DeepSeek-R1 — Intuitively and Exhaustively Explained

An “Aha” moment in Artificial Intelligence

“Deep Cycle” by Daniel Warfield using MidJourney, all images by the author unless otherwise specified. Brought to you by the subscribers of Intuitively and Exhaustively Explained.

In this article we’ll discuss DeepSeek-R1, the first open-source model that exhibits comparable performance to closed source LLMs, like those produced by Google, OpenAI, and Anthropic. This heightened performance is a major milestone in artificial intelligence, and is the reason DeepSeek-R1 is such a hot topic.

We’ll begin our exploration by briefly covering some of the fundamental machine learning ideas that DeepSeek builds off of, then we’ll describe some of the novel training strategies used to elevate DeepSeek-R1 past other open source LLMs. We’ll spend a fair amount of time digging into “Group Relative Policy Optimization”, which DeepSeek uses to elevate it’s reasoning ability, and is largely the source of it’s heightened performance over other open source models.

Once we have a thorough conceptual understanding of DeepSeek-R1, We’ll then discuss how the large DeepSeek-R1 model was distilled into smaller models. We’ll download one of those smaller DeepSeek models and use it to make inferences on consumer hardware. Finally, we’ll close with speculation as to how DeepSeek may impact the state of the art of AI moving forward.

By the end of this article you will understand what DeepSeek is, how it was…