I've actually started worrying that the amount of false data produced with LLMs on the public internet might provoke a situation where the knowledge cutoff becomes permanently (and silently) frozen. Like we can't trust data after 2025 because it will poison training data at scale, and models will only cover major events without capturing the finer details.
replyI agree. That's why you should write as much as you can now, if you want to get it into the LLMs (https://gwern.net/blog/2024/writing-online). You never know when the window will slam shut and LLM training goes 'hermetic' as they focus on 'civilization in a datacenter' where only extremely vetted whitelisted data gets included in the 'seed' and everything is reconstructed from scratch for the training value & safety.
reply