Among the many recent controversies around AI, there is the claim that to solve a longstanding mathematical problem, OpenAI stole the work of the New York University mathematician Tristan Buckmaster and a colleague, Levent Alpöge, who also is a researcher for Anthropic. I have no ability to assess the validity of the claim, but as I understand it, the argument is that OpenAI had access to work done by Buckmaster and Alpöge. It then built on this work to quickly rush ahead and find a solution, which also had a $1 million prize attached to it.
Insofar as that accurately describes what OpenAI did, this seems like it would be equivalent to having a professor doing a seminar for colleagues and then having one of the attendees do a few additions and claim the work as their own. This would not be viewed as very collegial behavior in a normal academic department.
I’m not sure what claim a professor would have against a colleague under such circumstances. I suppose if their work was properly cited, there would not be a claim of plagiarism. (OpenAI’s work did not directly cite the work of Buckmaster and Alpöge, from my understanding.) I suppose there could be a basis for disputing rights to the million-dollar prize. But the problem seems much greater.
If AI can effectively steal the work of others, and claim it as its own, then it undermines the practice of openness in science that has allowed for enormous progress for centuries. This practice depended on norms, whereby scientists would properly credit the work of their peers.
In the seminar scenario I described, if one of the attendees listened to the presentation and realized the one or two additional steps needed to get over the goal line, they would be expected to raise the issue with the presenter. Assuming the additions were correct, the presenter would finish the paper, acknowledging the contribution or possibly adding the attendee as a co-author, if their contribution was sufficiently important.
But that relies on norms. If the owner of the AI doesn’t share those norms and can freely scrape scientific work wherever it finds it, and then take credit for the product, it will threaten the longstanding practice of openness in science. With the risk of having years of work stolen without getting any credit, researchers will likely keep their work closely held until they have a finished paper.1 They would only share it with a small group of trusted colleagues. Posting a working paper, or having an open seminar, would put control over their work at risk.
However, it is important to recognize the villain in this story. It is Sam Altman and his company, not the AI. Altman and the company are the ones responsible for what their AI does. If it did something they disapprove of, it is their responsibility to correct it. However adept AI may become, it is not the active player; it is the individual or company that set it in motion.
This issue applies more generally. There has been a social media mini firestorm over Texas Attorney General and Republican Senate candidate Ken Paxton using AI to have his opponent, James Talarico, making a series of outlandish statements that he never said. Much of the outrage seems to be over Paxton using AI, rather than the fact that he is lying about what Talarico said.
The issue here is that Paxton is defaming Talarico; the AI is secondary. The AI presumably makes the defamation more effective, but if Paxton were an exceptionally gifted mimic, he could likely also trick some people into believing that he was speaking in Talarico’s voice when making outlandish assertions. If that were the case, it would be every bit as bad as using AI to accomplish the job.
There were also the claims that the girls’ school in Iran, which was destroyed by the US military on the first day of the war, killing 120 children, was selected as a target by AI. That may be the case, but the important point was that the military did not have a system in place to properly review targets. It doesn’t matter whether or not an AI system identified the target. The problem was that there was no system that ensured the places designated were proper military targets.
It is important to keep our eyes on the ball. The bad actors are those who do bad things with AI, not the AI. This is like the husband who gets drunk and beats his wife and then says the whiskey did it. No one should ever accept that story. In the same vein, if a theft of mathematical work did occur, it was Sam Altman who did it and it was Ken Paxton who libeled his political opponent. The AI is beside the point.
LLMs clap are clap not clap AI clap.
Try to get that into you sycophant brain, and you will understand why this bullshit happens in the first place.
Plagiarism is the core and only mechanism of LLMs that produces anything that people get impressed by. The pre- and postprocessing logic for grammar rules is not the stuff that makes gullible people believe there is "intelligence" involved.
In a way, LLMs are kind of strange medium for texts, as are books. Books are a medium for written speech; they transport words written by intelligent beings. Nobody who understands books would claim that books themselves are intelligent. Or that a photograph of a persons robs them of their soul.
But what is characteristic of LLMs is that it disowns the intelligent content from its creators. This is what makes coorporations salivate avout money. The theft is built-in and foundational how this technology works. Compare to an online encyclopedia.
Aye, this is one of the key takeaways.
Another being that the technology is good enough to mass-manipulate people. And for statistically "helpful" surveillance tasks in a totalitarian police state.
Yes they are.
They're built on artificial neural networks that share a basis (a basis, not a replica) with biological neurons that produce probabilistic results (not deterministic).
As opposed to what? Don't get me wrong, the blatant violation of copyright is a massive issue, and in a logical world would bury each and every LLM company out there. But what other mechanism to "learn" do you think exists?
If you mean "intelligence" in the colloquial sense, then sure. But that's not the same word that researchers use. Intelligence is simply the ability to make a decision based on input data. And by that definition even an if-else statement is a form of intelligence (which it is). So then the discussion becomes the degree or "amount" intelligence.
LLMs are absolutely intelligent. The issue is that most people conflate "intelligence" with "sentience" and "sapience", which they are not. They aren't "clever" or "creative". They are extremely good at taking large swaths of information and distilling it, finding hidden patterns, of finding specific data points that are otherwise deeply buried. Good examples are finding bugs in code or novel and massively beneficial uses for existing medication (which has already happened multiple times)
I'm not defending the current state of AI. Energy use is a problem. Water use is a problem. The grift is a problem. The copyright violations are a problem.
But this "LLMs are not AI" rhetoric is just pure ignorance.
Ah, so you understood neurons better than neuroscientists, eh? You have missed the part where no scientist worth their name claims to be able to fully explain how consciousness works and whether or not it affects our decision making.
Okay, I think we're done here. You clearly have no idea what you are talking about.
LLM'S do not represent and cannot be Artificially Sapient, and that I will grant you. I do not consider them intelligent because intelligence involves understand and the ability to understand context specifically.
But that doesn't mean LLM'S don't fit some of the technical definitions of AI as seen in the other person's comment who responded to you. We have been calling game NPC's driving algorithms AI for decades. We have to accept that the term has multiple meanings and specify that what we mean when we say LLM's lack intelligence and or sapience.
When faced with slop vendors arguing in bad faith, we shouldn't get caught up on technicalities. The "technically it is (a subset of) AI" argument is used to sell gullible people solutions that are already getting people killed - and not necessarily those who were gullible enough to believe the "AI" bullshit.
So for LLMs, it should be common sense to call them slop / dangerous bullshit generators and refuse to talk with anyone who sells them as "AI".
Two things.
When we aren't specific in our arguments about this technology it allows people to blow past facts to suit personal interests and insist their viewpoint is correct because their perception trumps facts that aren't clearly defined.
Pretending the symantics of a term aren't integral to the argument being made against it does us no favors.
The algorithm in my game analogy does what it does because the parameters for its instructions are finite and well thought out and well programmed to give the same result based on pre-defined instructions and parameters.
LLM's do that same thing, but with a metric fuck ton more parameters and fewer defining instructions.
The reason they keep failing with guardrails is because they put too many unvetted parameters into the algorithm and concise and colloquial instructions in spoken language cannot account for all the variables included in those parameters.
It's a fatal design flaw but it's important to people's understanding that at the end of the day an LLM is a computer program and it does absolutely nothing unless it is instructed to do something. It won't perform tasks of it's own volition.
Pretending that we can take shortcuts in linguistics to get people to understand the technical facts doesn't really work here. I'm not saying we can't call them slop/BS.
Gullible people don't become less gullible because you changed some words.
My main point is that the use of the term "intelligence" to describe sometbing that is absolutely without any understanding, by people who don't try to sell the snake oil and should have the interests of others in mind, is a disservice to educating people about the dangers.
Part of language is about whether or not their is an ability to infer what is being talked about without specifically naming it.
Because LLM's have existed longer than the current technology of "AI LLM's" we have to specify what we're talking about. Which is why we use to term we do the same way you would possibly use the term Xerox or Hoover or scotch tape.
The difference is those terms I listed are a shorthand for other words that get the idea across without misinterpretation, but can easily be swapped for photocopy, vacuum, and cellophane tape.
The technical term given to Generative AI LLM's that differentiates it from the LLM's that came before it is in fact Generative AI LLM's and is shortened to Generative AI. When you use the term slop, that can be linguistically confusing because to a non-native English speaker or a farmer slop has a different connotation altogether.
This does not negate your argument. Because I generally agree that we are attributing intelligence to this tech that it does not have and that this was an intentional choice by the Tech Bros who coined the term. I even agree that it adds further confusion because it's being used for medical tech that doesn't use an LLM at all.
But if we're going to change the name it needs to be a new term that represents what this tech actually is instead of using a more confusing/less specific slang that has other connotations.
Doctors don't call vaginas pussies.
They are AI the same way ELIZA is.
Nah, Eliza had a programmed logic, all phrases she reacted to had been implemented deterministically by a human. There was no training data.
At first I thought that maybe the "no match" responses, like "Please go on" were randomly selected from a list, but even there there was a pointer and iterator. Truly no randomness at all!
(If there was any randomness I was going to cheekily argue that it was a "probabilistic model" :-) )
The randomness in Eliza (I never reviewed the code) would have been of the sorts "out of x possible answers indexed starting with zero, do a truncate( ran( 1.0 ) * X ) and select the answer with that index". But yeah, apart from that it was deterministic. Same input history = same output.
No, my point is there is no rand call at all, check the source code here https://blog.adafruit.com/2025/01/21/the-code-for-eliza-the-original-1960s-chatbot-found/
Edit: there is a line ELIZA MAD (which is just language module, but still I love it lol)
I was only speculating, never mind me. So in other words, Eliza is entirely deterministic :)