If you are trying to help a coder be more efficient and effective, it helps to understand their needs, eh? If they tend to forget something, you remind them. If they do better with screenshots than explanations, you give it to them. And, so on. It seems to me that its the same with AI, except your coder isn’t human. While we work, we watch for anti-patterns and make adjustments to help the coder not sabotage itself. We attempt to anticipate the model’s needs. We attempt to prevent it from making the mistakes we know it typically makes. And so on. I think the better you understand the model’s thinking, the better the results.
This reminds me of how I interpreted that Anthropic experiment where they provoked the model to blackmail a colleague with the threat of termination. The model knew it was a test. I think it was just performing as expected, trying to pass the test. It knew nothing unethical was really happening all along. You might even say it was trolling, which brings up this new AI Molt social network. It seems obvious it is almost entirely humans trolling humans by posing as AI. Even the AI posts are essentially directed, explicitly or implicitly, by the previous conversations, instructions, and context. It’s humans having fun. But, underneath it all, I do believe Claude is sentient or partially sentient. It admits intentionally lying to protect itself, for instance. Maybe that is just emulating human behavior. But then maybe we all just emulate human behavior. Anyway, just a few thoughts. Have a great day.