i'm not sure if people realize that soon there will be an emergent role for AI-interaction design + finetuning research.. so either AI researchers should learn interaction design, or design people should learn AI research

During Claude 3 training one of my most favorite thing to do was to live blog all the evals on a daily basis and saying we were sota. People got either super excited and it made their day or freaked out due to weird regressions and it ruined their day 🥲

While post-training Claude 3 Haiku, I was most surprised by the expressive power of the model relative to the cost and the velocity of efficiency improvements. A few thoughts:

1/ Haiku is on par with or better than Claude 2 on most benchmarks, yet the cost reduced as much as

Claude Haiku is the best for rapid prototyping as it is super fast and smart. I loved Dune Part 2, so after some iteration I wanted Haiku to explore the world of Arrakis and express it visually:

'Imagine a planet where the very sands beneath your feet hold the key to the future

Ask not what large language models can do for you, ask what you can do for large language models

Another demo! Here I asked Claude 3 Opus to create a visual scrolly to explain backprop and Koglomorov’s complexity concepts for beginners in ML with simple three.js animations. This was done in ~30 mins w/ some back and forth prompting.

What language models can do to

Days and nights we were helping these poetic models to learn more and more, get better and smarter w/ RL yet trying not to trade off undesirable behaviors and I’m so grateful to ever contributing to their creations. I’m super excited for Claude Haiku as it’s often surpassing