Keep pulling the thread on Akshay Kothari & Ryan Nystrom.
Notion's AI agent can access and utilize context from integrated platforms including Slack, Google Drive, Microsoft email, and Jira.
A key technical breakthrough for Notion AI was deciding to represent Notion pages as Markdown for the LLM, rather than using Notion's internal JSON-based block data model.
Notion's AI evaluation process uses two distinct sets of tests: a "golden set" where the model must achieve 100% accuracy, and a "challenging set" where it is expected to fail around 50% of the time to measure improvements.
Notion's upcoming custom agents will be shareable, customizable, and can be triggered to run autonomously based on events in Notion, on a schedule, or via integrations like Slack.
Internal AI usage at Notion quadrupled following the internal launch of its custom agents feature.
Notion's new AI agent product was in development for over two years before its release.
Notion AI agents are capable of performing web searches to gather information.
Akshay Kothari, Notion's co-founder, states that he now uses the Notion agent for 90% of his work, shifting away from manually creating documents or databases.
Notion plans to integrate Salesforce as a context source for its AI agent.
Notion had early access to OpenAI's GPT-4 model, which influenced its AI product development.
Notion's initial attempt to build its AI agent with custom RLHF-trained models failed because the underlying models were not reliable enough, forcing a project reset.
Notion's strategy is to be an "applied AI company" that uses the best available third-party models rather than training its own.