Patel says DeepSeek r one zero shows reasoning behaviors emerge naturally from RL on verifiable rewards without human data.
“So it's the remarkable thing about these reasoning results, and especially the DeepSeek r one paper, is this result that they call DeepSeek r one zero, which is they took one of these pretrained models.”
Patel explains reasoning models dramatically increase serving costs due to long output context and memory constraints.
“your memory usage is going way up with these reasoning models, and you still have a lot of users. So effectively, the cost to serve multiplies by a ton.”
Patel explains reasoning models dramatically increase memory usage and reduce batch size, multiplying serving costs.
“So your your memory usage is going way up with these reasoning models, and you still have a lot of users.”
Patel reports reasoning models improved college math test scores from 4% to 80% in six months.
“Basically, our college level math tests were initially 4% for before this innovation was happened to now they're scoring 80%, the best leading edge models. Right? And this has just happened in six months.”
Patel says reasoning capability allowing models to think before answering dramatically increased capabilities in six months.
“Right? And this has just happened in six months. And the same applies to various certain coding tasks. The same applies to yeah.”
Patel reports alternative data shows reasoning models have low actual API usage despite availability.
“So this is something that I've found very interesting is that we've been trying to build a lot of alternative data sources for token usage, who's using tokens, what models, where, etcetera, why, and it's very clear that people aren't actually using the reasoning models that much in API.”