Inside Notion’s strategic bet on generative AI

There’s a new item on SF-based productivity startup Notion’s to-do list: Get tens of millions of users to try its brand-new AI tool.

On February 22, the company debuted Notion AI, a generative AI feature aimed at helping users do things like summarize notes, identify action items from a meeting, and create and edit text. It’s an optional add-on to the platform’s existing pricing plans, costing an extra $10 per user per month, and Ivan Zhao, co-founder and CEO of Notion, told us he sees the addition of generative AI as “a new type of Lego brick we’re introducing to our user tool set.”

A week into its debut, the tool had seen 14 million interactions, according to the company.

Notion’s new AI bet comes after the company raised its most recent funding round at a $10 billion valuation in 2021. The platform currently has more than 20 million users but declined to confirm how many are paying customers. For Notion AI, Zhao told us the company is working with multiple different model providers like OpenAI and Anthropic, and he views it as similar to how Notion uses Amazon Web Services as a cloud provider.

“AI is happening largely on [the] model-provider layer,” Zhao told us in an interview.

Taking note of Notion’s AI push

A key step to make Notion AI a reality came about at a company retreat in Cancun, where the team spent four days building a prototype of the tool; after Notion began testing it in November, the waiting list grew to 2 million, Zhao told us.

Zhao declined to share the size of the company’s financial investment in the AI tool or the number of employees directly working on it, although he did say he and his co-founder Simon Last are directly involved.

”Simon is my co-founder; he’s programming this every day,” Zhao said. “He’s personally coding on this.”

In an effort to keep the tool’s development as efficient and inexpensive as possible, Zhao said the company went with different model providers for different aspects of Notion AI: “Some models are fast, some models are cheap, some models are good for writing, some models you can go deep. It’s kind of like different ingredients for you to cook with.”

But for at least one of Notion’s business goals—the ability to use generative AI to, say, identify tasks or action items from a meeting transcript—the team had to do some additional “tuning” of the models.

Zhao told us that the “first layer” of tuning for a model like GPT-3 involves prompting, or asking the model a thoughtfully worded question in order to get the desired output.

“You don’t need too much investment, but you need to be pretty clever with what you ask the model,” Zhao said. “A language model is almost like a little animal—it’s almost like a dog. It doesn’t work exactly like what you want, but it’s like a Shepherd: If you train it well, you can do roughly what you want all the time. Prompting it is kind of like that: How do you train a dog to get what you want?”

If that method doesn’t do the trick, there are two other AI training methods a team can try, one of which is reinforcement learning with human feedback. But with Notion’s model providers, like OpenAI and Anthropic, Zhao said the team stopped at the first method if it was successful.

“It’s the fastest way to iterate [and] the cheapest way,” he said.

Zhao also believes that adding generative AI in an existing platform, rather than in an external tool, could help set Notion’s platform apart.

“Imagine you are in the 1890s [and] you’ve never seen electric lights,” Zhao said, adding, “Electric light bulbs…are much better than your kerosene-gas lights. It’s natural—you have to go there. You have to make it more mass-market. That’s the feeling that, when you see AI, what the generation of large language models can do, it’s just so natural, [to] apply right into what we’re already doing.”