Notion's Token Town — Sarah Sachs, Notion

Sarah emphasizes that Notion’s core mission is to serve as a durable system of record, not just for human collaboration but now also encompassing the interaction between humans and AI agents. This means providing a central place where both humans and AI can collaborate effectively, creating a more streamlined and efficient workflow. The company believes this foundational element is crucial for successful software factories and overall business operations.

Sarah outlines the evolution of AI usage, starting with 'AI as a thought partner' (early experimentation) progressing to 'AI as an assistant' (executing individual tasks), and culminating in 'AI as teammates' (integrating into critical workflows). The transition to AI as a teammate involves repetitive work automation and process interfacing, marking a significant step towards more sophisticated AI integration.

Sarah argues that cost represents a major obstacle for companies aiming to implement large-scale AI systems. She highlights examples of companies struggling with expensive commitments and inefficient processes, leading to frustration and ultimately hindering progress. This understanding is crucial for anyone working in applied AI to navigate the trade-offs involved in building durable products.

Sarah points out a concerning trend of vendor pricing instability, where model upgrades don't necessarily bring cost savings and can even increase token usage or introduce entirely new pricing tiers. She uses examples like upgraded reasoning models with unchanged per-token pricing and new model families with significantly higher costs to illustrate this problem, which Notion actively faces.

Sarah Sachs explains that open-weight models significantly reduce the cost barrier for customers, making AI solutions more accessible. These models also provide negotiation leverage by offering a credible alternative to proprietary providers, putting downward pressure on pricing which would otherwise be unavailable. This shift challenges the dominance of a few top providers and democratizes access to advanced AI capabilities.

The speaker emphasizes that relying solely on external benchmarks is insufficient for evaluating model performance. Instead, organizations need internal expertise to understand their specific system's error rates and latency requirements. She highlights the importance of tailoring AI solutions to unique operational needs rather than blindly following generic benchmark scores.

Notion has recently launched 'workers' which demonstrates that GPUs aren’t always necessary for every AI task. Many jobs, such as converting CSVs to PDFs or interacting with Notion tool calls via a CLI, can be efficiently handled by CPUs. This approach helps avoid becoming ‘token poor’ by optimizing resource utilization and reducing unnecessary GPU costs.

Sarah Sachs introduces the concept of a “lethal trifecta” for AI security: access to private data, exposure to untrusted content (like emails or web searches), and the ability to communicate externally. The combination creates significant risk, which is amplified by autonomous systems. She argues that building secure and trustworthy AI products requires addressing these risks proactively.