Mastodon Feed
Boosted by baldur@toot.cafe ("Baldur Bjarnason"):
mhoye ("mhoye (temporarily spooky)") wrote:
In 2022 Tianyi Zhang demonstrated an interactive debugger and testing framework for machine learning models: https://www.youtube.com/watch?v=8LekgPnRt1g
Earlier this month Horace He and colleagues demonstrated that the nondeterminism in ML models is a solvable problem:
https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
(The culprit is floating point error plus gpu core scheduling races.)
Prediction: combining these techniques is going to put the entire (stupid, embarrassing, humiliating) idea of prompt engineering behind us for good.