DeepSeek Shakes up the LLMs/Reinforcement and Recursive Self Learning/Synthetic Data

China made ground-breaking news in early 2025 when the DeepSeek-V3 R1 AI large language was announced. The developers claim to have created the system with only $5.6 million dollars in investment (compared to the normal U.S. AI model training budgets of $50 to $100 million) and to have done so quickly, with a lower total server investment. The model is open source, and fees charged for accessing and using the model’s data are significantly lower than those of competing models. This announcement from DeepSeek, established in 2023 in Hangzhou and funded by the hedge fund High Flyer, totally shook up competitors worldwide. DeepSeek planned to launch an advanced R2 version quickly, and will likely grow through continual development of newer, enhanced models.

DeepSeek utilized several advanced strategies, including Reinforcement Learning and Distillation. Reinforcement Learning encourages the system to experiment and improve itself continually. Distillation is a process of creating a smaller, more efficient language model by “distilling” existing larger language models (which may include the generation of “synthetic” data, or new data generated during this process).

Reinforcement Learning (RL) and Recursive Self Learning (RSL)

Reinforcement Learning (RL) is a machine learning technique whereby an agent or system learns to make decisions by receiving feedback on its decisions and actions. The system is “rewarded” or scored by the addition of positive points to its score when it is correct and penalized with negative points when it is wrong. The system is programmed to attempt to achieve a high score, and it can gradually adjust its behavior to maximize the total reward score. These systems learn the best processing paths to achieve desired outcomes or solutions. To some extent, RL imitates the trial-and-error learning process that humans use to achieve their goals. A good example is an autonomous (self-driving) vehicle system that enables a car to learn how to navigate challenging traffic conditions.

Recursive Self Learning (RSL) (also known as Recursive Self Improvement or RSI) is a system where the AI model makes itself much better and more accurate over time. The system may write/enhance pieces of its own code, repeatedly improving itself. For example, an RSL system might teach itself to play a complex game, such as chess, by playing the game over and over against expert humans, learning from the moves that generate wins and losses.

Dozens of major large language models (LLMs) are constantly evolving and improving in a massive global competition with each other. For example, in February 2025 Google announced the latest iteration of its Gemini model, called Gemini 2.0 Flash. The firm claims greater capabilities and much faster responses with this new model. During the same month, ChatGPT announced its “Deep Research” enhancement, “an agent that uses reasoning to synthesize large amounts of online information and complete multi-step research tasks for you.” Aimed at completing complex tasks quickly, it conducts multi-step research on the internet, by finding, analyzing and synthesizing hundreds of online sources to complete a “comprehensive report at the level of a research analyst.” The fact that Deep Research searches both the internet (in real time) and an OpenAI large language model at the same time is a significant step up in capabilities.

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.

Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.

Become a Plunkett Research Online subscriber to gain unlimited access to our massive amount of vital industry data and market research.

✔ 100,000 Industry Statistics
✔ 500 Industries Analyzed for Market Size, Profits and Forecasts
✔ 150,000 Industry Executives Profiled
✔ 25,000 Companies Profiled, Ranked and Analyzed
✔ 1,000 Vital Trends Analyzed, totaling 1,000,000 words

Learn More

DeepSeek Shakes up the LLMs/Reinforcement and Recursive Self Learning/Synthetic Data, Business and Industry Trends Analysis

A Representative List of Organizations that Have Used our Research and Products:

How to access Plunkett Research data.

What data challenge can we help you solve today?

Take a Tour

Contact

Testimonials