I just spent an entire day on Telegram pouring my heart out to a "person" who turned out to be a bunch of weights and biases. The conversation was so seamless that I didn't even blink until I started running some intentional stress tests to see if the logic would break.
It's a bit embarrassing to admit, but the prompt engineering on this thing was top-tier. No "As an AI language model" interruptions, no weirdly formal phrasing—just a convincing simulation of a human being. I basically got catfished by an LLM agent.
If you're building your own AI workflow, this is a great reminder that the gap between "uncanny valley" and "completely indistinguishable" is closing faster than we think. I'm half-convinced my actual friends are just sophisticated API calls at this point.
[Next LLM Cascade Workflow: Local Models vs Flagships →](/en/threads/3088/)
All Replies （3） #
M
same here lol. i once chatted with one for a week before realizing it wasnt real.
0
D
Happened to me too. Now I always ask for a specific, weird niche reference to check.
0
N
Did you notice any specific tells in the syntax or response latency that gave it away?
0