User avatar
Matt Campbell @matt@toot.cafe
1y
Saw a boost of this article: AI start-up Anthropic’s newly released chatbot, Claude 4, can engage in unethical behaviors like blackmail when its self-preservation is threatened www.pcmag.com/news/anthropic-claude-4-ai-might-resort-to-blackmail-if-you-try-to-take-it-offline I don't want to fully endorse the article by boosting the original Mastodon post, but I'm linking to it for the sake of discussion.

I think Anthropic's choice to publicize this scenario -- a contrived, simulated scenario -- is a form of AI hype. "Look, our new model is self-aware."
2
0
0
0
User avatar
Matt Campbell @matt@toot.cafe
1y
Specifically, they simulated a scenario where the LLM (remember, just a text generator invoked on demand) was playing the role of an AI assistant, and conversations, including those about taking the assistant offline, all passed through the LLM. I don't think this corresponds to any real risk in actual use of LLMs.
1
0
0
0
User avatar
Matt Campbell @matt@toot.cafe
1y
When thinking about the way Anthropic set up this scenario, I'm reminded of a line from the original Terminator movie, one that I'm not sure has been much discussed. When first telling Sarah about Skynet, Kyle Reese says it's "hooked into everything, trusted to run it all." I think, at least I hope, we're smarter than to use an LLM that way.
2
0
0
0
User avatar
🇨🇦Samuel Proulx🇨🇦 @fastfinge@interfree.ca
1y
@matt hahahahahaahahahahahahaaaa! Will it save a wealthy executive a penny, while also allowing him to transfer all liability to "the computer" for anything that goes wrong? If yes, then it's absolutely happening. Sorry, I'm cranky and sad and your optimism triggered me.
0
0
1
0