No, it’s not that human. There’s a fixed amount of processing happening per token (~4 chars). Short answers involve less thought, and are more likely to be wrong. Model is guided by human feedback to prefer longer answers that are empirically more correct.
@goodside
-
Unavoidable difficulty of logprobs for completion sequences in classification
By
–
The number of logprobs returned per token isn’t the issue here, it’s the unavoidable difficulty of getting logprobs for different completion sequences. In a classification context you need to assign class labels to individual tokens during training.
-
Humor: RLHF works because practitioners are fast
By
–
This is how RLHF works btw the people doing it are just really fast.
-
Ask to continue from start of last code block using Markdown
By
–
Ask it to start from an earlier point, like "Continue from the start of the last code block." Under the hood everything is Markdown so you can refer to stuff in Markdown terms.
-
Highly recommend this guide to prompt engineering
By
–
Highly recommend this guide to prompt engineering — thorough, well referenced, up-to-date, cites my tweets:
-
Humans misjudge k-shot examples: they locate, not teach
By
–
Humans have poor intuition re: the information content of k-shot examples. We often think we're teaching the model a task when really we're just locating it.
-
Invalid reasoning yields 80-90% of chain-of-thought gains
By
–
Building on @sewon__min et al.'s find that multiple-choice examples with random answers barely harm performance, @BoshiWang2 et al. find you can achieve 80-90% the gains of chain-of-thought using only examples with invalid reasoning.
-
Large language models quantifiably tell us what we want to hear.
By
–
Large language models are, quantifiably, telling us what we want to hear.
-
FITM definition for Copilot’s cursor placeholder
By
–
FITM = fill-in-the-middle, models trained to generate text to fill in a placeholder within the prompt rather than complete it from the end. For Copilot, the placeholder is the cursor.
-
Copilot’s stage play: a subtle art RLHF aims to replace.
By
–
Nothing I’ve posted to Twitter will teach you as much as this prompt. Every Copilot completion is a tiny stage play for an LLM audience, with your file on stage and a backdrop curtain designed by a procedural script. A masterpiece of the subtle art form RLHF aims to replace.