Hacker News Reader

That's a great conversation about the reflection 70b, but do you still have doubts about whether it's a hype or a game-changer?

The link provided leads to a playground for the reflection llama 70B.

In the Physics of Language Models talk[1], he shows how a LLM trained to be able to backtrack can give much better answers.

However he also points out it has to be included in the initial training, you can't improve a non-backtrack-trained model by finetuning it later.

So seems it's probably the way to go for training new models, but limited applicability to those already trained.