DeepSeek R1: Scientific Progress vs User Experience
DeepSeek R1 shows in detail how problem-solving capabilities emerge, but in daily use it remains less coherent than GPT-o1. A scientific milestone, not a user experience one.
This article is more than a year old. Some of its findings may have changed since.
Why breakthrough research doesn't always translate to practical usability
Not impressed by DeepSeek R1 for practical use.
Scientific Significance vs Practical Reality
Yes, it confirms that human-intelligible problem-solving capabilities (what some might call reasoning, depending on the definition) naturally emerge when large language models are trained to find their own path toward solving verifiable problems. In contrast, GPT-4o and similar models "only" developed what some might call intuition about our world when trained on anticipating human-written text. Some might say "duh."
While DeepSeek R1 is the first to demonstrate in great detail how problem-solving capabilities can emerge—and will therefore dramatically accelerate progress in this space—we shouldn't underestimate the complexity of making these models practically usable.
The User Experience Gap
Supervised fine-tuning and other tricks of the trade mastered by major players like OpenAI are currently what make these models work in practice. After days of trying to leverage DeepSeek R1 in my usual co-creations with AI, I have yet to find a single interaction where everything was as coherent as I typically experience with GPT-o1.
DeepSeek R1 is a significant milestone in scientific progress, not in user experience. It is a stark reminder of the human effort—and cost—still required to make these models truly functional.
Beyond the Media Hype
The "DeepSeek threat" is merely a matter of perception disconnected from empirical evidence that media outlets could easily decode instead of amplifying. The tool is primarily a demonstration of a known training method—reinforcement learning on verifiable problems—never before implemented in open source at this scale.
The result isn't convincing at the user experience level, clearly showing the complexity and head start of other major players. Regarding the massive recent investments, they don't represent any concrete advance in AI; just an advance in creating large language models.
The real challenge isn't wasting money to create titanic models but using existing models that are amply sufficient (along with "old" deep learning or numerical optimization methods) to solve concrete problems. On this front, the United States can still invest trillions in models without advancing an inch.
Where Real Value Lies
University students would be much more useful creating value with models than creating models themselves. The focus should be on practical applications that solve real-world problems, not on chasing benchmark numbers or building ever-larger systems.
This is why user experience matters as much as scientific progress. True AI advancement comes from bridging the gap between research breakthroughs and tools that people can actually use effectively in their daily work.