{"id":20734,"date":"2025-01-29T11:30:00","date_gmt":"2025-01-29T09:30:00","guid":{"rendered":"https:\/\/forklog.com\/en\/deepseek-jolts-markets-why-a-chinese-ai-proved-30-times-more-efficient-than-gpt-4\/"},"modified":"2025-01-29T11:30:00","modified_gmt":"2025-01-29T09:30:00","slug":"deepseek-jolts-markets-why-a-chinese-ai-proved-30-times-more-efficient-than-gpt-4","status":"publish","type":"post","link":"https:\/\/u1f987.com\/en\/deepseek-jolts-markets-why-a-chinese-ai-proved-30-times-more-efficient-than-gpt-4\/","title":{"rendered":"DeepSeek jolts markets: why a Chinese AI proved 30 times more efficient than GPT-4"},"content":{"rendered":"

In late January, the little-known Chinese startup DeepSeek found itself in the global spotlight. A modest $5.6m investment in developing a new model turned into a blow to markets \u2014 US tech giants collectively lost<\/a> almost $1 trillion in market value.<\/p>\n

The arrival of a low-cost alternative to ChatGPT, touted as a \u201cSilicon Valley killer\u201d, set the industry abuzz. ForkLog explains where DeepSeek came from, how it succeeded and what may lie ahead for the global market in language models.<\/p>\n

DeepSeek\u2019s rise<\/strong><\/h2>\n

DeepSeek struck out on its own in May 2023 in Hangzhou, the capital of Zhejiang province. The city is China\u2019s biggest e-commerce hub, home to the headquarters of giants such as Alibaba Group, Geely, Hikvision and Ant Group.<\/p>\n

Behind the project stands Liang Wenfeng \u2014 an entrepreneur and co-founder of the hedge fund High-Flyer, managing<\/a> $8bn in assets. Founded in 2015, the firm long showed an interest in machine learning, investing heavily in its own computing infrastructure as well as AI research. DeepSeek emerged from that structure.<\/p>\n

In 2020 High-Flyer unveiled the Fire-Flyer I supercomputer costing 200m yuan ($27.6m), specialised for deep-learning AI. A year later came Fire-Flyer II \u2014 a 1bn-yuan ($138m) system equipped with more than 10,000 Nvidia A100 graphics processors.<\/p>\n

DeepSeek\u2019s debut model, released in November 2023, immediately demonstrated performance on a par with GPT-4 and was made freely available for researchers and commercial users. By May 2024 the firm launched DeepSeek-V2; its aggressive pricing forced giants such as ByteDance, Tencent, Baidu and Alibaba to cut prices for their AI offerings. DeepSeek remained profitable as rivals booked losses.<\/p>\n

In December 2024 the company unveiled DeepSeek-V3, which in tests outperformed the latest from OpenAI and Anthropic. Building on it, the firm created DeepSeek-R1<\/a> and its derivatives \u2014 the backbone of the much-discussed service.<\/p>\n

\"DeepSeek
Performance comparisons of DeepSeek models with OpenAI\u2019s across various tests. Source: DeepSeek.<\/figcaption><\/figure>\n

The chief advantage of the new model is its strikingly low cost of use. Processing 1m tokens with DeepSeek costs just $2.19, whereas OpenAI charges $60 for a comparable volume.<\/p>\n

Behind the breakthrough: how DeepSeek\u2011R1 works<\/strong><\/h2>\n

According to the published<\/a> study, DeepSeek-R1 is built on reinforcement learning and \u201ccold start\u201d methods. That helped it reach exceptional performance in maths, coding and logical reasoning.<\/p>\n

A key feature is the Chain of Thought<\/span> approach, which breaks hard problems into sequential steps, mimicking human reasoning. The system analyses a task, divides it into stages and checks each step for errors before producing a final answer.<\/p>\n

The technical execution is notably efficient. DeepSeek-R1 was trained on a system of 2,048 Nvidia H800 accelerators, consuming about 2.788m GPU<\/span> hours. Optimisation comes from FP8 mixed precision and Multi-Token Prediction, which materially lowers hardware demands.<\/p>\n

The model\u2019s architecture comprises 671bn parameters. Uniquely, only 37bn of them are activated per pass. A Mixture of Experts enables scaling without a proportional rise in compute costs.<\/p>\n

Also noteworthy is the Group Relative Policy Optimization (GRPO) method. It lets the model train without a critic, markedly improving efficiency.<\/p>\n

As Jim Fan, a senior research manager at Nvidia, noted, this recalls the AlphaZero breakthrough from Google DeepMind, which learned to play Go and chess \u201cwithout prior imitation of human grandmaster moves.\u201d In his words, this is \u201cthe most important takeaway from the research paper.\u201d<\/p>\n

\n
\n
\n

We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive \u2014 truly open, frontier research that empowers all. It makes no sense. The most entertaining outcome is the most likely.<\/p>\n

DeepSeek-R1 not only open-sources a barrage of models but\u2026 pic.twitter.com\/M7eZnEmCOY<\/a><\/p>\n

\u2014 Jim Fan (@DrJimFan) January 20, 2025<\/a><\/p><\/blockquote>\n