On LLM AIs
I have an irrational aversion to them.
I do not want to use them, they take all joy I have from my work.
More pertinently to this article, I am not an expert at machine learning. Most of my knowledge comes from attempting to understand research papers, and from incredible channels like 3Blue1Brown and sentdex. But I have not actually worked at a large language model company, or even a machine learning company of any kind, I have made some toy classification models, but that does not give me credence to talk authoritatively on this subject. Now with that out of the way, we can start with my rational reasoning behind why I don't like them.
An AI doesn't have to be sentient to ruin the world
The effects of AI in it's current form have already massively affected the world, from young kids using it to do their homework for them, to even more degraded customer support. There have been cases of professionals skimming their work by taking the output of these things as though it's a search enginge that gives perfect answers, and attempted to, or actually acted on the horrible advice it provides. Google has been inserting these in place of the old quick results feature, so now you can learn such awesome facts like, 9.9 is smaller than 9.11, or that 1 meter per second is 2.237 miles per hour.
Script writers, character writers, book authors are being replaced everywhere. We need food, and water, and sensory organs to survive. A model may never die of thirst. To quote an IBM internal presentation slide from the 1970s, "An AI can never be held accountable, therefore AI must never make a management decision".
They are just compression algorithms with a powerful retrieval mechanism.
It is a strong, but hopefully incontroversial opinion of mine, that LLMs are not sentient.
I've been trying to chase down an article or a video that explained this very well, but gist is this, a leading theory for how humans remember things is very similar to how a neural network does it as well. Instead of storing memories directly, the state of the brain is adjusted such that given a particular sensory input, a process is initiated where groups of neurons fire to change the recall space of the mind to an approximation of what the memory was, similar to hamming codes for error correction, just over much bigger space.
The reason this is possible, as is beautifully illustrated by a recent 3b1b video How might LLMs store facts, a vector in a high dimensional space has a capacity to point in a unique enough direction to form an n exponentially larger basis for a far more massive state space of vectors that are almost orthogonal to each other if is allowed to point in a direction just 2% away 0.02 > | ⃗a . ⃗b | > 0 creating space to store unimaginably large amount of information in, imperfectly.
The copyright issue
To create these models, massive swaths of the public internet have been indiscriminately raked for any morsel of fresh training data, irregardless of licenses, ignoring pleas of website hosts to not scrape their servers, piling millions of automated requests over weeks causing immense strain to small or self hosted services. The main reason this has not been litigated into oblivion, is largely because prominent book publishers, and media publishers at large in case of generative image models are being avoided, or detected and hastily censored from the model output, while everyone else is not.