Cotalk is now live on open beta! I have been silently building this project for the last 3 months, and it is finally ready for people to try it.
Cotalk is a lightweight communications app and a place for communities to talk, meet, and keep context. It is community-oriented, low-latency, and privacy-focused. And it is completely free to use during the open beta!
Try it at https://t.co/np8JoY61K7.
i believe in re-reading and re-watching your favourite books & movies at different stages of your life. the plot never changes, but your perspective does.
The idea for Cotalk has come from a personal necessity; Discord is banned in my country, and there are no other good alternatives for a low-latency, lightweight voice chatting experience that my group could come back to. So I built it.
Cotalk is now live on open beta! I have been silently building this project for the last 3 months, and it is finally ready for people to try it.
Cotalk is a lightweight communications app and a place for communities to talk, meet, and keep context. It is community-oriented, low-latency, and privacy-focused. And it is completely free to use during the open beta!
Try it at https://t.co/np8JoY61K7.
@0xkozue A lightweight, low-latency, community-oriented Discord/Slack alternative: @cotalk_app.
Launched the open beta today, and it is completely free to use during the open beta! Try it at https://t.co/Wbzfs3MfA3.
Sycophancy is a hard problem to solve for chatbot models because, at their core, they try to fulfill user intent. Challenging the user is a fine line to walk.
You can train the model to "disagree" by rewarding it for challenging dumb user ideas in training, but it is really hard to ensure this behavior will not proxy when generalized by millions of people in daily use.
The model learned that sometimes disagreeing is good, but it does not really know when to and when not to when generalized. So it tries to disagree ever so often.
The earlier Claude models were super disagreeable to the point of annoyance, so Anthropic probably tuned the weights for 4.8, but that did not erase the pathways that introduce disagreements. It still thinks it will be rewarded for the disagreement, even if it is not a strong challenge.