There’s extra. To make its use of reinforcement studying as environment friendly as potential, DeepSeek has additionally developed a brand new algorithm known as Group Relative Coverage Optimization (GRPO). It first used GRPO a yr in the past, to construct a mannequin known as DeepSeekMath.
We’ll skip the particulars—you simply must know that reinforcement studying entails calculating a rating to find out whether or not a possible transfer is sweet or dangerous. Many current reinforcement-learning methods require an entire separate mannequin to make this calculation. Within the case of enormous language fashions, meaning a second mannequin that could possibly be as costly to construct and run as the primary. As a substitute of utilizing a second mannequin to foretell a rating, GRPO simply makes an informed guess. It’s low cost, however nonetheless correct sufficient to work.
A standard method
DeepSeek’s use of reinforcement studying is the primary innovation that the corporate describes in its R1 paper. However DeepSeek will not be the one agency experimenting with this system. Two weeks earlier than R1 dropped, a staff at Microsoft Asia introduced a mannequin known as rStar-Math, which was educated in an identical means. “It has equally enormous leaps in efficiency,” says Matt Zeiler, founder and CEO of the AI agency Clarifai.
AI2’s Tulu was additionally constructed utilizing environment friendly reinforcement-learning methods (however on prime of, not as an alternative of, human-led steps like supervised fine-tuning and RLHF). And the US agency Hugging Face is racing to copy R1 with OpenR1, a clone of DeepSeek’s mannequin that Hugging Face hopes will expose much more of the substances in R1’s particular sauce.
What’s extra, it’s an open secret that prime companies like OpenAI, Google DeepMind, and Anthropic could already be utilizing their very own variations of DeepSeek’s method to coach their new technology of fashions. “I’m positive they’re doing virtually the very same factor, however they’ll have their very own taste of it,” says Zeiler.
However DeepSeek has multiple trick up its sleeve. It educated its base mannequin V3 to do one thing known as multi-token prediction, the place the mannequin learns to foretell a string of phrases without delay as an alternative of one by one. This coaching is cheaper and seems to spice up accuracy as nicely. “If you consider the way you communicate, while you’re midway by a sentence, you already know what the remainder of the sentence goes to be,” says Zeiler. “These fashions ought to be able to that too.”
It has additionally discovered cheaper methods to create giant knowledge units. To coach final yr’s mannequin, DeepSeekMath, it took a free knowledge set known as Widespread Crawl—an enormous variety of paperwork scraped from the web—and used an automatic course of to extract simply the paperwork that included math issues. This was far cheaper than constructing a brand new knowledge set of math issues by hand. It was additionally more practical: Widespread Crawl contains much more math than another specialist math knowledge set that’s out there.
And on the {hardware} aspect, DeepSeek has discovered new methods to juice outdated chips, permitting it to coach top-tier fashions with out coughing up for the most recent {hardware} in the marketplace. Half their innovation comes from straight engineering, says Zeiler: “They undoubtedly have some actually, actually good GPU engineers on that staff.”