From Note-Taking App to AI Workspace: The Simon Last Interview
In a Nutshell
Notion evolved from a note-taking app into an AI workspace where humans collaborate with agents, starting with GPT-4 experiments in 2022 and launching features like AI Writer (Feb 2023), Q&A (Oct 2023), and full personal/custom agents (Aug/Sep 2024) that autonomously handle tasks via agent-friendly APIs like Markdown and SQLite. Simon Last highlights rewriting AI harnesses every 6 months, using coding agents to boost engineering ambition (e.g., 100x productivity for top users), and personal agents for email triage and feedback routing that learn from memory pages. The core shift: Notion now optimizes for humans managing agents, not directly performing work, with primitives like databases enabling coordination.
These notes were generated by AI and may contain inaccuracies.
From Note-Taking App to AI Workspace: The Simon Last Interview
Hi listeners, welcome back to No Priors. Today with Simon Last, co-founder at Notion. Discussion covers Notion's new vision for the AI age as a platform for humans and agents to collaborate, changes in engineering and product org at Notion, and new tools for thought.
Notion is at scale with lots of users, started quite a while ago. Notion has embraced AI aggressively. Simon first got hands on GPT-4 at a company offsite in Mexico in 2022.
I've been watching what's going on in general, super curious about the technology, fascinated to try everything and think about how we can apply it.
It became real when playing with GPT-4, accessed via a prototype interface. Simon and co-founder Ivan both got access. Two big realizations: (1) pretty smart, follows reasonably complicated instructions, writes and edits things; (2) scope of knowledge extremely deep and broad world knowledge. Instantly clear: time is now to apply it, only going to get better.
Talking about Mexico, GPT-4: clearly the time. Started with long-term and short-term vision.
Short-term: obvious writing assistant in documents—select text to rewrite, write text, look something up with sources. Started tiger team, launched in 2-3 months.
Long-term: general assistant with all tools inside Notion humans have—create databases, query/manipulate them, create/edit documents, weave together for longer-range tasks. Started both immediately; short-term shipped quickly, long-term took much longer.
First launch of AI-specific Notion features: February 2023.
Timeline of shipped features:
- First: AI Writer (writing system)—single-step tasks like rewriting/editing text, no retrieval, raw model access. Easiest to get working.
- Next: Q&A with semantic index of entire workspace, ask question for grounded answer with sources. Obvious utility. Beta before, GA in October 2023. Bigger effort: real-time updating index, serious evals/quality.
Q&A multi-year journey. Once Notion index working, indexed Slack, Google Drive, launching new ones regularly—now fairly complete.
Those are difficult problems; native products haven't solved perfectly.
What right do we even have to do this?
Most companies bad at indexes. Key: AI-pilled savviness, craft, attention to detail. Empirical/iterative: try queries per data source (e.g., Slack vs Google Drive different). Craft/love: try queries, use daily, iterate retrieval.
Diversity in workspace organization (e.g., messy vs organized). With embeddings, doesn't matter as much—AI cares about text snippet with context, retrieves it. Advise: don't worry about organization, pipe it all in.
Decisions like chunking strategy change performance, transparent to user, independent of organization.
Difficult challenge given content differences. Lots of iteration: chunk sizing, retrieval steps/pipeline.
Ivan said ask how many times rebuilt Notion and harnesses.
Running joke: rewrite AI harness every 6 months, time decreasing as progress accelerates.
Key: keenly aware of current models/technology state, design harness/system/product around it—rewrite every 6 months. Fun process to restart/rethink. About to release new harness in next 1-2 weeks, already thinking next.
Leads to how Notion's engineering/product/research org works with coding agents—willingness to rewrite harness up dramatically with agent help.
Fun to use coding agents; ambition of what to build increased a lot.
Changes over 2-3 years, multiple eras: tab autocomplete, inserting/rewriting code. Big shift early 2023 with agents (started using Claude Code ~April 2023). Push agents to end-to-end implement/verify/maintain, requires architecting/verification loop thought. Upshot: more ambitious, robust builds if done well; slop if badly.
Doesn't change team lens much (size/seniority). Individual impact higher, depends on tool use/willingness. Still smallish tiger teams—better if small, true before and now.
Main: harness tools.
Gap bigger: median engineer vs 10x engineer (or 10x willing to use tools). Can be 100x or 1000x engineer using tools right. Minimum bar same, maximum extremely increased.
Impact: more messy/chaotic (loves it)—more prototypes. E.g., design team made design playground Git repo: simplified Notion with UI primitives, agent inside, high-fidelity prototypes quickly (URL to deployed prototype). True across stack: more chaotic, ambitious PRs.
Draw line on dangerous/sensitive areas (e.g., data loss risk)? Still reviews on all PRs. PRs now agent-written, larger/complex (worst part), but better tested—demand full unit tests. High confidence, but think carefully: what change, verification, safe deploy, enlist agent.
General assistant doesn't quite exist yet. Struggled years to build agent—worked but not useful (too early). Tried 3-4 times, launched last fall (August/September 2024): Notion AI full agent with access to everything in Notion. Original vision works, fully shipped.
Personal agent: every user has one, access to all user has access to—create database, update/create documents, search web/do research.
Just launched (last week): custom agents—create new, name it, grant access (default none). Runs autonomously in background. Examples:
- Access own database/file tasks, attach to Slack channel, responds/filing tasks.
- Access weekly reports database, search web/workspace.
Represents work/job/knowledge task autonomously.
Excited: extremely good at bootstrapping capabilities from initial kernel to do anything (e.g., build unsupported integration, deploy/use it).
Notion agents: coding as tool, close to broader agent definition.
Coding agents kernel of AGI. AGI will be a coding agent. Code useful primitive for deterministic logic.
Exciting for knowledge work agent: bootstrap capabilities (build integration, connect new data source).
Notion at scale, but productivity/platform players at more scale, all building agents/integrating sources.
Landscape: labs, software platforms, infrastructure. Notion as Switzerland for models—customers don't want lock-in, labs release new versions monthly (one better). Easy access to all best models, switch around.
Open source plays in—getting really good (e.g., 4 Chinese models good/cheap). Released one in agent last week, will do all four. Many use cases for cheaper than frontier.
Role: take best models, high-quality state-of-the-art agent implementations, easy access, collaborative workspace for humans/agents to coordinate. Very needed, doing tastefully/well-executed.
Need index for good agents, give agents human tools in Notion (blocks, databases). All still useful.
Challenge: make convenient for agents (new customer). Past: convenient for humans, then APIs for humans coding.
E.g., API JSON for blocks verbose/horrible. Solution: convenient APIs for agents.
- Markdown dialect: normal markdown enhanced with Notion blocks—models great at it, reads/writes pages.
- Databases: SQLite—models good at SQL + speak.
Default didn't work; engineering challenge solved—now extremely convenient, agents naturally good.
How figured what makes API better: empirical (try things, notice token waste, iterate), first principles (model training priors, natural fit, agent loop patterns), play around, user research with agent as user (chat/script/scale), ongoing evals.
Infinite access to agent wonderful.
Simon uses Claude Code or Cody (CLI tools, comfortable in CLI). Goal: as many running as possible, all the time. E.g., every night give enough work so still running by morning (victory). Done last 5 nights. Personal record: coding agent 13 days straight without stopping, working tasks.
Well-prompted; nerve-wracking, check before bed.
Uses personal agent all time—company context. E.g., last night asked about custom agents launch signals.
Many custom agents running. Personal favorite: I have a
My personal favorite is an email triage agent with access to all work and personal emails. It wakes up every day, archives stuff I don't need to see, and trains over time to learn preferences.
Do you actually label data for it?
It's pretty easy. Make the agent, give it email access, create a blank page as its memory, and let it edit that page. Say "go look at my emails and interview me"—it proposes things to archive, you correct it, and it generates rules on what's correct.
For the first couple days, I corrected it. After a couple weeks, I dropped approval entirely; it now automatically archives everything I don't need to see.
Wow. It completely solved my email problems cuz for me, like I don't use email that much for work stuff. Like it's mostly in Slack. 95% of personal and work emails I don't need to see. It's a waste of time. Now my inbox only has stuff I need to see.
Another agent triages customer feedback, internal feedback, and bugs from a Slack channel where people post random product feedback and bugs. Previously, some got answered, others ignored haphazardly across teams.
Its job is to route to the right place. It uses a similar memory pattern, learning on the fly where to file bugs. Over time, it built hundreds of rules. For example, a mobile app bug routes to the mobile team and files a task in their database.
Do you look at the generated and updated memory to check if it makes sense?
I did at first, but once trusted, I ignore it. If it breaks, I fix it. The benefit is not reading email.
General pattern: build as prototype in approval mode, watch closely, then trust after it runs many times.
Is there anything you do internally at Notion to make sure non-technical teams have intuition for building agents or expressing productivity?
We do workshops and hackathons frequently. A month ago, I did one with the people team—they're among the highest adopters of custom agents. They handle workflows in Slack and Notion manually.
People are excited to try but need a push for intuition and getting started. Once past the technical barrier—what is a prompt, agent, trigger, wake-up—the concept is intuitive and humanlike.
The biggest barrier is getting people to try it, assuming it will work.
You and Ivan met in the internet tools for thought community. Tools for thinking are different now. Has your core conception of Notion changed? What does the tool do for thinking? What should agents do?
It changed a lot. Before AI, goal was best tool for humans to directly perform work.
Now, goal is best tool for humans to manage agents to do the work. That's a fundamental shift.
Most primitives remain useful: documents for unstructured writing (agents love markdown), databases for structured data. With 100 background coding agents, you want a Kanban board, not 100 chat threads—for coordination.
New primitives needed: representing agents and how they interact with pages/databases.
What's changed about how you personally build?
Completely changed. Haven't written code since last summer; don't type code anymore.
Shifted from humans typing all code, to tab-complete, to talking to agents for tasks (humans in outer loop), to now: design end-to-end task (make change, verify), then outer verifier—double-check at end, monitor if off rails.
Now I'm the agent manager instead of the coder.
Amazing. Thanks Simon. Great discussion about becoming agent managers, hopefully in Notion.
Find us on Twitter at @nopriorpod. Subscribe to YouTube. Follow on Apple Podcasts, Spotify, or wherever. New episode weekly. Sign up for emails or transcripts at nopriors.com.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
