Quote of the Day
Newer AI models are now being trained, more and more, on content either generated by itself or other AI models.
A researcher that works with Monash University, Jathan Sadowski, recently described a hypothetical self-trained AI as an “an inbred mutant, likely with exaggerated, grotesque features.”
More recently, a study was published in Nature that tested the capabilities of AI trained on itself, and the results were concerning.
By the 5th cycle, the degradation had become stark, and by the 9th, it was almost completely nonsensical.
Benjamin Carter
August 27, 2024
Model Collapse: Why AI Is Probably Getting Worse, And It May Be Fatal (msn.com)
See also AI Appears to Be Slowly Killing Itself (msn.com).
One of the truisms of machine learning is that result cannot be any better than the data used for learning. And, in fact, it can only be some fraction as good as the input data. Hence, if your learning is 90% as good as in the input and your input was based on AI output which was only 90% as good as the original source, your second-generation AI cannot be expected to be any better than 81% as good as the original.
Proponents of LLMs and other AI confuse the map for the terrain.
The other failing of AI is that it doesn’t have a feedback system connected with the terrain (that is, reality).
Put another way, words in a database are not congruent with reality.
Kurt
What an excellent analogy: Confuse the map for the terrain
Imagine if all cartographers stopped building their maps from measurements of terrain, and just sped things up by making them from other peoples maps?
The only way AI can “improve” is if you can keep *ALL* AI generated data *out* of the training data, and even then it only improves if the human generated data it is trained on improves.
Unfortunately, the only value in AI is as an “assistant” to people, so the AI generated data is being directly fed to the people who are the source of the data the AI is trained on. In this case, feedback is noise! Further, people, being lazy, will lose creativity because they can just *steal* it from other people people via AI. So the data will drop in value.
This is the same trap as socialism: This is people stealing information and “creativity” from other people instead of producing something themselves. The result is that all production grinds to a halt, all creativity ceases, as everthing becomes an arbitrary and capricious derivative of everything else.
The only way AI can become smarter than people, is to do what people are doing, better and/or faster. Large Language Models cannot do this, it’s a dead-end technology. If they work out how people do it, and do that, maybe they’ll get somewhere…. or maybe they’ll discover that Penrose was right, and that human creativity is not sourced in neurons, but in quantum functions of microtubules. In other words the source of human creativity is not physical, nor of this “universe”.
Ummm, should that read “81%” rather than “0.81%”?
Yes.
Fixed.
Thank you.
Agreed.
Pure conjecture on my part.
But could it be the real problem (other than just being ignorant enough to try and build a machine that thinks, in the first place), Is that a screwed up human is dictating “truth” to it?
Making historical American founders black was truth to AI? Or was AI in on the joke?
You can ask a political question and it will go completely hypocritical without a second “thought”.
One small data center holds the entire library of congress and the patents office. So excess to truthful information should not be the problem.
To me, something/someone seems to be blocking “truth”?
You’re alluding to another inherent problem with generative AI models, which is that the output is always run through the sociopolitical lens of the developers.
A perfect, non-biased developer could (in theory) build an AI or LLM that “learns” only from verifiable facts and truth, and generates output based upon raw logic that applies those facts and truth.
But developers are human, with all our flaws, and perfect, non-biased developers don’t exist. Add in that facts and truth are difficult to verify on the information jungle we call the Internet, and that even authoritative sources have their biases. Given all that, how can an AI or LLM know what to absorb and what to reject, except to follow the directions of its developers?
The “black Founding Fathers” images were generated because the developers programmed the AI to include more “people of color” in generated images — after it was “discovered” that a lot of AI-generated images depicted “too many” white people — and that parameter didn’t differentiate between pure fiction and historical fact.
It’s not the AI’s fault; it’s the programmers’. Their bias got incorporated into the generative parameters, and thus the output was historically inaccurate.
But every AI and LLM is going to have that problem to some extent or another. The developers’ sociopolitical views WILL bleed over into the learning models and output data.
So, AI didn’t search all the available data and see there were no black founders? Or very few?
And report that to it’s programmer?
If it can’t do that, it’s almost all artificial, and very little or no intelligence.
Which makes it seem like a lot of hype?
But with the huge amounts of money being dumped into data centers? That’s a giant bet on just hype.
It ain’t passing the smell test.
I mean, nobody dumps that much money into something that stupid?
Don’t answer that, I’m going to shut up now.
the output of AI will always be *exactly* as hypocritical, *exactly* as concerned with “the truth” or “reality”, as the data is is trained on.
If you hear a presenter use any anthropomorphic terms like “understand” in relation to any text LLM grab your wallet and activate your BS detector. LLMs are a statistical model of the text they have been given to analyze. There is no human understanding of irony or sarcasm and no generalization of responses.
If you don’t add domain-specific information to the database all you ever get is (almost) grammatically correct responses to your prompt that are consistent with the statistics of the training set.
Both of these (the original post and the first comment by Rick T) Sound a LOT like some of the folks I know that live in certain major metros and/or go to any number of Universities.
Perhaps we have had AI among us for longer than we thought, creating new generations from themselves and teaching them from their own knowledge?
The whole notion of AI is based on a defective understanding of what computers and programs are and how they operate. A computer is a flexible automaton, and a program is a precise specification of a specific automaton. That’s true for all computers (give or take noise and hardware defects) and all programs. “Bug” is a slang term to describe a case where the precise specification was different from the “intended” specification, for some suitable definition of “intended”.
The above is true for all programs including “AI” programs such as “learning” or “large language model” programs. But there is a practical difference. “Conventional” programs can be fairly well understood as data transformers: for a given input they produce a particular output. That’s not fully true because in general they have internal state, which affects the transformation. For many conventional programs the influence of that internal state is small compared to the influence of the inputs. For example, a word processor, for a given set of inputs (UI actions and perhaps input files) produces a particular output.
AI programs have very large internal state, that’s what the “large” in “LLM” refers to. The program takes a given input, and those terabytes of internal state, performs a transformation that is of course precisely specified but not at all understood by any human, and produces some output. The programmers hack on the program until the output starts to look sort of like what they want it to look like in the small body of test cases they try. So AI programs are like Microsoft products, only vastly more so. This is why AI cannot serve, and never will be able to serve, in any safety critical application: no one knows and no one can tell you what the actual properties of the system are.
Apart from the obvious craziness of “training” an AI on the outputs of some other AI, the whole notion of training from external inputs has a problem too: a whole lot of the body of material used is fiction, either intentionally (novels etc.) or unintentionally (wrong notions written by someone). But the inputs aren’t tagged, so how can an LLM distinguish between, say, the writings of Albert Einstein and Larry Niven?
I came across a very good and normie-friendly post about this exact thing last week. https://wattsupwiththat.com/2024/08/27/ai-model-collapse/
The image it has illustrates the problem clearly.
I think if the AI programmers were honest about the data input, used only verifiably human inputs, and were not trying to stamp their own world-view on it by deliberately omitting “inconvenient facts” while also force-feeding it material of dubious quality (but of the “right” political perspective, then it could be hugely beneficial to humans. However, the deep pockets funding it all are far more idiologically driven than they are good (in the good v evil sense.)
Didn’t French royalty prove a couple hundred years ago that incest, at any degree, constitutes the opposite of “improving the breed”?
European royalty in general. It was worse than the stereotypes about hillbillies. When the Russians found the bodies suspected to be those of their executed royal family, they turned to DNA. After some genealogical research, the determined that Prince Philip (of England) was the closest living relative of both Nicholas and Alexandra. George V, Nicholas II and Wilhelm II were first cousins, grandchildren of Queen Victoria. The Romanovs often acted like they were French but actually they were Germans. Depending on who actually was the father of Catherine the Great’s son, they may have been all German.
I was thinking of haemophilia propagated by frequent intermarriage and the occasional incest, but OK.