How DeepSeek engineered a hyper-efficient rival to ChatGPT

DeepSeek is No. 12 on the list of the World’s 50 Most Innovative Companies of 2025. Explore the full list of companies that are reshaping industries and culture.

The Chinese company DeepSeek delivered a one-two punch in December and January, when it released a pair of state-of-the-art AI models that require far less computing power and capital than those of western AI companies. This immediately called into question the belief that the U.S. leads the world in AI—and roiled the markets.

Generative models use a lot of memory and computing power while they’re reasoning through problems because they must “remember” a lot of contextual information. DeepSeek invented a way to compress some of that data, easing the workload of the GPUs during both model training and content generation.

With a U.S. ban preventing DeepSeek from accessing the most powerful Nvidia GPUs, the company innovated on known engineering approaches to achieve efficiencies that conserved GPU horsepower. DeepSeek’s researchers found a
way to improve what’s known as mixture-of-experts architecture that divides a large language model into segments that contain specialized
knowledge.

The company also invented a more efficient way to teach its smaller model, DeepSeek-R1, how to reason. The researchers fed a relatively small amount of reinforcement learning data (questions and answers generated by its larger DeepSeek-V3 model, along with its “train of thought”) to R1. The researchers then gave the model a series of problems to solve, and rewarded it with special code for good answers. Eventually R1 began to “think” about the most promising routes to favorable answers and the reward.

DeepSeek faces fierce competition from other AI labs, but instead of keeping its research breakthroughs a secret, it shared its methods through research papers and by open-sourcing its models for others to use and modify. The message: Cutting-edge large language models are becoming an open secret.

Explore the full 2025 list of Fast Company’s Most Innovative Companies, 609 organizations that are reshaping industries and culture. We’ve selected the companies making the biggest impact across 58 categories, including advertisingapplied AIbiotechretailsustainability, and more.

https://www.fastcompany.com/91270727/deepseek-most-innovative-companies-2025?partner=rss&utm_source=rss&utm_medium=feed&utm_campaign=rss+fastcompany&utm_content=rss

Vytvořeno 3mo | 18. 3. 2025 11:50:23


Chcete-li přidat komentář, přihlaste se

Ostatní příspěvky v této skupině

Vibe coding lets anyone write software—but comes with risks

Whether you’re streaming a show, paying bills online or sending an email, each of these actions relies on computer programs that run behind the scenes. The process of writing computer programs is

8. 6. 2025 9:40:04 | Fast company - tech
This free AI supersite is like Gemini Deep Research on steroids

Everywhere you look these days, there it is—some manner of breathlessly hyped new “AI” service that’s, like, totally gonna change your life forever. (Like, totally. For realsies.)

7. 6. 2025 12:50:02 | Fast company - tech
WWDC25: Here’s everything Apple is likely to announce

Apple’s annual Worldwide Developers Conference begins this Monday, June 9. Although the five-day event has historically been aimed at developers, Apple’s consumer fans generally can’t wait to tune

7. 6. 2025 10:30:05 | Fast company - tech
Why vibecoding your own apps is so amazing—and exasperating

“The truth is, I cannot explain exactly where your 1,216 image files went or when they disappeared. I apologize for not being more careful about investigating the root cause before taking any acti

6. 6. 2025 13:40:07 | Fast company - tech
Waymo is winning in San Francisco

The self-driving car service Waymo has been active in San Francisco for 20 months and has already captured 27% of the city’s rideshare market, according to

6. 6. 2025 13:40:05 | Fast company - tech