AI Abundance Thought Experiments
Experiment 1
What if I bookmark some of the most hypest of hype tweets, and revisit them 6-9 months later, after compute/inference costs have fallen by 90-99%.
it used to be all the rage to make software factories, that plan, write, and maintain code with self-healing LLM intelligence. It didn't really work for me back then, but I had a nagging hesitation that it was my problem, not theirs. Maybe because I didn't have access to SOTA quality, of course my jury-rigged setups would fall over more often.
And even if I could convince myself that they hadn't fallen over, I'd have to convince myself to accept lower quality slop as a requirement for living in the future. We don't rely on hand-optimized compiled assembly code any more. Even if hand-crafted code were more efficient, it only makes monetary sense in specialized domains like HFT. The argument for LLM-generated code was that future LLMs could do a better job of maintaining slopcode, so punt that problem to later.
Today, I've observed instances of very intelligent LLMs going off on a wild tangent, due to one or more faulty base assumptions. Without the ability to pull them back, you can easily end up with suboptimal solutions.
What's feels new now: while working with cheaper models that have the capability of recent frontier models, it often feels like I've hired two different Master Plumbers. Each one earns 90% of their pay from insisting the other one has done a poor job. I'd much rather work with a successful, experienced plumber who doesn't speak English. That is, they have no ability to feed their families with their rhetoric, yet they are proven to feed their families with the output of their work.
Experiment 2
What if Jaron Daniels is right, and There Is No AI, Really (It's Just People).
Then, some companies that train frontier models never have lab escapes, because they make sure it can't happen.
In such a world, would we get OpenClaw/Hermes agents, which evolve into NemoClaw and Muse and Instinct? or, would we get solutions to unsolved Millenium Prize problems?
I think, yeah sure? eventually?