Notion's Token Town — Sarah Sachs, Notion

AI Engineer · 23 min · 350 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:01[music]
2. 00:21Okay.
3. 00:22Hello.
4. 00:22Okay.
5. 00:23Before I get started, you guys, this is a huge keynote room.
6. 00:26Can everyone like come forward because I'm talking to like four empty rows and disperse people.
7. 00:32Do me a favor.
8. 00:32I'm spending 30 minutes telling you all of our secrets.
9. 00:35I can see you still.
10. 00:36Thank you.
11. 00:37Thank you.
12. 00:37Thank you.
13. 00:38Thank you.
14. 00:38We're just going to chat.
15. 00:39It's a giant room and there's 500 of us.
16. 00:42This room is way larger than that.
17. 00:44Thank you.
18. 00:45Honestly, I knew you guys had it in you.
19. 00:48It's really not so hard.
20. 00:49Thank you.
21. 00:50I also sit in the back.
22. 00:51I also work during talks.
23. 00:53I get it.
24. 00:55I totally get it.
25. 00:55I did all day, but not for me.
26. 00:58Okay, I'm gonna start, but I'm gonna still point at you if you're in the back like you.
27. 01:02Okay, I'm Sarah.
28. 01:05I'm lead our engineering teams for AI at Notion.
29. 01:09Um, welcome to my talk.
30. 01:10Um, it's about token town.
31. 01:13How do you go from not go from AI pled to AI poor?
32. 01:18Okay.
33. 01:19Um, I know that today is all about software factories.
34. 01:22We're going to talk about that, but we're going to talk about how to do it sustainably.
35. 01:25This is me.
36. 01:26This is on my first day at Notion in a very sweaty subway.
37. 01:30Um, like I said, I lead our AI teams at Notion.
38. 01:33Um, and I negotiate AI contracts for a living.
39. 01:36My team jokes I act like Anna Winter.
40. 01:38So, this is a nice a nice image of me with AI Anna Winter hair after a press article referred to me that externally.
41. 01:46Um, and that's kind of the idea, right?
42. 01:48Uh, how do you think about negotiating between different vendors?
43. 01:52um making sure that you maintain taste for your company.
44. 01:56I don't do it alone.
45. 01:57Um this is launch day of one of our recent launches.
46. 02:00This is just a subset.
47. 02:01Any good engineering manager points out that we have a whole company of people building this.
48. 02:05I'm just the one that gets to come talk to you about it.
49. 02:08So, we've been building a lot.
50. 02:10Um this is um an example of our AI usage um just in 2026.
51. 02:15Um and we've been really proud of how we've been able to grow that usage.
52. 02:19and I'm going to talk to you about how you can build an AI native product and an AI native company.
53. 02:24Um, but this is just to give me some credit that that we're doing it kind of well.
54. 02:29Okay, so for those of you that don't know, notion's always been that durable system of record.
55. 02:34It's always been the place where you can collaborate with your peers.
56. 02:37Um, but today that point of collaboration is a little bit different.
57. 02:40It's not just humans.
58. 02:42Notion's always been the place for collaboration.
59. 02:44And today that collaboration happens between humans and agents.
60. 02:47Humans and humans, agents and agents.
61. 02:52And we like to think about AI transformations going through this journey.
62. 02:56And I'm sure some of you are looking at the slide and wondering where you are.
63. 03:00AI as a thought partner is when we all started tinkering.
64. 03:03We all started just going to the very first version of chat on Thanksgiving when it came out three years ago, four years ago.
65. 03:10And we started saying like how can I send this email to my landlord to say that I shouldn't pay for repainting right then we'd copy paste it enter it into our email
66. 03:21eventually we started getting to a place where we could use AI like an assistant AI was able to maybe execute individual tasks that's how notion AI really took off in the beginning
67. 03:32um and it was able to save employee time but functionally was limited in its capabilities based on what humans asked it to do AI as teammates is what we were really excited to launch almost a year ago now.
68. 03:43Um, but this is true in many products where we can do repetitive work and think about a process and have AI do that process.
69. 03:51What I think is really interesting is when AI actually becomes that critical workflow where processes are interfacing with each other and you have entire systems running.
70. 03:59How many of you guys feel like you have AI as a system down?
71. 04:03Aren't you sad you came up now?
72. 04:06Great.
73. 04:06None of you.
74. 04:07Exactly.
75. 04:08We have found that no one has figured out how to do this well.
76. 04:1188% of people can't even get past AI as an assistant.
77. 04:16And why is that?
78. 04:17We have a thesis at notion.
79. 04:18It's because there's too much siloed data and not a sturable system of record for that point of collaboration.
80. 04:25And we believe that for your software factory to work, for your company to work, and for your systems to work, you need that durable system of record.
81. 04:33And that is notion's mission.
82. 04:36So doing that is expensive.
83. 04:38Um you see a lot of companies that try and commit themselves to this vision and these are just a series of headlines all within a week of how that's painful.
84. 04:49So you can put all of your money into a process to try and make a system and you end up feeling like this, right?
85. 04:56You end up using a blowtorrch to light what is actually a large cigar.
86. 05:00But you kind of get the idea.
87. 05:04Cost is a structural barrier to entry.
88. 05:07It makes it hard for you to serve products.
89. 05:09It makes it hard for you to build factories.
90. 05:11And it is ultimately, I would posit, one of the largest reasons why things do not happen at scale successfully today.
91. 05:18And I would argue for anyone working in an applied AI company, it's something for them to be really familiar with to understand the trade-offs that they're making to build durable
92. 05:27and exciting and enlightening product for their customers.
93. 05:32But that's not really how the market is today, right?
94. 05:35I'm not going to name names here, but you guys have search engines.
95. 05:39You can figure it out.
96. 05:40Exhibit A. A reasoning model gets upgraded.
97. 05:43Amazing.
98. 05:44The per token pricing is the same.
99. 05:46What's not to love, you try it out, it uses three times as many output tokens, right?
100. 05:53Exhibit B. A model gets upgraded, but it has an entire new digit, right?
101. 05:59Whatever demarcation system that model family likes, it's brand new.
102. 06:02It's 40% more than its predecessor, which is being deprecated in the next four months.
103. 06:07These are real scenarios that we face at notion.
104. 06:09All of you are nodding because these are common pretty much monthly now.
105. 06:15But here's the problem.
106. 06:16Are you growing 40% in that time period?
107. 06:20Are you making 30 3x more revenue?
108. 06:23No. So, how do you navigate the system?
109. 06:26If you just auto upgrade your model and everything that you're doing, you're you're giving someone a bad deal.
110. 06:31Either your customers or your investors, depending on how you charge and where you get your money, neither are good.
111. 06:39Fortune 5 million companies have the capability to navigate this.
112. 06:43They can hire large consulting teams, have durable teams on their own, and build expertise on how to navigate these trade-offs.
113. 06:51Um, most people don't.
114. 06:53everyone else has no ability to negotiate with leverage and they're stuck in these scenarios, right?
115. 06:59Part of my job as that Anna Winter joke is to think about advocating for the Fortune 5 million, the non-fortune 500 companies that don't have the mass to have leverage and negotiate, but need to think about how.
116. 07:12And I'm going to share some of the lessons that I've learned when I have kind of large amounts of traffic behind me that I think scale to those who don't.
117. 07:22Um, this is probably less of a secret now than it was when I started giving talks like this u maybe four months ago.
118. 07:30Um, your supplier is your competitor.
119. 07:33I know very few people who have convinced me that that's not true.
120. 07:38Um, you will always be getting a bad deal on tokens with someone who builds them natively, right?
121. 07:43Sometimes the cost of goods served is extremely different.
122. 07:46you're basically they're serving a first-party product and then you're buying those tokens at a huge search charge and then selling them again at another search charge.
123. 07:54Um, that's not really value you can defend.
124. 07:56You're getting a really bad deal.
125. 07:58And if you tie yourself to one provider, you have no exit.
126. 08:01If you build an AI product that you're selling with this structure, you are crossing your fingers and hoping that you are a viable business.
127. 08:10I do not encourage that.
128. 08:14This is really interesting.
129. 08:15Dylan in some analysis posted this.
130. 08:17I think it's it says eight hours ago.
131. 08:19It wasn't at this point.
132. 08:20It was probably a month ago.
133. 08:21Um they purchased a subscription plan and they just highlighted right how different what Frontier Labs charge customers for first-party products are versus what they sell.
134. 08:33It's a bad deal.
135. 08:34Don't play this game or try and let me know how you win.
136. 08:38I don't recommend.
137. 08:41Think about everyone else.
138. 08:43Think about what that structure means and where you have expertise.
139. 08:48I don't think that that's winning on the token economics.
140. 08:51I think it's about product.
141. 08:54It's about building data flywheels and understanding your customers better than anyone else.
142. 09:00Understanding when you need capability, when you need low price, when you need latency improvements.
143. 09:06I promise you, you don't always need what is usually the slowest but the most capable model out there.
144. 09:12and then build compelling UI and orchestration.
145. 09:15And I'll show you some examples of that to justify the cost on the bad deal tokens that you do resell.
146. 09:23The job is not to train.
147. 09:25I mean, some of you might be training the best model and I'd love to serve it and come talk to me afterwards, but most of you are not doing that.
148. 09:31Stop trying to win that game and think about the best product that uses many models.
149. 09:36Help your customers, help your team.
150. 09:39Bet on the frontier, not on the lab.
151. 09:42and we'll talk about what it looks like to do that.
152. 09:45This cost per capability per second trade-off is actually really intense.
153. 09:50Um, Citadel came out with this memo um a while ago, maybe two weeks ago.
154. 09:55I loved it.
155. 09:56The idea is that for the economy at large, simpler models might be the most cost-effective productivity augmenting pathway.
156. 10:04They talk about this bifurcation on frontier versus everyday usage.
157. 10:08I really believe that.
158. 10:09And for every product, the definition of frontier versus everyday, the definition of saturated capabilities or model capability overhangs depends on your expertise on your product.
159. 10:21No one can replace that.
160. 10:24And not all traffic is equal.
161. 10:26It is a huge miss to send all of these to the latest opus model.
162. 10:30Some of these absolutely large scale data analysis when you do it on notion will recommend Opus, right?
163. 10:36When you triage an email inbox, if we're charging you to do that on Opus, we're ripping you off and ourselves.
164. 10:44Think about where your traffic patterns are.
165. 10:47And then think about how Frontier Lab model providers are structured today.
166. 10:51I mean, it's functionally an oligopoly, right?
167. 10:54And that's fine because they're racing to the top.
168. 10:57And I think the top is really hard and really important.
169. 11:00This is not to say that products don't have a place for frontier difficult tasks.
170. 11:04I want everyone to nod and understand that's not what this talk is about.
171. 11:08Understand when you need those tasks and it's not everything.
172. 11:11The problem with those tasks are is keep in mind how pricing is incentivized.
173. 11:16You can figure out who these players are.
174. 11:18Either you are the best model.
175. 11:21Everything above what AI can't do today is your market.
176. 11:24You can basically price it as high as you kind of want.
177. 11:28If you're slightly behind that best model, all you need to be is like a dollar per million tokens cheaper and you have the rest of the market.
178. 11:36You know that economic theory about gas stations where the best gas stations are the ones that are right next to each other because they cover east and west the most.
179. 11:43Yeah.
180. 11:44It's the same with model pricing, which means that price does not correlate with capability growth.
181. 11:53So for this complex task, understand what capabilities you need, but be the expert on what complexity is.
182. 12:01And keep in mind that who handles complexity changes.
183. 12:05Um, oftentimes you'll see applied AI companies really be super outspoken on marketing with a specific lab.
184. 12:11That's always kind of a red flag for me when they're not model agnostic because if you look at this graph, it basically shows that they're behind every month, right?
185. 12:20the new model and the new model provider of the best Frontier capabilities change and if you hit your ride with one particular provider in exchange for for instance a larger discount
186. 12:32um you're doing a disservice to your customers like half of the time right so really think about if that discount is worth not actually having a Frontier product and remember that that optionality
187. 12:44is your leverage if you don't have the capability to walk at any point you are stuck.
188. 12:51And again, I think that's probably the most expensive decision you'll make regardless of what discount you get or the engineering work to have model interoperability.
189. 13:01One option to navigate this is stay model agnostic.
190. 13:05Have different models and capabilities in your system so that at any point if pricing seems unfair or untenable, you are not out of business.
191. 13:15Notion's auto model does this really well.
192. 13:18We have state-of-the-art models available always.
193. 13:21Um, but we also have an auto model there at the top that handles about 75% of our traffic.
194. 13:27Right?
195. 13:29We have the ability to switch between models in our product and we also offer it to our customers so that they have access to these models without vendor lockin.
196. 13:36That's part of our AI Switzerland approach.
197. 13:41You guys love taking photos of slides.
198. 13:43This is the slide.
199. 13:44Okay.
200. 13:46Model agnostic playbook.
201. 13:47This is how you do it.
202. 13:48Build for multimodal.
203. 13:50It is hard to kill the cache and switch models mid-transcript.
204. 13:53I understand that we invest in that technology.
205. 13:56It doesn't even have to be per thread.
206. 13:58Just think about your harness as model interoperability.
207. 14:02Think about the cost per capability per second, not just the tokens.
208. 14:06Here's a great example.
209. 14:07We posted this review when we um announced our partnership with Parallel as our web search provider.
210. 14:13If you were to look at just latency of a single call or just cost, parallel might not be the cheapest.
211. 14:19But if you have expertise in entire web search trajectories, you'll see how it differs.
212. 14:24The granularity of this eval is what lets us make the best decisions for our customers because we understand all of the trade-offs on entire trajectories, not just single calls.
213. 14:35Switch fast and often.
214. 14:36I think we talked about that.
215. 14:38And give them something back.
216. 14:39That expertise on use cases is also very valuable to Frontier Labs.
217. 14:43We find that our eval program partnerships actually help us a lot with Frontier Labs and is something that we can exchange instead of extraordinarily large commits and I don't think the discount is ever worth the loss in optionality.
218. 14:58That's a perspective you can choose to keep or not.
219. 15:01The second option is moderate tasks understanding open weights place there.
220. 15:06um openweight models are really strong enough to handle these tasks and the possibility to RL on top of them has also kind of expanded the upmarket growth that they can cover.
221. 15:17I view openw weight models as basically lowering the barrier to entry on cost for our customers and they also give you negotiation leverage.
222. 15:25So it's kind of a credible alternative that's putting that downward pressure on pricing that if there's an igopoly of two or three providers at the top is unavailable right now otherwise.
223. 15:35I think Kimmy 26 was probably the first time that we really saw a model that outperformed 52, GPT 52, GLM 52 now is another 52 bombshell in the villa that also probably does best here.
224. 15:49But it's no longer the case where openway models are good for just SFT on small tasks.
225. 15:54Really think about without RL if they're capable enough for what you need.
226. 16:00Um, and again, don't just think about external benchmarks.
227. 16:06Be able to have expertise on your system.
228. 16:08What are your tool errors?
229. 16:10What's the actual latency that you need?
230. 16:12Right?
231. 16:13Here's an example of a benchmark that we posted.
232. 16:15It's a little bit stale on purpose, right?
233. 16:18But you get the idea.
234. 16:21Philip at BE 10 this showed this slide once and I've stolen it ever since.
235. 16:25Um, thank you.
236. 16:28Are you here, buddy?
237. 16:29Okay.
238. 16:29Well, chat, hi.
239. 16:31Um, well, he could come up and say it better, but the idea is that you don't have to be at the top, right?
240. 16:37I'm not trying to make a case that open weight is the best model out there.
241. 16:41Um, the case being made, however, is that um the gap gets covered eventually.
242. 16:48So, if the tasks that you're having today are good enough, then in six months, they're probably covered by open weight.
243. 16:54So, be prepared now.
244. 16:57And the last thing is CPUs over GPUs.
245. 17:00Um, we've we've recently launched something at notion called workers.
246. 17:03I don't think that the GPU is necessary for every job.
247. 17:07A lot of the jobs that we have are actually serving um discrete pieces of code.
248. 17:12Like you don't need an LLM to turn a CSV into a PDF.
249. 17:16You don't need an LLM to talk to notion tool calls if we have a CLI.
250. 17:19You definitely don't need an LLM to do deterministic SQL queries.
251. 17:23This is where people become token poor very quick.
252. 17:28And I think the last option here besides openw weight CPUs and optionality is actually governance.
253. 17:34Um there's a lot of AI governance.
254. 17:36Um one is visibility.
255. 17:38Um understanding who's using the data, understanding its maintainability and control.
256. 17:44When you have model optionality, you can offer a lot more to your customers.
257. 17:48Um here's an example of how that governance works in notion.
258. 17:53So final tips again, think about architecture, think about open weight, and build value that transcends tokens.
259. 18:00So we're going to depart Token Town.
260. 18:02I know I said welcome to Token Town.
261. 18:04We're going to spend the next 10 minutes really thinking about what to do next.
262. 18:09So I think the challenge of the next six months doesn't have to do with capabilities.
263. 18:13I think it has to do with security.
264. 18:14Let's start there.
265. 18:16There's this concept called the lethal trifecta.
266. 18:18Simon Wilson, I think, crafted this.
267. 18:21If you have access to private data, exposure to untrusted content, whether it be through ingestion, MCP, email, right?
268. 18:29And the ability to ex to communicate externally, and that can include like payloads in a web search.
269. 18:34The second you have that system, you're exposing risk.
270. 18:37And in fact, the more autonomous your system is, the more unsupervised this risk is.
271. 18:42I think that this is what builds valuable product, not just capability.
272. 18:48Same with sandboxes and computers.
273. 18:50We talked about this, but it really is something that builds better determinism in your product and also better token economics for your customers in multi-agent orchestration.
274. 19:03Understanding what agents see and do and what persists.
275. 19:06I think persistence of enterprise knowledge is something that's actually really not discussed enough.
276. 19:13It's starting to be with some recent launches, you know.
277. 19:17Oh, there is audio.
278. 19:19So don't have your workflows look like this.
279. 19:21And I think this is where most software factories are today, right?
280. 19:24It's like actually your entire engineering time just spends time babysitting the factory, right?
281. 19:30I mean, I get it.
282. 19:31Ours started off like this.
283. 19:33Agent orchestration is one of the most difficult tasks of making factories work.
284. 19:38So okay, this is me telling Tay to tell people to buy notion AI.
285. 19:42And the reason I included this slide is I am going to sell notion for a second.
286. 19:45It's my job.
287. 19:46Always be closing.
288. 19:47Always be selling.
289. 19:48Always be hiring.
290. 19:48come find me.
291. 19:49But I'm going to talk for a second about how notion does this today.
292. 19:54We already have the ability to inspect tasks.
293. 19:57And you can imagine any task that you look at um in a notion document.
294. 20:03You can have Claude actually go ahead and scope out what you need.
295. 20:07We've launched this manage agent capability today.
296. 20:10So if I go ahead to the top of this task, I can actually ask Claude agent to scope out the task.
297. 20:15Right?
298. 20:18Ideally, it's working.
299. 20:20Um, and you'll see it'll actually populate um an entire spec of what needs to be done.
300. 20:27In this example, it's not ready.
301. 20:28It's going to ask me a question.
302. 20:30Keep in mind, this isn't a markdown file.
303. 20:32This is an active document.
304. 20:34Um, let's say I don't actually know the question and I go ahead and I ask my team um what to do.
305. 20:40Imagine that you can kind of tag in your team into these systems.
306. 20:44MJ's our PM.
307. 20:50So in this example, she doesn't know.
308. 20:51Usually she does, but multi-agent orchestration is important.
309. 20:55Maybe Claude Code isn't the best at customer voice, but Decagon is, right?
310. 21:01You can ask Decagon agents, we're proud partners of them as well, to collect the right data that you need.
311. 21:08Okay, in this example, we think we know enough.
312. 21:13We're going to go ahead and actually um iterate through some of this flow.
313. 21:16I'm going to skip ahead a little bit.
314. 21:18We asked our TL what we needed.
315. 21:20He replied again.
316. 21:21It's a collaborative file, not just a markdown.
317. 21:24And we can have Claude actually go ahead and spin up the PR.
318. 21:27Hopefully, this is looking a little familiar now.
319. 21:29This is kind of the vision of software factories.
320. 21:31It's what we're trying to host.
321. 21:33Okay, Claude put up a PR.
322. 21:36Maybe that's not enough.
323. 21:37Um maybe I want to go ahead and ask Codeex what it thinks.
324. 21:48Great.
325. 21:48Found two issues.
326. 21:51You can think about this scaling in an actual factory.
327. 21:54So today in notion, you're actually able to orchestrate these agents together and you're not committing to a lab.
328. 22:01You're committing to the concept that AI is augmenting and automating what you do.
329. 22:08This is real.
330. 22:09I asked Rejieve if I could post this.
331. 22:12This is how it works today internally at notion.
332. 22:14Almost all of our polish and large feedback like this is actually coordinated um through our software factories both in terms of writing to the right teams and also having coding agents take the first step.
333. 22:26Verscell does this as well from staging to shipping to closing.
334. 22:33And we see massive ROI gains from our customers.
335. 22:37That's over three minutes saved on a given task.
336. 22:39Imagine that at scale.
337. 22:40So I think we're trying our hardest to think about the factory lens.
338. 22:44We cannot do this without optionality and we cannot do this without conviction that we understand what models are required for which tasks.
339. 22:53It's really wild out there you guys.
340. 22:56I get it.
341. 22:57The market is really young.
342. 22:58It's exceptionally opaque.
343. 23:00It's moving fast.
344. 23:00I'm super grateful for communities like AI engineer to bring us together and like talk openly about these things and how we navigate it.
345. 23:08Um I think we owe it to all of our customers to get it right and to be critical thinkers about how we navigate this together.
346. 23:15I'm chronically online fortunately.
347. 23:18Um you can always DM me on Twitter, you can email me, you can find me after this.
348. 23:22Um but thank you for yapping with me and thinking about this problem and have a good day.
349. 23:32[music]
350. 23:38[music]