Wan Weigang on DeepSeek, with Robin’s Commentary

Verification notes (2026-10-28): 1. R1 was released on 2025-01-20 under the MIT license, with distillation permitted. Its API output price was US$2.19 per million tokens, about 4% of o1's US$60. Sources: https://www.deepseek.com/en/news/deepseek-r1/ https://venturebeat.com/business/deepseek-r1-is-a-boon-for-enterprises-making-ai-apps-cheaper-easier-to-build-and-more-innovative 2. Jim Fan's comment about a non-US company carrying on OpenAI's original mission: his own post of 2025-01-20 is consistent in substance. Source: https://www.linkedin.com/posts/drjimfan_we-are-living-in-a-timeline-where-a-non-us-activity-7287125475280265217-dUT3 3. The Stargate Project was announced on 2025-01-21. Source: https://openai.com/index/announcing-the-stargate-project/ 4. R1's model weights were released under the MIT license (open-weight), alongside a paper. The training data and full code were not released. It has 671 billion total parameters and activates 37 billion per use, according to the official model card, matching the figures in the post. Kimi k1.5 was also announced on R1's release date, 2025-01-20, and claimed o1-level performance. Claims of being the “only” one or the “first paper” therefore do not hold; the wording here has been revised accordingly. This post reflects views held in January 2025. Sources: https://github.com/deepseek-ai/DeepSeek-R1 https://www.iss.europa.eu/publications/briefs/challenging-us-dominance-chinas-deepseek-model-and-pluralisation-ai-development https://www.deeplearning.ai/the-batch/how-deepseek-r1-and-kimi-k1-5-use-reinforcement-learning-to-improve-reasoning https://arxiv.org/html/2501.12948v1 5. The claim “GPT-4b mini…could add ten years to life expectancy” is incorrect. The model developed by OpenAI and Retro Biosciences is GPT-4b micro, for protein engineering. The reported result was an approximately fiftyfold improvement across multiple laboratory markers for engineered variants of Yamanaka factors, not an additional ten years of life. Sources: https://openai.com/index/accelerating-life-sciences-research-with-retro-biosciences/ https://www.technologyreview.com/2025/01/17/1110086/openai-has-created-an-ai-model-for-longevity-science/ ---- Wan Weigang on DeepSeek On January 20, DeepSeek released its latest large model with reasoning capabilities, DeepSeek R1. You can now use it free on DeepSeek's website— http://chat.deepseek.com DeepSeek R1— - Achieves reasoning capabilities comparable to, or at least approaching, o1, according to benchmarks published by DeepSeek; - Reaches that level with far fewer resources, making it extremely inexpensive; - Makes its model weights publicly available under the MIT license; - Comes with a paper detailing all the steps and techniques used in training—whereas OpenAI has so far kept o1's algorithms and training methods secret. (Commentators can look online for flaws in the relevant paper.) - Is made by a Chinese company. Approaching o1's performance, extremely low prices, publicly available weights and a paper: any one of these would be AI news, and R1 delivers them all. Even someone writing fiction to boast about Chinese AI might not dare imagine this much—but it is real! The American AI community on X was stunned and incredulous, yet hands-on testing showed it really was exceptionally capable. Praise is now pouring in. ------ Nvidia senior researcher Jim Fan has insightful things to say about every AI advance. Here is his comment on R1— “We are living in a timeline where a non-US company is carrying on OpenAI's original mission—truly open frontier research that empowers everyone. This is incredible. The most entertaining ending has actually come true.” He not only praised DeepSeek's openness, but also mocked OpenAI's opaque style of dropping a hint today and a code name tomorrow. The contrast is striking. We have always said open source embodies the spirit of Silicon Valley. Openness was OpenAI's original vision, and Meta is also firmly committed to open source, although its technology is currently less advanced than OpenAI's… Now a Chinese company is not only approaching OpenAI technologically, but also publishing model weights and papers—even the technical details! ---- R1 still falls considerably short of o1 pro in generating insights, creative work, and reasoned argument. But remember that this is a relatively small model. It has only 671 billion parameters in total and is built from a series of mixture-of-experts (MoE) models. Each inference activates just 37 billion parameters. R1 thinks quickly and is very inexpensive. Using it directly on the website is free. If you call its API in your own application, the output price is US$2.19 per million tokens—4% of o1's US$60! That means you can use R1 for a great many everyday tasks. R1 also supports web search and PDF reading—features o1 currently lacks. My own experience was not that astonishing, but it was very usable. Some American users say it feels more capable than both ChatGPT search and Perplexity— ---- How far does DeepSeek's openness go? It not only publishes weights for free download and its training methods, but also lets anyone use R1 for data distillation to train their own models—and commercialize them. DeepSeek has even used two open-source models on the market, Alibaba's Qwen and Meta's Llama, to distill six smaller models for anyone to use. They all score quite highly on benchmarks— These distilled models are no small achievement. One, with 32 billion parameters, directly outperforms o1-mini in mathematics and programming. Another miniature model has just 1.5 billion parameters, yet its mathematics and programming performance already exceeds the two most mainstream non-reasoning models today, GPT-4o and Claude 3.5 Sonnet. It is small enough to run on your personal computer, or even your phone! Some people are already using it— That is incredible! Just six months ago, those two models seemed almost godlike… Now your own phone, without an internet connection, can have their capabilities, at least in mathematics and programming. A former DeepMind researcher also used R1 to distill a model of his own that outperformed o1-preview in mathematics and programming— What does empowerment mean? This is empowerment. ---- DeepSeek's paper introducing R1 makes the methods and details of training reasoning models public. OpenAI keeps how o1 reasons secret, leaving outsiders to guess… That is why some people say every AI lab is now reading DeepSeek's paper— ---- What is DeepSeek's secret? Reinforcement learning without human intervention. Just as AlphaZero once learned Go by playing itself without studying game records, engineers did not tell the model how to reason. They simply rewarded correct answers. The model explored on its own and acquired reasoning methods. The researchers first trained a base model called DeepSeek-R1-Zero, which spontaneously developed several problem-solving methods during training— - It writes out steps for mathematics problems and automatically checks whether each is correct; - If it recognizes a mistake midway through, it interrupts its thinking and starts the derivation again; - After solving a problem, it reflects on its steps, tries different approaches, and seeks the best solution; - It can automatically generate very detailed solution steps; - If it senses that a problem is harder, it automatically extends the reasoning steps and spends more time thinking… Simply put, it solves problems like a person. Trainers did not tell it to solve problems this way. These were abilities the model discovered for itself! Even more strikingly, an “aha moment” emerged during reasoning: the instant when a crucial step falls into place and everything suddenly makes sense— The researchers present were stunned to witness it for the first time. The model seemed almost alive, producing humanlike flashes of thought. You could even say its intelligence upgraded itself. R1-Zero sometimes mixes Chinese and English in its output, making the interface less friendly. The researchers made it more user-friendly, producing R1. Recently, OpenAI researchers have also said their approach involves abilities emerging automatically through reinforcement learning, which sounds consistent with the approach in DeepSeek's paper. But OpenAI has never provided details. DeepSeek has both explored the approach independently and made it public. Another interesting feature is that each R1 response provides its thought process—something OpenAI also does not provide in full. Many people say simply reading those thought processes is rewarding. Wharton professor Ethan Mollick, for example, remarked that watching R1's first-person thought process makes it hard not to feel strongly that it is a person… None of this was known to us before R1. Now a small company from China has contributed decisive new AI knowledge to humanity. ---- The mainstream coding application Cursor can already call R1 directly, and large numbers of programmers will use it daily. China has imposed no restrictions on American users; you can even sign in to DeepSeek's website with a Google account. Compare that with OpenAI's various restrictions: who is more open? ----- DeepSeek's ability to achieve so much with so little compute offers an important lesson for everyone. I believe OpenAI will also study why. If DeepSeek has opened the door to high efficiency with little compute, that is good news for humanity. But OpenAI's massive compute investment is not wasted. R1 is specifically optimized for mathematics and programming, and is not yet strong enough at other tasks. In investigating a scientific question, for example, it is less capable than o1-pro. There may be an 80/20 rule here. With relatively little compute, you can address, say, 80% of problems at work. But covering the remaining 20% requires far more resources. That latter part is precisely what OpenAI aims to do. Another major story in the past two days is that OpenAI, SoftBank, and other companies are joining forces on a US$500 billion undertaking called the Stargate Project, modeled on America's Manhattan Project to build the atomic bomb and its later Moon program. This project is not about programming or solving mathematics problems, or even artificial general intelligence (AGI)—it is about artificial superintelligence (ASI), accelerating research across disciplines and exploring the unknown. We must not underestimate the role of compute, or America's ambition. This is a contest over national fortunes. OpenAI remains far ahead. But China now has at least DeepSeek. The US–China AI contest is only beginning! ------ Robin: “Beyond the US–China AI contest, there is competition between fully open-source and closed AI systems. Personally, I believe creativity, innovation, openness, and breadth of vision will produce the eventual winner. We will see whether America and American companies, or China and Chinese companies, are more open. Rather than devote resources and energy single-mindedly to blocking others from catching up, we should find ways to move faster and achieve innovative breakthroughs ourselves.” [Editorial note: Here, “fully open-source” refers to the degree of openness. What R1 actually made public were model weights under the MIT license and a paper. Its training data and full code were not released.]