[status: rough draft]

One type of post I'm going to include in my meta blog is to record one-liner ideas that I plan to post and potentially try to grow into more full blogs. Some of these may be more along the lines of repeating / trying to socialize I point someone has already made. Others may be testing out an idea.

Here's what I've got! I have 3 big things to tweet about.

  • I think it's worth noting that "sovereign ai" (for certain operational definitions) has some deep inherent tensions with the whole foundational model and LLM paradigm. Parametric knowledge including core reasoning is a collectivization across borders. Everyone using foundation models is benefiting from a very global public/club good. That said, it can be reasonable to build sovereign data stores on top of this (but this isn't really what most people are talking about)!

  • Worth noting that any attempts to build systems that "augment but don’t replace" is fundamentally about privacy and friction in information flow; we must withhold certain information such that a person's brain or API keys provide that information at run time to enable task completion. Otherwise, a adequately capable model (especial if this model has built in reasoning and exploration capabilities) can always just do the thing itself. If organizations want to take this mission very seriously, they're going to need to engage with differential privacy, membership inference attacks, etc.

  • I’m feeling a bit more hopeful about training data economics. Evaluation involves data labor — I think the eval ecosystem (which is seeing lots of momentum and attention) can create a high-potential entry point for build healthy markets for data labor broadly. Big question is: can we make eval “good jobs” or doomed to precarity?

  • I want to revisit https://dataleverage.substack.com/p/measuring-relative-ai-alignment-in-terms-of-data-pipelines

  • I think in economic impacts of AI discussions, one perspective that could use more attention is viewing AI progress as equivalent to "just open sourcing everything".

    • Thought experiment: what if the medical field just open-sourced everything (e.g. made uptodate free, etc.)? How is this different from current AI progress

    • If AI labs fail to capture gains, they may end up effectively doing this...

    • Here quasi-enclosure framing is helpful, I think?