Archive link: https://archive.ph/GtA4Q
The complete destruction of Google Search via forced AI adoption and the carnage it is wreaking on the internet is deeply depressing, but there are bright spots. For example, as the prophecy foretold, we are learning exactly what Google is paying Reddit $60 million annually for. And that is to confidently serve its customers ideas like, to make cheese stick on a pizza, “you can also add about 1/8 cup of non-toxic glue” to pizza sauce, which comes directly from the mind of a Reddit user who calls themselves “Fucksmith” and posted about putting glue on pizza 11 years ago.
A joke that people made when Google and Reddit announced their data sharing agreement was that Google’s AI would become dumber and/or “poisoned” by scraping various Reddit shitposts and would eventually regurgitate them to the internet. (This is the same joke people made about AI scraping Tumblr). Giving people the verbatim wisdom of Fucksmith as a legitimate answer to a basic cooking question shows that Google’s AI is actually being poisoned by random shit people say on the internet.
Because Google is one of the largest companies on Earth and operates with near impunity and because its stock continues to skyrocket behind the exciting news that AI will continue to be shoved into every aspect of all of its products until morale improves, it is looking like the user experience for the foreseeable future will be one where searches are random mishmashes of Reddit shitposts, actual information, and hallucinations. Sundar Pichai will continue to use his own product and say “this is good.”
Is this real though? Does ChatGPT just literally take whole snippets of texts like that? I thought it used some aggregate or probability based on the whole corpus of text it was trained on.
It does, but the thing with the probability is that it doesn’t always pick the most likely next bit of text, it basically rolls dice and picks maybe the second or third or in rare cases hundredth most likely continuation. This chaotic behaviour is part of what makes it feel “intelligent” and why it’s possible to reroll responses to the same prompt.
I remember doing ghetto text generation in my NLP (Natural Language Processing) class, and the logic was basically this:
This is a rough explanation of Baysian nets, which I think are what’s used in LLMs. We used a very simple n-gram model (e.g. n words are considered for the statistics, e.g. “to my math” is much more likely to generate “class” than “homework”), but they’re probably doing fancy things with text categorization and whatnot to generate more relevant text.
The LLM isn’t really “thinking” here, it’s just associating input text and the training data to generate output text.
Sounds quite similar to Markov chains which made me think of this story:
https://thedailywtf.com/articles/the-automated-curse-generator
Still gets a snort out of me every time Markov chains are mentioned.
Yup, and I’m guessing LLMs use Markov chains, which are also a really old concept (the idea is >100 years old, and it’s used in compression algorithms like LZMA).
Yeah I’m not an AI expert, or even really someone who studies it as my primary role. But my understanding is that part of the “innovation” of modern LLMs is that they generate tokens, which are not necessarily full words, but simply small linguistic units. So basically with enough training the model can learn to predict the most likely next couple of characters and the words just generate themselves.
I haven’t looked too much into it either, but from that very brief description, it sounds like that would help to mostly make it sound more natural by abstracting a bit over word roots and considering grammar structures, without actually baking those into the model as logic.
AI text does read pretty naturally, so hopefully my interpretation is correct. But it’s also very verbose, and can repeat itself a lot.
Most LLMs are transformers, in fact GPT stands for Generative Pre-trained Transformer. They are a different to Bayesian networks as transformers are not state machines, but rather assign importance according to learned attention based on their training. The main upside of this approach is scalability because it can be easily parallelized due to not relying on states.
Back in my day, we called that “hard-mode plagiarism.” They can’t punish you if they can’t find a specific plagiarized source!
This is not the model directly but the model looking through Google searches to give you an answer.