On the other hand, if a piece of something I made similarly wound up in someone’s own “creation” because of an LLM, I’d feel empty, as it means nothing. Less than nothing, in fact. I’d be really upset by it.

But why though? Personally, I would just feel nothing, maybe amusement that my work got mentioned at all. LLMs are exactly that — large language models, i.e. massive-scale mathematical algorithms whose inputs and outputs are natural language. They’re tools to solve problems.

In particular, I use LLMs extensively as a LaTeX assistant. (LaTeX is a programming language for typesetting technical mathematical documents.) I do the underlying theory myself, but I use an LLM to refactor my code and debug obscure errors. Learning the ins and outs of LaTeX is neither a very interesting nor enlightening nor useful exercise. I am happy to let an LLM take care of that crap while I go do something fun or useful with my time.

I absolutely recognize that many, if not most, people will not have any use for these tools, and I think it is obscene that the capitalists have hijacked society to put these tools at the center of our existence. That does not make the tools useless.

Additionally, I oppose the “artificial intelligence” framing when talking about LLMs, the broader machine learning field, and frankly the entire lineup of applied statistics starting from the 1950s that have lead us to this point. I do not need nor want truly intelligent machines, and capitulating to the capitalist framing that these machines are even converging towards intelligence is doing free advertising for the capitalists. In my view, these tools ought to be called “statistical learning” algorithms, where “learning” is defined as a group of related mathematical problems wholly and openly distinct from the processes that psychology, philosophy, and biology gives the same title.

And I really don’t see the moral problem with building machines in general to “learn” important tasks, especially ones that are not interesting or enlightening to do as a human. As the probabilistic framework of statistical learning and its offshoots make transparent, you don’t pick a statistical learning solution when your problem has a well-defined algorithmic solution process. You also don’t pick a statistical learning solution when you absolutely require an exact solution in every single run. Typically, you pick a statistical learning solution when a near answer is better than none at all. E.g., if I want to add a custom environment to a LaTeX document, I don’t really care about the exact trajectory that the LLM takes in its state space, as long as it produces a useful approximation of what I asked for in finite time.

Lastly, I want to remark that large language models are, nowadays, part of more complex systems. In particular, LLMs can chain together “classical” bots, classical deterministic algorithms, pre-existing code generation technologies, and even a Linux container to perform certain tasks. So when the technology first came out, if you, say, told the machine to work on a piece of LaTeX code, it would actually just use its large language model to auto complete the most statistically likely response, and it would do this based on the fact that it has LaTeX language references in its training. So it was a roll of the dice whether or not the code would compile, and the machine had no way to check it. Nowadays, I can actually give an LLM the ability to compile LaTeX code. Claude actually has a minimal LaTeX compiler in its Linux container, but I can actually give it access to my actual local LaTeX compiler (I do this inside my own virtual machine, I wouldn’t do this on my actual hardware). So the LLM can literally check its “work”. Even locally hosted LLMs can be given access to external tools. In particular, if you give it SymPy access (SymPy is a Python library for symbolic mathematics), LLMs can “like magic” be decent at math (although they still make mistakes!!!).