4-hour closed door meeting! What did DeepSeek Liang Wenfeng talk about 1、 I have no intention of building a super app, and the current traffic is just sesame seeds There is no plan to create the next ByteDance or Tencent, and they will not follow the trend to build a traffic oriented super app Not competing for traffic or short-term profits is because there are still greater opportunities ahead, and there is no need to allocate core R&D resources for short-term gains. In terms of business track selection, DeepSeek voluntarily gave up on a number of popular directions: 3D generation, video generation, and world models are all not laid out; Multimodal is only used as a supporting auxiliary component and cannot determine the upper limit of intelligence; The model illusion problem, which is a common headache in the industry, is internally defined as a common product optimization problem that will be iteratively fixed, but it is not part of the current core goal The product is just a byproduct on the AGI road, naturally possessing the advantage of dimensionality reduction Currently, focusing on product development and traffic management is not the optimal choice for maximizing profits The paid services for C-end clients and B-end APIs are just byproducts that emerged during the development of a universal base model. Standing at the forefront of general intelligence technology and applying it downwards is itself a dimensionality reduction blow. Last year, DeepSeek unexpectedly went viral. The team never took the initiative to invest in traffic or monetize, but users spontaneously retained it, which is a practical proof of this strategy. The company does not have a dedicated sales and customer service team, with only a small number of personnel maintaining basic services. Customers rely on the cost-effectiveness of the model to proactively approach them 2、 Clear AGI ladder route: CoT → Agent → Continuous learning, the core breakthrough point of the next generation model has been determined Compared to the strategic trade-offs that are not made layer by layer, what the market is most concerned about is DeepSeek's complete and implementable general intelligent evolution path. Completed Phase: CoT Thinking Chain Relying on the chain of thought to enable models to autonomously and deeply reason, surpassing top humans in professional tasks such as Olympiad mathematics and code, but there are natural shortcomings: complete context must be inputted to complete tasks, and long-term experience accumulation cannot be achieved like humans. There is still a clear gap between general intelligence and this 2026 Core Challenge: General Agent Intelligent Agent This year, the research and development focus is fully tilted towards agents, expanding the executable task boundaries of the model. However, the current agents are limited by their inability to learn autonomously for a long time, and their capabilities have a ceiling Clear priority of the track: At present, we will make every effort to tackle the Coding Agent, and all vertical industry agents such as finance and healthcare will be placed in the rear without allocating core R&D resources Simultaneously establish an internal R&D ranking: the first service target of the self-developed model is internal researchers within the company, prioritizing the improvement of team R&D efficiency to accelerate the implementation of AGI. The external commercial value is only an additional benefit Next Generation Core Bottleneck: Continuous Learning (a Global AI Common Challenge) The current AI does not lack taste and intuition in text creation and logical judgment, but its biggest weakness is its ability to continuously learn, which is also the core criterion for distinguishing the next generation of models After breaking through continuous learning, the industry will usher in a gradual intelligent singularity, not an instant mutation, but a long-term gradual process: models can autonomously train and iterate the next generation of large models, achieving AI reverse acceleration of AI research and development, significantly shortening the development cycle of general intelligence Long term finale: embodied intelligence After completing the stage of self evolutionary singularity, the ultimate landing point of technology is embodied intelligence, landing robots, physical service scenarios, covering practical needs such as household chores, elderly care, and human resource substitution 3、 Open source+low-priced underlying logic: only make reasonable profits, fully compatible with open source and commercialization Doubts about the long-term existence of open source in the industry and its impact on paid services 1. Sincerely open source, models are not castrated, and there is no fear of competition from peers The strongest main model of the future company will also be fully open sourced to the outside world, and the online commercial version will have the same weight as the open source, without deliberately weakening the ability of the open source model. Even if peers enter the market competition based on open source models, DeepSeek is not worried at all. The core basis is threefold: The scale of the AI industry is unprecedentedly large, and it may occupy more than 10% of the human society's GDP in the long run. A single enterprise cannot monopolize the market. Open sharing and co building ecology are objective laws for long-term survival Under the same effect, low-cost deployment and efficient inference operation and maintenance are extremely strong engineering barriers. Startup companies lack computing power, large companies have bloated organizational processes, and the vast majority of vendors cannot replicate DeepSeek's cost control capabilities The current scale belongs to the exclusive dessert area of the company: if the scale is too small, there is insufficient research and development strength, and if the scale is too large, the management and cost burden will increase sharply 2. The iron rule of pricing: recoup hardware costs within 10 months and refuse to maximize profits There is a fixed pricing standard within the company: the server hardware investment takes 10 months to recoup the cost, which is a reasonable profit. Even if the price doubles, user demand is almost inelastic, and the team still insists on the low price route. Liang Wenfeng shared a case on site: Previously, the DDCP model was initially priced too high due to concerns about demand overload, and the internal team generally did not approve of it. Later, the price was directly reduced to 1/4 of the original price, and all employees of the company cheered together Our goal is not to make the most money, but to make high-quality models affordable for everyone while ensuring reasonable profits The ultimate competition in the industry only focuses on three core dimensions: cost, landing time, and user experience, among which cost is the first barrier. Based on a 10 month payback period, the company only earns about 6 times the hardware revenue. Within this profit range, third-party vendors cannot compete on price through private deployment, and open source will not erode their own API revenue