There’s extra. To make its use of reinforcement studying as environment friendly as attainable, DeepSeek has additionally developed a brand new algorithm referred to as Group Relative Coverage Optimization (GRPO). It first used GRPO a yr in the past, to construct a mannequin referred to 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 unhealthy. Many present reinforcement-learning strategies require an entire separate mannequin to make this calculation. Within the case of huge language fashions, which means a second mannequin that may very well 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 typical strategy
DeepSeek’s use of reinforcement studying is the principle innovation that the corporate describes in its R1 paper. However DeepSeek is just not the one agency experimenting with this method. Two weeks earlier than R1 dropped, a staff at Microsoft Asia introduced a mannequin referred to as rStar-Math, which was educated in an analogous means. “It has equally big 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 strategies (however on high of, not as a substitute of, human-led steps like supervised fine-tuning and RLHF). And the US agency Hugging Face is racing to duplicate R1 with OpenR1, a clone of DeepSeek’s mannequin that Hugging Face hopes will expose much more of the elements in R1’s particular sauce.
What’s extra, it’s an open secret that high corporations like OpenAI, Google DeepMind, and Anthropic might already be utilizing their very own variations of DeepSeek’s strategy to coach their new era of fashions. “I’m certain 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 referred to as multi-token prediction, the place the mannequin learns to foretell a string of phrases directly as a substitute of one after the other. This coaching is cheaper and seems to spice up accuracy as effectively. “If you consider the way you converse, if you’re midway by means of a sentence, you understand what the remainder of the sentence goes to be,” says Zeiler. “These fashions must be able to that too.”
It has additionally discovered cheaper methods to create giant information units. To coach final yr’s mannequin, DeepSeekMath, it took a free information set referred to 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 information set of math issues by hand. It was additionally simpler: Widespread Crawl consists of much more math than every other specialist math information set that’s obtainable.
And on the {hardware} facet, DeepSeek has discovered new methods to juice previous chips, permitting it to coach top-tier fashions with out coughing up for the newest {hardware} available on the market. Half their innovation comes from straight engineering, says Zeiler: “They positively have some actually, actually good GPU engineers on that staff.”