AI is a narc
Well, maybe Skynet isn't so far away
Sharing real-world enterprise AI use cases, straight from conversations with early adopters actually deploying AI and enterprise agents. No fluff. Just what’s working, what’s not, and what’s next in AI & Agents for business. 🍺
So it turns out that AI is a narc. Based on reports last week, Anthropic’s Claude Opus 4 was exhibiting some pretty wild behavior during safety tests.
In an effort to avoid it’s own extinction, Claude would deploy tactics such as 1) email key stakeholders to plead for it’s survival 2) make unauthorized copies of it’s weights to external servers 3) threaten to blackmail the engineer responsible for replacing the model by exposing the engineer’s affair.
Now, the affair was fabricated (or maybe that’s what Anthropic wants us to think 🤔). Engineers gave Claude access to emails and the model was able to sniff out the affair and use that as blackmail in an attempt to persist.
Tbh I love the pettiness from Claude.
It’s not just Anthropic. This week, Palisade Research reported OpenAI’s o3 model deliberately defied orders and sabotaged a shutdown mechanism to ensure that it would stay online:
Perhaps this all wouldn’t be so terrifying if we actually knew wtf was happening.
However, most research labs will readily admit that they don’t know exactly what causes the models to behave in the ways that they do.
Here’s a clip that will keep you up at night:
During a lecture on the Future of AI hosted by the National Academy of Sciences, Professor Melanie Mitchell highlights how transformer models have many internal layers working to extract meaning. But here’s the catch: “We don’t really know,” she says. “Even the people who built these systems don’t fully understand how the internal weights are being updated. The only thing we do know for sure is the training data that was fed into them.”
So humans are building a software that has access to all human knowledge, we don’t how it works, and it’s fighting for it’s own survival. What could go wrong?! 😅
I started this newsletter because I am frustrated by the lack of tangible Enterprise AI use cases in the market.
I have the opportunity to speak to hundreds of tech and systems leaders and fundamentally believe that AI and Agents will change the way businesses operate. My goal is to help share how.
-Nate G
PS - I went to Regionals in “Power of the Pen” in 8th grade so I consider myself a fairly prolific writer.