Why Notion’s greatest AI breakthrough got here from simplifying all the pieces

When initially experimenting with LLMs and agentic AI, software program engineers at Notion AI utilized superior code technology, advanced schemas, and heavy instructioning. 

Shortly, although, trial and error taught the group that it might eliminate all of that difficult information modeling. Notion’s AI engineering lead Ryan Nystrom and his group pivoted to easy prompts, human-readable representations, minimal abstraction, and acquainted markdown codecs. The consequence was dramatically improved mannequin efficiency. 

Making use of this re-wired strategy, the AI-native firm launched V3 of its productiveness software program in September. Its notable characteristic: Cutomizable AI brokers — which have shortly develop into Notion’s most profitable AI device so far. Primarily based on utilization patterns in comparison with earlier variations, Nystrom calls it a “step perform enchancment.”

“It's that feeling of when the product is being pulled out of you relatively than you attempting to push,” Nystrom explains in a VB Past the Pilot podcast. “We knew from that second, actually early on, that we had one thing. Now it's, ‘How might I ever use Notion with out this characteristic?’”

‘Rewiring’ for the AI agent period

As a conventional software program engineer, Nystrom was used to “extraordinarily deterministic” experiences. However a lightweight bulb second got here when a colleague suggested him to easily describe his AI immediate as he would to a human, relatively than codify guidelines of how brokers ought to behave in varied eventualities. The rationale: LLMs are designed to grasp, “see” and motive about content material the identical manner people can.

“Now, every time I'm working with AI, I’ll reread the prompts and gear descriptions and [ask myself] is that this one thing I might give to an individual with no context and so they might perceive what's happening?” Nystrom mentioned on the podcast. “If not, it's going to do a foul job.”

Stepping again from “fairly difficult rendering” of information inside Notion (resembling JSON or XML) Nystrom and his group represented Notion pages as markdown, the favored device-agnostic markup language that defines construction and which means utilizing plain textual content with out the necessity for HTML tags or formal editors. This enables the mannequin to work together with, learn, search and make adjustments to textual content recordsdata.

Finally, this required Notion to rewire its techniques, with Nystrom’s group focusing largely on the middleware transition layer. 

Additionally they recognized early on the significance of exercising restraint relating to context. It’s tempting to load as a lot info right into a mannequin as potential, however that may gradual issues down and confuse the mannequin. For Notion, Nystrom described a 100,000 to 150,000 token restrict because the “candy spot.” 

“There are instances the place you’ll be able to load tons and tons of content material into your context window and the mannequin will battle,” he mentioned. “The extra you place into the context window, you do see a degradation in efficiency, latency, and likewise accuracy.” 

A spartan strategy can be vital within the case of tooling; this can assist groups keep away from the “slippery slope” of countless options, Nystrom suggested. Notion focuses on a “curated menu” of instruments relatively than a voluminous Cheesecake Manufacturing unit-like menu that creates a paradox of alternative for customers.  

“When individuals ask for brand new options, we might simply add a device to the mannequin or the agent,” he mentioned. However, “the extra instruments we add, the extra choices the mannequin has to make.”

The underside line: Channel the mannequin. Use APIs the best way they had been meant for use. Don't attempt to be fancy, don't attempt to overcomplicate it. Use plain English.

Take heed to the total podcast to listen to about: 

• Why AI remains to be within the pre-Blackberry, pre-iPhone period; 

• The significance of "dogfooding" in product growth;

• Why you shouldn’t fear about how price efficient your AI characteristic is within the early levels — that may be optimized later; 

• How engineering groups can maintain instruments minimal within the age of MCP; 

• Notion’s evolution from wikis to full-blown AI assistants. 

Subscribe to Past the Pilot on Apple Podcasts, Spotify, and YouTube.