DeepSeek is also a design story
The company’s decision to expose chain of thought is a surprise hit — and a reminder of how far AI product teams have left to go
This is a column about AI. My boyfriend works at Anthropic. See my full ethics disclosure here.
Since just after ChatGPT was released, I’ve been fascinated by artificial intelligence’s missing user interface. Chatbots offer increasing amounts of utility to those who master them, and yet — as with predecessors like Siri and Alexa — most of that utility is invisible to the user. For more than two years now, getting a sense of what chatbots can do, and how to guide them, has required a frustrating process of trial and error.
In 2023, I wondered what this frustration might mean for Inflection, a new AI company that had just raised $225 million. Co-founder Mustafa Suleyman told me at the time that the company’s chatbot, Pi, would stand out because of the way it had been trained to handle sensitive, emotional queries. And while Pi did show real aptitude for that, Pi’s chatbot was an empty box like any other. It might have been an excellent companion, but the product itself made no effort to communicate that.
A year later, Microsoft purchased most of Inflection’s assets in a Biden-era hackquisition. What remains of Inflection now sells a white-labeled version of Pi to enterprises.
Inflection’s failure to catch fire as a consumer product had many causes: a huge number of well funded competitors; a commoditized product; and a weak go-to-market strategy. But its most glaring failure, I think, may have been its design: it was positioned as a softer, kinder chatbot but attempted to convey this information only through colors and font choices. Whatever more sophisticated help it might have offered was left to the user to discover.
“Of course, users can learn over time what prompts work well and which don't,” wrote Amelia Wattenberger, a research engineer at Github, in a 2023 post about the limitations of chatbot design. “But the burden to learn what works still lies with every single user. When it could instead be baked into the interface.”
I thought Inflection and AI’s missing user interface this week during our weeklong industrywide freakout about DeepSeek.
Just as Inflection’s failure had many causes, so too did DeepSeek’s success. First and foremost is that it offered for free a reasoning model, r1, that is about as good as OpenAI’s much more expensive o1. Second is the fact that the model was made by a Chinese company, at a time when many Americans believed that the country was much further behind in AI development than the high quality of r1 suggests.
But there is a third, less discussed aspect of DeepSeek’s success, and it has implications for everyone who builds AI products. Unlike other reasoning models before it, DeepSeek responds to your queries by telling you what it thinks you want, and what it thinks it should do in response. It’s as if you had an assistant who, before completing the task you requested, first tells you their complete understanding of your desires and their plan for fulfilling them. The technical term for what r1 is doing is exposing its chain of thought, and it might prove to be as influential on other companies as anything else that DeepSeek has done.
