ื—ื‘ืจื• ืืช OpenTG ืœ-ChatGPT ืื• ืœ-Claude ืชื•ืš 60 ืฉื ื™ื•ืช

ื ืกื• ืื•ืชื•

Two days ago we put ๐Ÿ‘/๐Ÿ‘Ž under every character reply in @luciddreams.

platov.ai

platov.ai

@platov_ai

Founder of @luciddreams

33,300 ืžื ื•ื™ื™ื
ืคืชื— ื‘ื˜ืœื’ืจื
Two days ago we put ๐Ÿ‘/๐Ÿ‘Ž under every character reply in @luciddreams. Within hours, our community had a theory about how they work.

"It does take them into account," one user explained to another. "Rate the reply and it will move in that direction. How long it remembers, who knows. Probably you have to rate every new message." Someone else called it a brilliant idea: finally a way to tell the bot what you like and what you don't.

The funny part: the buttons don't change your chat. Every rating comes to us, and we use them to compare models and decide which one gets better.

Researchers have a name for this. Folk theories are the stories people build to explain how a system works when nobody showed them the inside. If you've ever "trained" your feed by scrolling past things faster, you have one. Same if you've pressed the crosswalk button five times so the light would change sooner. In a lot of cities that button isn't even wired to anything ๐Ÿ˜

The obvious reaction is to correct people: pin a message, explain how it really works. I think that's the wrong move.

Urban planners know this as the desire path. You lay a nice paved road, and a month later there's a dirt trail cutting across the lawn, because that's where people actually walk. You can put up a fence and a sign. Or you can pave the trail.

A folk theory is a desire path for a product. It shows you what people expect the thing to do, in their own words, before you even asked. Our users just told us they want to steer their character with a single tap. That's a better spec than anything I would have written alone.

So we're paving this one. We're now working out how to make your ๐Ÿ‘/๐Ÿ‘Ž actually shape the character in your chat, and we want it in the next update. The theory our community came up with is about to become true ๐Ÿ›ค

This is why I want everything in production as fast as possible. Likes and dislikes are the easy part. What I really want to see is how people understand a feature, and no mockup or meeting gives you that.

Every guess you make about users without checking carries some error. Sit on an idea long enough and you start stacking new guesses on top of old ones, and the errors compound. Ship, watch what people actually do, update your picture of the world, and make the next call on solid ground. Then do it again ๐ŸŽฏ
ืคืชื— ืืช ื”ืคื•ืกื˜ ื‘ื˜ืœื’ืจื