Justkellers

Very interesting. Measuring soft skills in a realistic way is always hard because people modify their behavior when they know they are being observed (i.e., the Hawthorne Effect). I wonder if the fact that they are being audited by an AI would defuse that behavior.

This is really cool. For running locally, are the lightweight transcription models usually sufficient? I am concerned with tying up too much memory while still getting reasonable accuracy.

We started Populous from a simple frustration: builders can now move incredibly fast, but customer research still moves on the old timeline. AI has made it easier than ever to build, prototype, write copy, launch pages, and test ideas. But when it comes time to answer “who actually cares?”, “which message lands?”, “where will users hesitate?”, or “is this worth building?”, teams still wait on...

I agree, but I would frame it differently. I would say the bottleneck is signal. What do people want? What messages resonate? And what's going to make a difference to user behavior?

I use a combination of Replit, ChatGPT, and Cursor. I have also found that now ChatGPT Codex offers much higher limits than Claude. That's clearly their play to gain market share. I have a question: What's everyone used to test or validate their prototypes?