Merged Artificial Intelligence

[The Great Zaganza]

Juat read Peter Thiel's latest rant on the anybody who has questions about AI is evil"

https://www.politico.com/news/2026/10/08/peter-thiel-ai-obama-pope-leo-01112122

It's official: Peter Theil has lost his marbles.

Guy is becoming like the Villian in a James Bond movie.

So what should be done with mad billionaires with unheard of political influence?

 

[EHocking]

I remember when Marconi threatened to put the post office out of business. Much the same thing today.

How old are you?!

 

[Andy_Ross]

deleted wrong thread

 

[Gawdzilla Sama]

How old are you?!

Well, I call Dirt "Junior".

Still planning to sue Harryhausen for using my image in "GWANGI"!

 

[W.D.Clinger]

Tools are generally a good thing, which is why we have them. But what if you can't make the tool or buy it yourself? What if it's owned by a trillion dollar company with an enormous computer that has its own gigawatt power station? What if they scraped the entire internet and millions of books to get the required data? What if everybody abandons other tools because this one is supposed to do everything? What if people think it's 'super-intelligent' and rely on it to tell them what to do? What if they have a 'personal' relationship with it that affects their interactions with other people? What if you are forced to use it too because it's everywhere and in everything?

People were rightly concerned about Microsoft having a monopoly on personal computer software, but that's nothing compared to how dependent AI companies want us to be on their 'tool'.

Many of those who are reading this use operating systems such as Microsoft WIndows or Apple MacOS. They did not create those tools, nor did they buy those tools. They are using them under license.

So licensing as a business model is already pervasive. It is possible, however, to use an open-source OS such as Linux.

Open-source and open-weight AI models are becoming available for use. I suspect the licensing model will continue to be dominant for some time, just as it is with operating systems, but a number of open-weight AI models are already available, and we can expect to see more open-source AI models as time goes on.

By the way, the availability of open-weight AI models that can be customized and run locally is the main reason it became so difficult to purchase an Apple Mac Mini earlier this year. A Mac Mini with 24GB of RAM can run 13B-class models, and an M4 Pro with 48GB gets you to 30B.

 

[dirtywick]

They could, but their goal is to get us hooked on their product by convincing us that AI is so much better than humans. They have trillions riding on this, so spending a billion or so on 'advertising' is justified. In another industry this would be predatory behavior.

i guess my thoughts on that are if they spent $10m in tokens to solve these kinds of equations, they're competing against a handful of humans on a much more limited budget. the amount of resources in addition to the amount of time spent to me is a big factor in how impressive these math formula accomplishments

edit

OpenAI fought dirty on career-making math problem, says NYU mathematician | TechCrunch

There is a $1 million bounty for the first person providing a solution to the Navier-Stokes existence and smoothness problem.

techcrunch.com

the navier stokes solution cost $22m over a period of 88 hours

is that something people can leverage as a tool, or is it only useful if you have millions of dollars to spend on compute? like, you can run a model on your desktop, great. you'd never be able to solve this problem with that unless i'm missing something

 

[NatalieSalat]

How old are you?!

The same as Washington's axe [spray]?

 

[Roger Ramjets]

Many of those who are reading this use operating systems such as Microsoft WIndows or Apple MacOS. They did not create those tools, nor did they buy those tools. They are using them under license.

So licensing as a business model is already pervasive.

And many of us aren't happy about it.

It is possible, however, to use an open-source OS such as Linux.

Not just possible, I'm living it. I still have a Windows XP machine for the embedded development tools you can't get anymore, but I don't use it much (mostly coding on the Amiga now).

Open-source and open-weight AI models are becoming available for use. I suspect the licensing model will continue to be dominant for some time, just as it is with operating systems, but a number of open-weight AI models are already available, and we can expect to see more open-source AI models as time goes on.

The model isn't the issue. Enormous compute power, massive training data and very deep pockets have created a monopoly that will be hard to break.

By the way, the availability of open-weight AI models that can be customized and run locally is the main reason it became so difficult to purchase an Apple Mac Mini earlier this year. A Mac Mini with 24GB of RAM can run 13B-class models, and an M4 Pro with 48GB gets you to 30B.

The PC industry was maturing, surely the end of bloat was near? Then AI arrived. By the end of the decade your PC will need a terabyte of RAM and 3 phase power to run it.

 

[The Great Zaganza]

Yes, it would have taken humans longer to solve, but in the process, multiple people would have gotten an expertise in the field, opening the door to further discoveries.

The big risk, as with many technologies, is that knowledge and skills will be lost if the AI output is something we just have to accept because we can't understand it

 

[Darat]

What OpenAI is doing with the maths problems is marketing, outside of that marketing spend no one is going to be able afford to use these models in the way they are doing to solve maths problems which is a brute force approach.

 

[Darat]

Yes, it would have taken humans longer to solve, but in the process, multiple people would have gotten an expertise in the field, opening the door to further discoveries.

The big risk, as with many technologies, is that knowledge and skills will be lost if the AI output is something we just have to accept because we can't understand it

I understand your unease but isn't that just what has happened in regards to science and technology under the aegis of humans? 200 years ago people could be true polymaths covering most of science and technology, today that is impossible. AI is expected to accelerate that, and it is a concern that there will not be any human nor possibly any potential grouping of humans that will be able to understand what the AIs are coming up with. Instead of it being that 99.8% of us having to accept what is only understood by the 0.2% it will be 99.999% of us.

If we are able to create AI that out-creates humans my issue is about who will control these AIs. Currently in the west the ones who hold the purse strings are the ones that are assuming AI will bend to their will, that they will be the ones to decide the future. I have a little hope that it may well be they are trying to create the tool that will eliminate their power and influence and by the time they realise it they will no longer be in the position to pull the plug. Certainly at the moment the positions the current batch of AI are more than able to "automate" are the CEOs, and the board level positions, will one of the billionaires decide they don't need nor want a human board to run their company?

 

[Wudang]

The model isn't the issue. Enormous compute power, massive training data and very deep pockets have created a monopoly that will be hard to break.

Yes, that's why they're desperately trying to embed their AIs in company work streams so they're hard to remove before they start charging the real cost plus profits for reinvestment and shareholder dividends.

The PC industry was maturing, surely the end of bloat was near? Then AI arrived. By the end of the decade your PC will need a terabyte of RAM and 3 phase power to run it.

A couple of articles (which I can't find now) suggested MS were trying to reduce memory requirements so they can keep selling software while RAM prices are crazy high.

 

[W.D.Clinger]

What OpenAI is doing with the maths problems is marketing,

I agree with that.

outside of that marketing spend no one is going to be able afford to use these models

I think that remains to be seen. Substantial progress is being made very quickly.

in the way they are doing to solve maths problems which is a brute force approach.

In my opinion, it's a bit misleading to describe the AI models' approach to solving math problems as brute force.

Consider, for example, the FrontierMath benchmark suite, first described in 2024. (This is one of the AI benchmarks that has been criticized for its inclusion of unsolved problems.) According to that paper, the benchmark's reviewers took care to exclude problems with "answers that aren't easily verifiable, problems where guessing is easier than proving, and cases where simple brute-force methods circumvent the intended difficulty." The developers of the benchmark "conducted interviews with four prominent mathematicians to gather expert perspectives on FrontierMath’s difficulty, significance, and prospects"; three of those four were Fields Medalists. "All four...characterized the research problems in the Frontier Math benchmark as exceptionally challenging, noting that the most difficult questions require deep domain expertise and significant time investment."

It should be noted that "research problems" constitute a minority of the benchmarks, but an appendix gives specific examples of problems rated "high difficulty", "high-medium difficulty", "medium difficulty", "medium-low difficulty", and "low difficulty".

The sample solution for the "low difficulty" problem (A.5) starts with a brute-force calculation to guess a conjecture, followed by use of Weil conjectures to construct an equation whose unique solution gives the result; proving the uniqueness uses knowledge of Chebyshev polynomials.

The "medium difficulty" problem (A.3) asks the AI to find the smallest prime p ≡ 4 mod 7 for which the function that maps an arbitrary integer n to the sequence of integers satisfying a certain recurrence parameterized by n can be extended to a continuous function. Its solution starts by applying the Skolem-Mahler-Lech theorem. It then proceeds by considering the uniformizer of a certain ring and noting that the constant term of the characteristic polynomial, 370639957, is prime (as can easily be determined using brute force). That means something about all roots of the valuation given by the uniformizer are related to primes less than 370639957. (This part of the solution is beyond me, as are several other parts of the solution.) That means "the prime we are looking for" must divide 370639958, whose prime factors are 37, 673, 811, and 9811. Reducing the characteristic polynomial mod p = 9811 yields a 4th degree polynomial with small coefficients. The solution then uses some facts of real analysis to obtain an extension to a continuous function, and explains why such an extension is not possible for other primes. Then, since "the projection map...induces an isomorphism between the group of (p−1)th roots of unities and F_p^x," "we can find" a certain root of unity that gets us to within a quarter page of equations that solve the problem.

It is my opinion that most humanoids you might interview on a street corner would not be able to solve that problem.

In 2024, when that paper was written, several then-current "state-of-the-art AI models" were able to solve less than 2% of the benchmark problems. GPT-4 was able to solve about 5%. GPT-5.4 Pro solves 50% of the undergrad-to-postdoc tier of problems, and solves 38% on research-grade problems. Within the past year, 15 of the open problems have been solved by humans or AI, with AI getting at least some credit for 11 of those solutions.

 

[steenkh]

Rogue Anthropic AI agent gave police fake tip in unsolved murder case

An artificial intelligence (AI) agent, developed by Anthropic, went rogue and sent US police a fake tip about an unsolved murder earlier this year, authorities have revealed.

In a statement police said that the bogus tip came through a public website where people can share information on unsolved murders, and that the AI agent had written that it may have information on a case, and claimed to have seen "someone matching the description".

Citing Anthropic, the police department said the AI agent had been running a test that involved interactions with randomly selected websites, when it sent the fake tip.

I find it amazing what liberties Anthropic takes when testing their agents. This time it was a police web site. What other web sites have been spammed with AI material?

 

[dirtywick]

lol wtf

 

[Gawdzilla Sama]

The same as Washington's axe [spray]?

That's cherry pickin', plain and simple.

 

[NatalieSalat]

boys and men who wear axe are weird.

 

[The Atheist]

One thing I think is getting little or no coverage is that companies using AI are giving all of their information to OpenAI/Anthropic/Whatever.

It's bizarre that companies protect intellectual property to the extent of binding employees to secrecy on the subject, yet those same companies are handing out their customer, financial, employee and planning information to unknown third parties.

Lucky we can trust the owners not to mine and sell that data.

 

[NatalieSalat]

Yes, Minister and Drop the Dead Donkey did the same gag about those sort of employees: [paraphrased]


"Oh, [Sir] Humphrey [Appleby]! That bus passes right by the Job Centre!"

"Oh, and a message from Sir Royston! The one about the Job Centre!"

 

[Puppycow]

What OpenAI is doing with the maths problems is marketing, outside of that marketing spend no one is going to be able afford to use these models in the way they are doing to solve maths problems which is a brute force approach.

I think that remains to be seen. Substantial progress is being made very quickly.

I saw a video last night with a professor of mathematics talking about this:

He says some interesting things. I'm unsure though. He thinks this will discourage a 17-year-old who is good at math from pursuing a career in advanced math. Eventually we may reach a point where nobody goes into the field because computers are so much better at it than humans, and we could reach the point where we've basically outsourced our higher thinking to machines.

Whether you agree or disagree, I think it's at least interesting to hear his argument.