What Can We Learn from DeepSeek?

There’s been no shortage of ink spilled since China’s DeepSeek dropped their R1 reasoning model. It’s been making news, briefly tanked Nvidia’s share price, and reshaped the AI leaderboard. Let’s dive in.

Following the open-source philosophy championed by Meta’s Yann LeCun, DeepSeek released its model and published a paper detailing its groundbreaking learning methodology. Its reasoning capabilities rival OpenAI’s wallet-busting $200/month GPT-4 Omni, yet it operates under MIT’s permissive license—making it free for anyone to install and run. While allegations emerge about DeepSeek distilling from external models, R1’s real achievement is in how it was trained to reason. Already, projects like Open R1 have emerged to replicate and refine DeepSeek’s pipeline. This move has reset the AI playing field, allowing anyone to build from the cutting edge and push AI forward.

AHA! So What Was So Groundbreaking About It?

Previous models like ChatGPT and LLaMA used Reinforcement Learning from Human Feedback (RLHF). For them, reward was based on expected human answers, or answers we preferred—a method susceptible to bias. DeepSeek instead applied Reinforcement Learning (RL), giving the model rules for what its answers should be but allowing it to iterate until it arrived at a solution. The developers built a framework that would recognize when the model had a breakthrough in logic that led to a correct answer—an “aha moment.” It learned to use the reasoning pathways that led to aha moments, and better reasoning emerged. It discovers its own novel problem-solving approaches, which do not necessarily mimic human reasoning. History has shown that Reinforcement Learning often leads to solutions that defy human intuition—just as AlphaGo did in 2016.

In 2016, AlphaGo demonstrated the strength of Reinforcement Learning (RL) by mastering Go through self-play rather than imitating human strategies. As an advanced form of Artificial Narrow Intelligence (ANI),it made alien, seemingly flawed moves—only for them to later prove decisive, game-winning moves. As AI moves toward greater generalization and Artificial General Intelligence (AGI), RL-driven models may exhibit similarly unintuitive but ultimately prescient behaviors, challenging human expectations of intelligence.

Cognitive Space

Throughout history, humans have used reasoning to navigate challenges in our environment. In the pursuit of AI’s nearly limitless upside, the DeepSeek team had to innovate around constraints like chip embargoes and limited compute. What began as a necessity-driven innovation by DeepSeek’s team has now imbued their model with the ability to innovate itself, revealing that intelligence—whether human or artificial—emerges through problem-solving under pressure. This marks a shift—reasoning is no longer solely a human domain.

Like mathematics, reasoning may be something we discover rather than invent—a fundamental structure that becomes clearer as intelligence advances. AI and human cognition are now co-evolving, forming a hybrid intelligence system where machines contribute digital thought processes, expanding the boundaries of problem-solving in ways we have yet to fully understand.