Is it just me, or are the top 0.001% becoming increasingly annoying and dangerous? Given the effectiveness of today's family trusts in preserving wealth from taxes, stupidity, and fraud, this problem may only get worse. The risk free returns on billions is tens to hundreds of millions per year, mostly in tax free shelters. They can't buy yachts and tropical islands fast enough, and Congressmen are surprisingly affordable. Nor will their heirs.
It is most certainly very clear that Ackman has scarcely any grasp of what present neural network AI is truly all about and just how fundamental its limitations really are. It is possible that our "AI" friends will someday manage to find a way to truly intelligent AI but right now it is far from clear how, and Ackman is one of the last people whose advice on such matters should be heeded. Unfortunately, he and the thundering hordes of those of like mind suck a great deal of energy and resources out of serious research.
Another moron who thinks the fact he can play the market means he's smart. Time to start eating the rich. I hear they're tasty, lightly sauteed with a nice Chianti.
In effect, once the tables are turned, Ackman is joining the "no big deal" defenders of Claudine Gay, with whom I have some sympathy. He assumes that, with a careful check, everyone in academia will be found to have omitted quotation marks, or quoted Wikipedia improperly. He might be right.
If it's also true that students are being punished severely for small-scale errors of this kind (as Gay's critics have claimed), then maybe we should relax our standards a bit. But Turnitin usually picks up such things - the real issue with students involves essay writers for hire, reuse of essays from students who have taken the course previously and so on.
Speaking as somebody that used to actually do computational lingustics for a living: If you were only worried about a strict string-matching algorithm, the tiniest paraphrase would let you hide. There are various word-substitution and syntactic-analysis things you could do with a rules-based model that would let you identify, say, a string of 50 words that matches source other than the occasional swap. A "fuzzy" string-level match.
But I think it is almost certainly correct that the current generation of LLMs either can, or at least could, do a better job than this, at identifying strings of 50+ words that are likely to be "the same". You might need to train it on a few thousand examples first of not-identical texts that were, in the past, identified as significantly plagiaristic, and then after that, it can compare each paragraph of a paper, to its entire corpus, and spit out instances where it sees a likelihood of plagiarism.
The question would be, as you say, how many false positives it turns up, and how good it can get if you give it feedback when some calls are wrong. My _suspicion_ is that while it would make a lot of false-positives at first, it would get better after a few hundred of those were pointed out to it. And ultimately the goal with it would be that you don't bother to move on to the next level of human review unless it's found _multilple_ examples in a given person's body of work.
Fundamentally I think it is correct that the current-gen LLMs will be possible to train to be _much_ better at this kind of work than a mere fuzzy string match algorithm.
Of course, this all is an enormous and absurd waste of resources, but, well, Bill Ackman is an absurd man.
The strange part to me is that of course no one can survive AI interrogating their work, when LLM AI will simply say, "Yes, of course" until you say "No, you're wrong" in which case it will volte-face immediately into "Yes, of course I'm wrong".
I need to have ChatGpt rewrite every quote I failed to put quote marks around the sentence or paragraph I lifted without a footnote.
This is such a joke.
Particularly when they use AI to detect these failures to attribute or cite a source. ChatGpt does this all the time, using its training set.
In reality, what a researcher does is read a whole lot of stuff, and gets trained on it, as does CjatGpt. One get's outed for doing the work of understanding what was read, and the other does not.
The real question is innovation: will we be tied to the learning of the past, which serves as the training set of ChatGpt, and will we give recognition, or be able to recognize, innovation in a paper, even if there are places where someone didn't give a citation or place a quote mark.
This is an innovative comment which has not been generated by ChatGpt, although I am working on a literature review which certainly will be generated by ChatGpt.
I view Ackerman's tweet as a continuing deflection of the attention on his wife. Ackerman was the "bully" who forced hay's resignation by demanding her palgiarism was a firing offence. Then his wife was outed but he, reasonably, defended her. What he should have done is acknowledged her plagiarism and retracted his deman gay be fired. But no, he has deflected by trying to show everyone in academia has plagiarized work to some extent. But while academices shouldn't plagiarize, why not extend this analysis to everyone in power to extract their plagiarisms? If "everyone does it", why demand such a universal trait be be focused on an academic institutional leader?
If one wants to use an LLM at all, use it to detect outright lies? Then target it at politicians, institutional leaders, religious leaders, pundits etc. That would be far more important than plagiarism.
Is it just me, or are the top 0.001% becoming increasingly annoying and dangerous? Given the effectiveness of today's family trusts in preserving wealth from taxes, stupidity, and fraud, this problem may only get worse. The risk free returns on billions is tens to hundreds of millions per year, mostly in tax free shelters. They can't buy yachts and tropical islands fast enough, and Congressmen are surprisingly affordable. Nor will their heirs.
It is most certainly very clear that Ackman has scarcely any grasp of what present neural network AI is truly all about and just how fundamental its limitations really are. It is possible that our "AI" friends will someday manage to find a way to truly intelligent AI but right now it is far from clear how, and Ackman is one of the last people whose advice on such matters should be heeded. Unfortunately, he and the thundering hordes of those of like mind suck a great deal of energy and resources out of serious research.
Another moron who thinks the fact he can play the market means he's smart. Time to start eating the rich. I hear they're tasty, lightly sauteed with a nice Chianti.
https://yastreblyansky.substack.com/p/of-the-faking-of-brooks?utm_source=profile&utm_medium=reader2
In effect, once the tables are turned, Ackman is joining the "no big deal" defenders of Claudine Gay, with whom I have some sympathy. He assumes that, with a careful check, everyone in academia will be found to have omitted quotation marks, or quoted Wikipedia improperly. He might be right.
If it's also true that students are being punished severely for small-scale errors of this kind (as Gay's critics have claimed), then maybe we should relax our standards a bit. But Turnitin usually picks up such things - the real issue with students involves essay writers for hire, reuse of essays from students who have taken the course previously and so on.
Speaking as somebody that used to actually do computational lingustics for a living: If you were only worried about a strict string-matching algorithm, the tiniest paraphrase would let you hide. There are various word-substitution and syntactic-analysis things you could do with a rules-based model that would let you identify, say, a string of 50 words that matches source other than the occasional swap. A "fuzzy" string-level match.
But I think it is almost certainly correct that the current generation of LLMs either can, or at least could, do a better job than this, at identifying strings of 50+ words that are likely to be "the same". You might need to train it on a few thousand examples first of not-identical texts that were, in the past, identified as significantly plagiaristic, and then after that, it can compare each paragraph of a paper, to its entire corpus, and spit out instances where it sees a likelihood of plagiarism.
The question would be, as you say, how many false positives it turns up, and how good it can get if you give it feedback when some calls are wrong. My _suspicion_ is that while it would make a lot of false-positives at first, it would get better after a few hundred of those were pointed out to it. And ultimately the goal with it would be that you don't bother to move on to the next level of human review unless it's found _multilple_ examples in a given person's body of work.
Fundamentally I think it is correct that the current-gen LLMs will be possible to train to be _much_ better at this kind of work than a mere fuzzy string match algorithm.
Of course, this all is an enormous and absurd waste of resources, but, well, Bill Ackman is an absurd man.
He should call it The Lobochevsky Project.
Further proof that Hedge Fund managers are not smarter than the rest of us.
The strange part to me is that of course no one can survive AI interrogating their work, when LLM AI will simply say, "Yes, of course" until you say "No, you're wrong" in which case it will volte-face immediately into "Yes, of course I'm wrong".
I need to have ChatGpt rewrite every quote I failed to put quote marks around the sentence or paragraph I lifted without a footnote.
This is such a joke.
Particularly when they use AI to detect these failures to attribute or cite a source. ChatGpt does this all the time, using its training set.
In reality, what a researcher does is read a whole lot of stuff, and gets trained on it, as does CjatGpt. One get's outed for doing the work of understanding what was read, and the other does not.
The real question is innovation: will we be tied to the learning of the past, which serves as the training set of ChatGpt, and will we give recognition, or be able to recognize, innovation in a paper, even if there are places where someone didn't give a citation or place a quote mark.
This is an innovative comment which has not been generated by ChatGpt, although I am working on a literature review which certainly will be generated by ChatGpt.
As Newton said: I see so far because I stand on the shoulders of giants.
Wow.
My guess is that the repeated plagiarism of Caudine Gay and Neri Oxam is rare. Let's see what Bill finds. He is short academic integrity.
I view Ackerman's tweet as a continuing deflection of the attention on his wife. Ackerman was the "bully" who forced hay's resignation by demanding her palgiarism was a firing offence. Then his wife was outed but he, reasonably, defended her. What he should have done is acknowledged her plagiarism and retracted his deman gay be fired. But no, he has deflected by trying to show everyone in academia has plagiarized work to some extent. But while academices shouldn't plagiarize, why not extend this analysis to everyone in power to extract their plagiarisms? If "everyone does it", why demand such a universal trait be be focused on an academic institutional leader?
If one wants to use an LLM at all, use it to detect outright lies? Then target it at politicians, institutional leaders, religious leaders, pundits etc. That would be far more important than plagiarism.
Will Chat GPT plagiarize Chat GPT?