I wrote down some thoughts on how I'm approaching LLMs, AI skepticism, and software development.
Also, enjoy my not-AI-generated picture of yarn.
@jaredwhite thanks Jared, I appreciate the feedback! I was trying to remain nuanced most of all, so I'm happy to hear this. I'm still forming opinions about it and agree there are ethical concerns. One thing I considered discussing was copyright issues, it just didn't really fit into the theme of this article (or I got tired of writing about it after almost 3000 words, lol).
@jaredwhite I do think that if devs are going to use these tools either way, it's good to engage positively (returning to that quote from @simon about effective criticism).
@wood @simon I assume you mean this quote?
> finding ways to mitigate [LLMs' problems] and helping people learn how to use these tools responsibly in ways where the positive applications outweigh the negative
I definitely don't agree with that. Replace "LLM" in that sentence with "AR-15", or "racemic methamphetamine", and you get a sense of how I find the sentiment a little nonsensical.
I'm glad Simon welcomes criticism of this technology space, but on the scope of it, we are not aligned.
@wood @simon a person in Simon's position seems to be asking how we can help guide ChatGPT, Claude, Copilot, etc. to be built and used more responsibly.
I am saying these tools should be illegal. Just as I think assault rifles should be banned.
Right now we don't even have *any* meaningful regulation on commercial AI, which makes my brain melt. It's a great tragedy.
@jaredwhite @wood what should we ban here?
Tools that translate from English to Spanish?
My dumb little LLM that outputs garbage text that looks a bit like my blog? https://til.simonwillison.net/llms/training-nanogpt-on-my-blog
This 1.28GB openly licensed binary file that runs in Google Chrome and can solve surprisingly complex problems (given its size)? https://huggingface.co/spaces/webml-community/deepseek-r1-webgpu
@jaredwhite @wood I'm strongly in favor of regulations that govern the way these tools are used - things like Colorado's legislation around using black box AI models to make decisions about insurance claims for example https://www2.deloitte.com/us/en/pages/risk/articles/ai-regulation-colorado-dept-of-insurance.html