Recent discussions in AI research highlight earlier claims about a new scaling law referred to as 'inference-time search,' proposed by Google and UC Berkeley. This law focuses on generating multiple answers for a query and selecting the best one, potentially enhancing the performance of existing AI models. While traditional scaling laws like pre-training still dominate, the emergence of post-training and test-time scaling indicates an evolving landscape in AI development. Experts express skepticism about the touted capabilities of this new law and its implications for future AI performance.
"Our paper focuses on this search axis and its scaling trends. For example, by just randomly sampling 200 responses and self-verifying, Gemini 1.5 (an ancient early 2024 model!) beats o1-preview and approaches o1." - Eric Zhao
"Pre-training hasn't gone away, but two additional scaling laws, post-training scaling and test-time scaling, have emerged to complement it."
"The researchers claim it can boost the performance of a year-old model, like Google's Gemini 1.5 Pro, to a level that surpasses OpenAI's o1-preview 'reasoning' model on science and math benchmarks."
Collection
[
|
...
]