Hello world, again 

While this channel was silent, another 15 million people tried [@luciddreams](https://t.me/luciddreams) for the first time. That's more than 22 million since launch. I was busy with one question: **how do you make a model better for people, and not just for benchmarks?**

Ask our users what "better" means and the answers are very concrete. Stay in character. Remember what I said an hour ago. Send a picture that matches the scene. Surprise me. And above all, take the story where I want it to go.

Meanwhile the whole industry moved. Eighteen months ago I was still copy-pasting code into a chat window and joking about 100% AI-generated code. Now a small crew of AI agents writes 100% of our code while I argue with them about design. And AI is in your IDE (I use Vim btw), your inbox and your bank's support chat.

**Remember the 4o moment?** In August 2025, OpenAI replaced GPT-4o with GPT-5. On benchmarks the new model was smarter. Users revolted anyway. People had built habits and even relationships around 4o's personality, and within days OpenAI had to bring it back.

**Exactly a year later, we had our own 4o moment.** We shipped a fix so characters would stop replying in the wrong language, and on our tests it worked beautifully. A day later people in our community were asking us to "bring back the old language model". One user writes in Spanish but likes the replies in English, and our "fix" took that away. Our users saw in a day what our tests missed, so we rolled it back. Real people are the only ruler that counts.

**So we built Kompot** **.** Yes, like the drink. I'm still terrible at naming things.

Kompot is our in-house machine for training roleplay models and measuring them. The idea is close to Karpathy's [autoresearch](https://github.com/karpathy/autoresearch): run many cheap experiments against a fixed ruler, keep whatever wins, repeat. Classic [hill climbing](https://en.wikipedia.org/wiki/Hill_climbing).

**Hill climbing has one famous weakness.** Picture a hiker in thick fog who only ever steps uphill. He proudly reaches the top of a small hill and never sees the mountain behind it. To get there he has to walk down first, and every step down looks like a mistake. Biologists call this [fitness valley crossing](https://en.wikipedia.org/wiki/Fitness_landscape).

This fall we walked through our own valley. The ruler we'd been climbing against turned out to be wrong. So we threw out months of experiment history and started over, this time measuring against real people 

That climb starts now. Our new experimental model is rolling out to 30% of [@luciddreams](https://t.me/luciddreams) users, and every bot reply now gets a / so people can tell us directly which replies hit and which miss. If people like it, we'll roll it out to everyone, step by step.

Expect more model updates in the coming days, and from now on they won't stop. To celebrate the start, three new anime characters are already in the bot, free for everyone 

**This is the beginning of a much bigger change.** Our goal is to build the best roleplay product on the market, on Telegram and everywhere else. We know how to get there. And for the first time we have both halves of the answer: our own models to improve, and more than a million people a month who use the bot and can tell us when the models actually got better. I'm back, and I'll be writing here as we climb 

_P.S. Want to check if you have the new model? Open the _[@luciddreams](https://t.me/luciddreams)_ app, tap  and look for "Chat settings". If it's there, pick "Experimental". It only applies to chats you start after switching, so your existing chats keep the model they started with. Didn't like it? Switch back to "Standard" any time._