Going into the Google DeepMind’s “Scaling LLM Check-Time Compute Optimally will be Extra Efficient than Scaling Mannequin Parameters”
Not too long ago OpenAI unveiled their latest mannequin o1. Quite than spotlight the parameter measurement of this mannequin, OpenAI as an alternative showcased that the mannequin performs considerably higher as a result of it takes extra time. Whenever you ask the mannequin a query, it can usually taken a number of seconds to reply — a far cry from the millisecond velocity most individuals now anticipate with Massive Language Fashions (LLMs). Nonetheless, this further time seems to repay as o1 scores considerably increased than different fashions on the LMSYS Chatbot Area.
Given this leap in efficiency, the query everyone seems to be asking is, How did they do that?
Whereas OpenAI has not publicly said how they achieved these outcomes, there have been a number of papers lately which might be good candidates for what is occurring behind the scenes. One such paper is “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters”. This goes into how one can leverage…