As far as this site goes: I write my own content. When I use LLMs with regards to my blog, it’s to do a grammar/spelling checks, formatting image captions from a text file, or figure out CSS issues. But at the end of the day: my words are my own. First appeared in Patch Notes August.

A list of things I hold to be true right now:

  • LLMs are one of the most unexpected, fascinating, and divisive technologies to pop out of Silicon Valley in the past decade
  • LLMs as a whole have considerable capacity to significantly affect - in both positive and negative ways - white collar wages and organizational structures, copyright law, intellectual property, legal advice, design, advertising, research, and information access
    • But I do not believe them to be the path / “skill tree” that gets us to AGI / hard takeoff / paperclip maximizer scenarios. That is not to say the downside risk is limited however; companies like OpenAI and Anthropic need to be regulated more, and not in the classic regulatory capture playbook the big tech companies have gotten so good at.
  • Coding/programming/software engineering today already looks very different to how it did a year ago - and it will continue to change, rapidly. The days of writing code by hand are gone, but the days of developing expertise, craft, and standards in reducing complexity in reality are not.
    • If I am generating code with AI that other people may need to understand, maintain, update, or fix in the future: I read it and make sure that I understand it, and can stand by it.
    • If I am writing code that only my future self has to deal with, I don’t bother reading it closely.
  • Media (text, images, videos, music, code) I produce with generative AI is labeled as such, where appropriate. I do not believe using them to generate work - for fun or for profit - is wrong, but I want to make sure it is clear how a piece of media was created.
    • Partly this is because the inaccurate-but-generally-true relationship between output & time invested has been subverted.
    • Partly it’s to give credit to the people whose data was used in the training of the models; likely without their consent.
    • And partly because I do not think obscuring how we make things generally leads to creating better things.
  • LLMs are an amazing resource for understanding beyond one’s current capacity of knowledge, but they should not be relied on for teaching someone something by themselves, especially children (ie, the Primer in Neal Stephenson’s Diamond Age)
  • The most likely and interesting near-term future to me is open source models running on infrastructure somewhere between a very nice laptop and a server cluster in a data center - hardware stacks that a school, coworking space, or even a group of families in a neighborhood can afford to invest in and share, a bit like old modes of sharing time on a central computer.