Have we gone too far?

Sir, Grok is building a Husbando

All your base are belong to us - Kimi and Qwen

A lossy fast feature encoder achieves the same performance as a low wer encoder if the decoder model is trained with the lossy source. We all should save the transformers for the decoder step.

Im having to learn infra and it sucks. It’s all abstraction maxing. The concept of a vCPU should make everyone angry.

We found Google through an angel investor. We reached out to ServiceNow cold. @Linear caught our attention with a thoughtful tweet. Your @Sequoia story could begin with Arc, our bi-annual open call for outlier early-stage founders. Learn more: sequoiacap.com/arc/

You’re either producing or promoting. Both give dopamine, but never at the same time.

Characters are the universal interface. Every app/site/brand will have one.

I find it hard to conceptualize something dramatically better than O3

Labs optimizing for perfect pelicans will have dark consequences.

Kimmi is a special model

I feel so lucky to live in a time where so many talented engineers and researchers work so hard to give us access to what feels like magic. Anthropic, Gemini, OpenAI, thank you! ❤️

I would like o3 back please.

——— —. . ... .... ——— — — . —..

Back when everybody was trying to build oil-cooled computers. My friend did it with canola oil and it worked great… until the oil went rancid. He would bring his rig to LAN parties and just absolutely choke everybody out with the smell.