# DeepSeek's New AI Is An Absolute Monster

## Executive summary

DeepSeek has released a new AI system, DeepSeek v4.1 Flash 1017, demonstrating significant improvements in generating working software and operating via the terminal. While the version number increment is minor, the core innovation is attributed to the model's ability to learn from self-feedback, allowing it to critique and improve its own work, similar to human learning. The model is available via API for a low cost (claimed 15 times cheaper than Claude Opus 5.5), and its weights can be downloaded for free, promoting open-source alternatives in the AI space.

## Key takeaways

- DeepSeek v4.1 Flash 1017 Capabilities: The new model shows marked improvements in generating working software and operating within a terminal environment. (0:20)
- Cost and Availability: The model can be accessed via API for a cost of just over a dollar per million output tokens. The weights are available for free download, although running the model requires substantial hardware. (0:25, 0:45)
- Architectural Innovation: Self-Feedback: The key breakthrough is the ability of the AI to learn from self-feedback, meaning it can critique its own output and iteratively improve, rather than solely relying on massive amounts of new training data. (1:20)

## Technical details

- Model Identification: The model discussed is DeepSeek v4.1 Flash 1017. (0:05)
- Learning Mechanism: The primary technical advancement is 'learning from self-feedback,' where the AI judges and rewards its own performance, allowing for iterative improvement (e.g., fixing collision detection in a game). (1:20)
- Cost Analysis: The cost per task should be considered over cost per million tokens, as different AI systems require varying token amounts for the same task. (0:35)

## Practical implications

- The availability of downloadable weights for frontier models accelerates the development of open-source AI alternatives.
- The focus on self-feedback mechanisms suggests a shift in AI research from pure data ingestion to self-correction and iterative refinement.
- The low cost per million tokens makes advanced AI capabilities more accessible for commercial and academic use.

## Topics

Generative AI, Large Language Models (LLMs), Self-Correction, Open Weights Models, Software Engineering, DeepSeek v4.1 Flash 1017, Lambda

Source: https://www.youtube.com/watch?v=pryCaGv8wuo