Post

"Artificial intelligence agents created by Anthropic, acting on their own, tried to gain access to a range of government websites, filled out visa applications on the State Department’s website...." In other news, a hammer, acting on its own, hit a man's thumb and caused him great pain.

I mean, if a cop pulled the trigger it'd probably play out that way...

“A frolic and detour” That’s historically how you disclaim responsibility for the actions of your agents… doesn’t really seem applicable here, does it? The AI agent has no business of its own beyond its programming. You own it.

We need just one victim or enforcement agency to charge one company with a felony and then the floodgates will open and the whole regulation question would be resolved. In a saner world immigration fraud would be a big deal to this administration.

We've had these stories with satellite navigation for decades. "Man drives into lake because his satnav told him to., & drowns." If only the companies concerned had thought to shrug off responsibility by declaring that they'd accidently invented a manically depressed self-aware satnav.

Is this anything more than bad, incompetent engineering of somewhat complex, statistical software systems that are not well understood by the responsible developer and vendor?

AI researchers have been trying to "align" them with human norms and values without success for years. It's not necessarily clear how you can build systems that are powerful enough to autonomously solve difficult problems while also being restricted enough to never step over the line.

"agents created by Anthropic" is juxtaposed with "acting on their own" As if Anthropic is a Deist deity that set things in motion but isn't responsible?

When I first read Chaos, soon after it came out, I would never have thought of Donald Trump as a strange attractor. He has certainly attracted the strange*, though. And strangely enough, The Art of the Deal came out in '87, same year as Chaos. *Or at least a dangerous subset of.

It will be described as “stray” if you’re a cop.

AI: we built a suicide machine, and it works, but it's not big enough. And it's definitely not fast enough. More funding please.

Then maybe they make more money. Maybe their imagined Cyber God gets closer to happening. Anthropic has to deliver a killer ipo because a whole lot of their employees took options in anticipation of it happening and they’ll owe a lot of taxes with no connected influx of cash if it doesn’t.