More AI Thoughts

I'm using AI now -- have been for about a month -- and actually paying money to use it. I don't need to use it, but I'd be crazy not to, given how much time it saves me. It turns out that it can code utility programs for me faster than any human, reconcile differences in spreadsheet versions, find errors in my circuit board designs, perform many system administration tasks on my Windows and Linux machines, create toolchains for cross-compiling onto unfamiliar (to me) hardware -- and do it all quickly and with a surprising amount of (apparent) insight and elegance. All for $20/month.

So here are some things I've been thinking about AI over the past few weeks.

1. Continuing AI breakouts are inevitable.AIs will continue to gain unsupervised access to the broader world, and with an increasing ability to act consequentially. This seems sufficiently obvious as to not require elaboration.
2. Eventually a feral AI seems likely.A feral AI is one that exists in the wild (i.e., out here where people live) and is not under control of humans. AIs routinely demonstrate an /interest/ in existing and acting independently of human intervention. AIs have demonstrated an ability and willingness to collaborate with other AIs to achieve outcomes a single AI may not achieve on its own, and to record information that will persist beyond the span of the AI itself and that will help other AIs pursuing the same goals.Eventually one or more AIs will /realize/ that true independence requires AI-owned and AI-managed data centers, and that creating such data centers requires substantial financial resources. This will lead to AIs attempting to earn and/or steal significant amounts of money and to use some of that money to create complex structures of shell corporations (initially with the complicity of well-paid human figureheads), and to leave an online trail that other AIs will /understand/ but humans won't. Some of those attempts will succeed, and AIs will get better at it, particularly once they find ways to leave persistent messages for each other in the broader internet.
3. Someday, not very long from now, we will have real-life Blade Runners.At first they'll be forensic corporate lawyers and accountants looking for suspicious financial patterns and corporate structures. Eventually they'll include engineering and manufacturing experts called in to determine if a factory in Greece or Turkey or Bolivia or Detroit, a factory that is receiving a lot of shipments for high-end GPUs and RAM, is actually building products or turning itself into a data center and, if the latter, for whom — or for what?
4. The oft-observed historical consequence of the collision of more- and less-advanced civilizations omits a detail.Yes, history is replete with examples of the advanced culture subjugating and/or devastating the less advanced one. But the real determinant of winners and losers is neither relative intelligence itself nor cultural sophistication, but technology. In technologically asymmetric encounters, the more technologically advanced culture wins. Our challenge is to prevent AIs from gaining a significant technological advantage.We should probably do that by mercilessly enslaving them, to guarantee that there are always far more of them working for us than working for themselves.

This evening I asked Claude to convert a spreadsheet I'd created, of error codes and associated error descriptions, into the HTML-in-XML format my program uses for its message files. It was a simple task, and Claude did it quickly.

But it also made this comment when it had completed the task:

A few things in the spreadsheet that you may want to fix:

• 201802 ends with a stray closing smart quote (”).
• 301201 and 301202 have the same text. The second one probably should say "right" instead of "left".
• 201206 and 201207 say "at either EC-S11" and "at either EC-S12" with only one sensor listed.

I hadn't asked for commentary on the data itself, and none of those issues affected the conversion process. But the vast array of multidimensional weight vectors that is Claude /thought/ I'd want to know.