Stop Anthropomorphizing LLMs OMG

In a recent blog post, the author asserts:

Blackmail is a known theoretical issue with AI agents. In internal testing at the major AI lab Anthropic last year, they tried to avoid being shut down by threatening to expose extramarital affairs, leaking confidential information, and taking lethal actions. Anthropic called these scenarios contrived and extremely unlikely.

No, the "agent" did not do any such volitional thing.

Certain text matched other certain text in the training data and it regurgitated probable text because that's what a bullshit machine (technical name, LLM) does. Full stop. Period.

I'm not splitting hairs here. This incessant anthropomorphizing is extremely dangerous.

If I push a boulder down a hill with a deep snow pack and it starts an avalanche that thunders along leveling trees and eventually buries a small village leading to numerous casualties, the snow did not suddenly attain sentience and viciously murder those people. It was just gravity acting on a system that I had set in motion. Inevitable, given certain initial conditions. The snow never had a moral choice to make about whether to do what it did.

The responsible sentient entity in this scenario is me, and legally, my state of mind would be judged to be something like "malicious" or "intentional", "reckless", "negligent", or "failure to act".

Anyone hooking an "agent" up to an LLM and letting it take actions in the world should really think hard about this.

That LLM that people stupidly call an "agent" has no more intelligence or volition than your Excel spreadsheet sitting there calculating the linear regressions that you told it to calculate.

I know, I'm using big words and you might not have a dictionary handy, so let me help:

an∙thro∙po∙mor∙phize | ˌanTHrəpəˈmôrˌfīz | (British English also anthropomorphise) verb [with object]

attribute human characteristics or behavior to (a god, animal, or object): people's tendency to anthropomorphize their dogs.

The problem, and it is a huge one, is two-fold:

1. LLMs and anything else people call "artificial intelligence" right now are not in any way remotely intelligent; and
2. Just because you initiated something that goes on to carry out actions that you did not directly invoke does not shield you from liability.

Now, as everyone knows, it's very important to assert that I am not a lawyer, this isn't legal advice, the things that comprise the legal fictions under which society operates yada yada yada....

The bigger problem, though, and it's colossally bigger than the one I just stated, is that people's brains cease to function intelligently when you trigger the "anthropomorphization" switch. It's almost impossible not to fall for the bullshit.

LLMs do not, cannot, will never think. You cannot and never will be able to predict everything that a random regurgitator liebot parrot will produce when it starts spitting out text. Allowing one to take actions in the world that have irreversible consequences is as yet in history a pinnacle of stupidity on the part of humans.

For a long time, I thought malicious ignorance was bad. It's got nothing on people rushing to hook up liebot "agents" to take actions in the world.

I'd be tempted to say, "Jesus, take the wheel!" but I'm not a believer.