more modern version of early (and sometimes) modern pages on Wikipedia where people "source" something that has zero factual basis.
If the source of the data is garbage, the answer to the equation will be equally useless.
That's fine and all, but sounds like sour grapes coming from Apple, who still has NOTHING. In what they offer users, they are years behind even freaking Grammarly.
It's funny to me that we act like if AI isn't "sentient" then that means it's complete BS. Why not just recognize what it is good at and stop with the unnecessary hype?
So Apple is proving Apple Intelligence is crap before it even releases? Funny.
But in all seriousness, I've had a lot of luck with the ChatGPT o1-preview model in particular. It is definitely a step up. What I like is that it breaks things down into parts and then goes through each part to show how it arrived at that conclusion. Then it's pretty easy for me to follow since it breaks down something complicated into something simple and I can check the work for each step and ensure it's not doing something weird.
It also nailed a fairly complicated bit of programming for me the other day in an emergency situation and knew all of the proper API calls to make to an external service we use. Allowed me to react several times faster to the issue by rapidly building out a tool to help resolve an issue across 800+ websites. Checked all the code, modified a couple bits to better suit the need, tested it on a few sites, and then mass deployed. Everything went great, but it's still a tool. It's good to understand the limitations of your tools, and some tools are more effective in the hands of one person vs. another, based on experience and skill level. I also do woodworking, and it's even more apparent in that field. Never rely too heavily on your tools, or on any one tool.
But in all seriousness, I've had a lot of luck with the ChatGPT o1-preview model in particular. It is definitely a step up. What I like is that it breaks things down into parts and then goes through each part to show how it arrived at that conclusion. Then it's pretty easy for me to follow since it breaks down something complicated into something simple and I can check the work for each step and ensure it's not doing something weird.
It also nailed a fairly complicated bit of programming for me the other day in an emergency situation and knew all of the proper API calls to make to an external service we use. Allowed me to react several times faster to the issue by rapidly building out a tool to help resolve an issue across 800+ websites. Checked all the code, modified a couple bits to better suit the need, tested it on a few sites, and then mass deployed. Everything went great, but it's still a tool. It's good to understand the limitations of your tools, and some tools are more effective in the hands of one person vs. another, based on experience and skill level. I also do woodworking, and it's even more apparent in that field. Never rely too heavily on your tools, or on any one tool.
That’s the differentiator. Anyone truly following this industry knew that intuitively (or from experience). But Apple actually assimilated the actual data. So let’s not bash them too hard for this.
It seems you might be the one with the skewed perspective. Do you use it very often?
Because the money pushing it wants it be some HUGE ASTRONOMICAL PAYDAY
...not just the occasionally and contextually possibly useful small side feature it actually is
We live in a BS Hype world the likes of which have never before been seen
I hope a reversion to some modesty, honesty, hopeful vision and truth may return, but right now it just feels like everyone is trying to see what they can get away with in the name of "cashing out"
I mean even just look at Apple
Is anything they are doing really expressing a coherent and hopeful vision for the future and where we want to go as people and customers?
No -- it's just a huge rent extraction operation to juice the stock price
Their one "thing" of late (AVP) is an overpriced, underwhelming, hugely isolating and anti-social helmet that sales would indicate nobody even wants.
I don't know. Before I retired this year, if I could have got that level of consistency from the executive team, I would have been a happy camper.
The problem with showing that some other model gets the correct answer is the same as we've always had: how do we know that we're getting a good answer? (Of course this isn't limited to AI) Opinions are everywhere. Curation is not. When we believe an answer, it's because of its reasonableness, the confidence we have in the answerer, and the supporting data provided (and its provenance). Have you ever read the studies quoted at the ends of articles about a drug's or an air purifier's effectiveness? About 40% of the time, the quoted study states something irrelevant or even opposite.
(And why should you believe me? 😉)
(And why should you believe me? 😉)