Another symptom of the annoying tendency to mistify what AI actually does.I don't think this argument works:
Why AI gets stuck in infinite loops — but conscious minds don’t
Anil Seth suggests the difference is that living beings are rooted in time and entropy, a grounding that may be essential for consciousness.bigthink.com
Surely just tiny bit of code could detect looping and a bit more fix the problem?
Merged Artificial Intelligence
- Thread starter Puppycow
- Start date
- Featured
A little while later, I found an article by a real mathematician that filled in a bit of the details.
Then I found a YouTube video that filled in the major details. The AI doesn't do the heavy lifting as more than fifty percent of the job gets done by regular computer programs.
Dr.Sid
Philosopher
This is much more interesting:
Claim: gpt-5-pro can prove new interesting mathematics. Proof: I took a convex optimization paper with a clean open problem in it and asked gpt-5-pro to work on it. It proved a better bound than what is in the paper, and I checked the proof it's correct.Details below.
Claim: gpt-5-pro can prove new interesting mathematics. Proof: I took a convex optimization paper with a clean open problem in it and asked gpt-5-pro to work on it. It proved a better bound than what is in the paper, and I checked the proof it's correct.Details below.
I think it's more of an attempt to mystify and claim humans are special.
Yeah, but, but, but it's just a parrot!
What's the issue with that?
Has someone proven that it isn't just a parrot?
Wudang
BOFH
"It's not stuck in a loop, it's just resting."
It's in the post I quoted: Claim: gpt-5-pro can prove new interesting mathematics. Proof: I took a convex optimization paper with a clean open problem in it and asked gpt-5-pro to work on it. It proved a better bound than what is in the paper, and I checked the proof it's correct. Details below.
It can't be parroting that because that is something that 1) couldn't have been in its training data 2) couldn't have searched the web and found it elsewhere.
I think this article offers a very balanced and well-reasoned viewpoint:
Did GPT-5 Really Solve ‘New Math’? - Tech2Geek
Artificial intelligence is no longer just summarizing articles or generating code snippets — it’s stepping into the world of advanced mathematics. Recently, a claim surfaced...
www.tech2geek.net
Dr.Sid
Philosopher
It's also that Chat GPT is not designed as specialized math AI. It's an LLM. It knows math not because it was designed to know math, but because it has learned math. So far math used in AI typically mean special AI designed to some specific math approach, and could not do anything else.
Is it just me or are people thinking that the first AGIs will be geniuses in all fields etc?
I would expect the first to be rather like an average person, so for instance there is no way I could ever achieve "a gold medal-level score on the 2025 International Mathematical Olympiad", work out the structure of proteins, find a "better bound than what is in the paper". Same with other areas - there is no way I can be as good an artist as the latest AI art generators. I worry if researchers will get lost and think an AIG is a failure because it can't outthink, out-reason human geniuses in every field.
I would expect the first to be rather like an average person, so for instance there is no way I could ever achieve "a gold medal-level score on the 2025 International Mathematical Olympiad", work out the structure of proteins, find a "better bound than what is in the paper". Same with other areas - there is no way I can be as good an artist as the latest AI art generators. I worry if researchers will get lost and think an AIG is a failure because it can't outthink, out-reason human geniuses in every field.
Wudang
BOFH
I think it's more of a reaction to the hype. As usual a certain element wets it collective panties with excitement, sales people say it's a very stable genius and so on. To me it's still just a very interesting and potentially promising technology that should be watched.
I'm going to post something I posted in the "Cont: Trivia Quiz part 16" thread. There was a question asked: "In the 1700s, blood transfusion was used to treat psychosis. Oddly, sometimes it worked. Why", the answer given was:
My research:
I'm posting what I did above as it was the same as Copilot (Using GPT-5) did when I went to research the claim made about treating syphilis, albeit it did more research and checked more websites than I did. I started to look at websites, followed links, reframed my questions, separated the syphilis claim from the blood transfusion history etc.
In what way was its research different to what I did? In what way was it not reasoning the same way I was?
"It is syphilis.
A particularly common form of psychosis occurred in the final stage of infection with syphilis. Syphilis is an uncommon bacteria in that it is extremely sensitive to temperature - so sensitive that a high fever greatly reduced or even eliminated infection. In fact, once quinine became available to treat malaria, which caused repeated high fevers, intentional infection with malaria was used to cure syphilis, which would then be treated with quinine. But in the 1700s, Europeans didn't have access to quinine yet."
This struck me as being wrong as I had thought that blood transfusions were banned during the 1700s. So I went and did some "research" - and this will get back to AI again I promise:My research:
I'm posting what I did above as it was the same as Copilot (Using GPT-5) did when I went to research the claim made about treating syphilis, albeit it did more research and checked more websites than I did. I started to look at websites, followed links, reframed my questions, separated the syphilis claim from the blood transfusion history etc.
In what way was its research different to what I did? In what way was it not reasoning the same way I was?
Why couldn't it have been in its training data?It's in the post I quoted: Claim: gpt-5-pro can prove new interesting mathematics. Proof: I took a convex optimization paper with a clean open problem in it and asked gpt-5-pro to work on it. It proved a better bound than what is in the paper, and I checked the proof it's correct. Details below.
It can't be parroting that because that is something that 1) couldn't have been in its training data 2) couldn't have searched the web and found it elsewhere.
I think this article offers a very balanced and well-reasoned viewpoint:
Did GPT-5 Really Solve ‘New Math’? - Tech2Geek
Artificial intelligence is no longer just summarizing articles or generating code snippets — it’s stepping into the world of advanced mathematics. Recently, a claim surfaced...
Dr.Sid
Philosopher
Researchers will quietly push the boundaries no matter what. They talk about benchamrks, or worse, the underlying math. They don't use fancy names. That's marketing stuff.
AGI can mean lots of things, but it is still an improvement over term AI, which can basically mean anything, especially if you are in the field for some time. First the field itself is called AI. The old meaning I learned in school 30 years back was a set of problems. Note, not how to solve them. OCR and image processing were traditional AI problems, even when they did not contain neural networks or learning. Then there are computer games, where AI is code which controls NPC, even if it is just going left and right in a loop. Obviously anything neural network related and machine learning is AI, not mentioning LLMs.
So I understand new terms are needed. But I rather had lot more of them with lot more specific meaning.
AGI can mean lots of things, but it is still an improvement over term AI, which can basically mean anything, especially if you are in the field for some time. First the field itself is called AI. The old meaning I learned in school 30 years back was a set of problems. Note, not how to solve them. OCR and image processing were traditional AI problems, even when they did not contain neural networks or learning. Then there are computer games, where AI is code which controls NPC, even if it is just going left and right in a loop. Obviously anything neural network related and machine learning is AI, not mentioning LLMs.
So I understand new terms are needed. But I rather had lot more of them with lot more specific meaning.
Dr.Sid
Philosopher
It's new solution. It's solution to a problem the researcher described, and so far nobody was able to find a solution. It's been checked by the researcher who described the problem, and found not only to be correct, but also to be original.
Because it was a new problem - created cleanly.
That's not proof, that's a baseless assumption that there wasn't some sort of solution in the indiscriminate piles of training data.
And just when you thougth they started to show some sense:
Trying the new Gemini flash model:
I asked it where this photo was - it didn't get it but it made a stab at it:
"
It's actually a photo of British troops in Burma hearing about the Japanese surrender.
Then I thought - everyone is raving about this new "nano-Banana " image editor - which is in fact Google's new model, so I asked it to colourise the image, be prepared:
WTF!
Trying the new Gemini flash model:
I asked it where this photo was - it didn't get it but it made a stab at it:
"
- Query successful
It's actually a photo of British troops in Burma hearing about the Japanese surrender.
Then I thought - everyone is raving about this new "nano-Banana " image editor - which is in fact Google's new model, so I asked it to colourise the image, be prepared:
WTF!
Last edited: