# cloudflare is a neolab Source: https://www.youtube.com/watch?v=ZZ19zhYnDKY Machine-generated transcript; recognition errors remain. Source-recording timestamps. Boundary-crossing subtitle ends are clipped to this session. [00:40:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2409s) Hi everyone, my name is Michelle Chen. [00:40:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2411s) I'm so excited to be here today at AI [00:40:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2413s) Engineer Code Summit chatting a little [00:40:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2414s) bit more about why Cloudflare became a [00:40:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2417s) Neolab. So at Cloudflare I am the group [00:40:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2419s) PM here. I run on the product team [00:40:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2421s) working on products like workers AI, AI [00:40:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2424s) search, AI gateway as well. And so today [00:40:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2426s) we're going to go through this little [00:40:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2427s) journey of how Cloudflare came to be. [00:40:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2429s) You might have heard that we launched [00:40:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2430s) and trained some models recently. I'm [00:40:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2432s) going to tell you a little bit more [00:40:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2433s) about why and what you can be doing with [00:40:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2436s) it. And so we're going to go through a [00:40:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2437s) little bit of a philosophical journey. [00:40:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2439s) Uh when I was preparing for this talk, I [00:40:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2441s) wanted to use this quote and then [00:40:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2443s) Anthropic like launched a blog post last [00:40:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2445s) weekend and used my quote, but I'm still [00:40:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2447s) going to use it anyways. Um so the quote [00:40:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2450s) goes [00:40:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2453s) in one second. Um the quote is from [00:40:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2456s) Kirkagard and it says, "Life can only be [00:40:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2458s) understood backwards, but it must be [00:40:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2459s) lived forwards." And I think this is [00:41:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2461s) really really apt quote here for us [00:41:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2463s) because at Cloudflare we're a [00:41:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2465s) 16-year-old company and we've done so [00:41:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2467s) many things over the years but somehow [00:41:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2469s) we got all some of the pieces right and [00:41:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2471s) we ended up having to become a Neolab. [00:41:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2473s) So we're going to go through a little [00:41:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2474s) bit journey. The first thing we're going [00:41:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2475s) to do is look backwards and talk through [00:41:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2480s) why Cloudflare had to uh became a [00:41:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2483s) Neolab. [00:41:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2486s) Give me one second here. [00:41:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2489s) Perfect. We're going to go backwards [00:41:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2491s) [laughter] and talk through why [00:41:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2492s) Cloudflare became a Neolab. So, [00:41:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2495s) Cloudflare is a 16-year-old CDN company. [00:41:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2498s) We just celebrated our 16th birthday uh [00:41:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2501s) at birthday week this year uh in [00:41:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2504s) September. [00:41:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2506s) And we first started as a CDN. So, as a [00:41:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2508s) CDN company, what you have to do is you [00:41:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2510s) have to serve content as quick to your [00:41:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2512s) users as possible. you sit between the [00:41:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2515s) users request and the origin and you [00:41:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2517s) help uh sit in the middle and you can do [00:41:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2519s) things like cache and protect do [00:42:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2521s) security and things like that. So that's [00:42:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2522s) kind of how Cloudflare started. But [00:42:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2525s) Cloudflare's infrastructure had to be [00:42:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2527s) adaptable in order to become the CDN. [00:42:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2530s) And so what we did was 16 years ago in [00:42:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2532s) 2010, we launched with our own [00:42:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2533s) infrastructure. Our fast CDN had to be [00:42:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2536s) everywhere. And this looks really [00:42:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2537s) different than a traditional hyperscaler [00:42:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2539s) where compute is centralized in US- East [00:42:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2541s) one or US West one. Cloudflare's [00:42:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2543s) infrastructure is a completely different [00:42:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2545s) paradigm. We have data centers all [00:42:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2547s) around the world as close to the user as [00:42:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2549s) possible. And this is really important [00:42:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2551s) in becoming a CDN. But now we have these [00:42:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2554s) computers basically running as close to [00:42:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2556s) the user as possible. So we thought to [00:42:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2557s) ourselves, what else can we be doing [00:42:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2559s) with these computers that people would [00:42:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2560s) want to build with? And so came. We [00:42:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2564s) started launching these compute [00:42:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2566s) primitives as well. And so in 2017, we [00:42:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2568s) launched a product called workers, which [00:42:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2569s) is basically like serverless functions [00:42:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2572s) that you can run. It contrasted VMs at [00:42:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2574s) that time where you had to rent a whole [00:42:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2576s) machine and pay for the whole machine. [00:42:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2578s) Either you're using it or not using it. [00:43:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2580s) And so with workers, you just execute [00:43:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2582s) some code and then the worker goes away. [00:43:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2584s) You don't have to pay for it. It takes [00:43:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2585s) milliseconds to start up, not seconds as [00:43:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2587s) well. And then we followed with other [00:43:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2589s) compute primitives like durable objects [00:43:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2591s) which are more stateful. And then in [00:43:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2592s) 2025, we launched containers as well [00:43:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2594s) because there are some workloads that do [00:43:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2596s) need something like a VM and we want all [00:43:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2598s) of those workloads to be able to run on [00:43:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2599s) Cloudflare. So now you know 2017 start [00:43:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2602s) of our developer platform we have these [00:43:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2604s) compute primitives people started using [00:43:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2606s) it people started building more like [00:43:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2608s) static websites and things like that so [00:43:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2611s) we thought to ourselves how do we [00:43:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2612s) elevate this and take it to the next [00:43:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2614s) level so we introduced storage as well [00:43:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2618s) and so we have these storage primitives [00:43:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2619s) on cloudflare's developer platform we [00:43:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2621s) have KV we have R2 do object storage we [00:43:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2624s) have databases like D1 as well and the [00:43:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2626s) idea here is that people should be able [00:43:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2628s) to build full stack applications [00:43:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2630s) all on Cloudflare stateful applications [00:43:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2633s) where people can go on and start [00:43:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2634s) building um full applications and web [00:43:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2638s) apps and platforms on Cloudflare's [00:44:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2640s) developer platform. So we have this [00:44:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2642s) infrastructure, we added compute [00:44:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2644s) primitives to it, we added storage [00:44:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2645s) primitives to it and then came 2023. And [00:44:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2649s) if you guys remember what happened in [00:44:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2650s) 2023, chat GPT kind of blew up a little [00:44:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2653s) bit and it changed how we think about [00:44:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2655s) AI, right? And I remember it being [00:44:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2659s) birthday week and uh three weeks before [00:44:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2661s) birthday week a lot of us sat together [00:44:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2662s) in a room and we were like what should [00:44:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2664s) we do now that chat GPT is on the rise? [00:44:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2666s) What can Cloudflare play in here with [00:44:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2668s) our infrastructure and our primitives in [00:44:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2671s) order to be in this AI game? So in 2023 [00:44:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2675s) we launched a product called workers AI [00:44:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2677s) where we added GPUs to our network. And [00:44:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2679s) so again this is really distributed [00:44:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2681s) where GPUs are running as close to the [00:44:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2683s) user as possible. It's not centralized [00:44:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2685s) like a US East1 hyperscaler for example, [00:44:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2687s) but it's running as close to the edge as [00:44:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2689s) possible. And we had a hypothesis when [00:44:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2692s) we started workers AI. Our hypothesis [00:44:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2694s) was that models were going to get [00:44:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2696s) smaller, so that network latency was [00:44:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2698s) going to become more and more important. [00:45:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2701s) We launched with small models that could [00:45:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2702s) fit on our small hardware around the [00:45:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2704s) world that were super distributed. And [00:45:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2706s) we thought that people would have models [00:45:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2708s) get better and faster and smarter and [00:45:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2710s) that network latency was going to play a [00:45:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2712s) much more bigger impact here. [00:45:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2715s) Well, uh actually the opposite happened [00:45:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2718s) and models got larger and larger. So we [00:45:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2720s) went from something like 7 billion [00:45:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2722s) parameters which was frontier scale at [00:45:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2724s) the time to being the three trillion [00:45:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2726s) parameter models that were getting [00:45:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2727s) launched. And I think this is not so [00:45:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2730s) shocking when you look back in time [00:45:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2732s) because you know as we're on the way to [00:45:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2734s) AGI um if you're following scaling law [00:45:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2737s) as you train on more data as the models [00:45:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2739s) get larger they're more generalizable [00:45:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2741s) and so on the path to AGI it makes sense [00:45:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2743s) that we're pushing the frontier of [00:45:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2744s) research here in order to get larger and [00:45:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2747s) larger and smarter models. [00:45:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2749s) So this goes against our hypothesis [00:45:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2752s) right that the models were going to get [00:45:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2753s) smaller. the models actually ended up [00:45:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2755s) being larger and the network latency [00:45:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2756s) mattered less and less. So what did we [00:45:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2758s) do? [00:46:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2761s) We ended up [00:46:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2763s) pivoting. So cloud got really good at [00:46:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2765s) serving these giant models as well. So [00:46:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2767s) in March 2026, we launched Kimmy and [00:46:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2770s) then we followed up with GLM and [00:46:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2771s) DeepSeek and these large trillion [00:46:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2773s) parameter models um hosted on our [00:46:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2775s) infrastructure. And so this is [00:46:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2776s) serverless. People hit an API call and [00:46:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2778s) get a token back. It actually took us a [00:46:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2780s) while in order to decide to do this, [00:46:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2783s) right? We March 2026 is actually pretty [00:46:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2786s) late in the game to to serving these [00:46:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2787s) model, but we were trying to make a [00:46:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2789s) conscious decision here on does this [00:46:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2791s) actually make sense for us to be serving [00:46:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2792s) it. It didn't really fit with our [00:46:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2794s) infrastructure story, but it fit very [00:46:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2796s) well with the developer platform. [00:46:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2799s) So in 2026, we had all these primitives [00:46:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2802s) in place. We had the infrastructure [00:46:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2804s) which is unique to Cloudflare running [00:46:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2806s) close to users. We had compute [00:46:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2808s) primitives where people were building [00:46:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2809s) agents on top of so workers and durable [00:46:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2811s) objects right as your agent needs to [00:46:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2813s) complete a task it goes on and it spins [00:46:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2816s) and does a task and then sleeps when it [00:46:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2817s) doesn't. We had storage primitives so [00:46:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2819s) your agents can actually interact with [00:47:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2822s) data and storage and then we had models [00:47:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2824s) these large models that were powering [00:47:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2825s) agents. And so all of this fit into the [00:47:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2828s) Cloudflare agent cloud story. And so we [00:47:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2830s) started telling this narrative that we [00:47:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2832s) want to be the agent cloud. We want [00:47:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2834s) agents to be deployed on us powered by [00:47:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2836s) models on us using our compute and data [00:47:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2838s) primitives all built on our [00:47:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2839s) infrastructure and our network. But it [00:47:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2841s) never felt quite right. We felt like we [00:47:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2844s) were always missing a piece here because [00:47:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2846s) the infrastructure pattern didn't [00:47:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2848s) necessarily fit in with what workers AI [00:47:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2850s) was designed for. [00:47:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2853s) And so I'm going to sit here and give [00:47:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2855s) you my second philosophical quote which [00:47:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2856s) is a risk considering that I'm a girl in [00:47:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2859s) Chinese and AI, but I'm going to give [00:47:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2861s) you this Chinese quote anyways. Um, [00:47:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2863s) [laughter] [00:47:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2864s) and uh, it basically says, [00:47:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2868s) which means everything is ready here. [00:47:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2870s) All that's missing is the east wind. [00:47:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2873s) And one of the things uh about this is [00:47:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2875s) like you can't really predict where the [00:47:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2877s) wind is going to come from or how it's [00:47:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2879s) going to come, but you need to be very [00:48:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2880s) reactive when it happens. And so in [00:48:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2882s) September 2026, the east wind came. [00:48:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2885s) Typesafe launched a model called Jev and [00:48:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2888s) introduced this idea of system one [00:48:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2889s) models and it contrasted what we [00:48:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2891s) typically knew with LLMs which were auto [00:48:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2893s) reggressive and taking a long time to [00:48:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2895s) generate tokens. It scales quadratically [00:48:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2898s) etc. A system one model is a completely [00:48:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2900s) different paradigm where it actually [00:48:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2902s) just takes one forward pass through the [00:48:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2904s) inputs and generates an output that you [00:48:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2907s) can use to make decisions with it. [00:48:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2910s) And so I really think that this came [00:48:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2911s) down to perfect timing. I think people [00:48:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2913s) were starting to get fed up a little bit [00:48:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2915s) with LLMs because they were so costly. [00:48:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2917s) It cost so many tokens. It's s time [00:48:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2920s) consuming. And I don't know about you, [00:48:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2921s) but when I read a reasoning trace, [00:48:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2923s) sometimes my blood pressure rises when [00:48:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2924s) the LLM doesn't get the right the things [00:48:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2926s) right. And so I think we were sending [00:48:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2928s) everything to LLMs and started getting [00:48:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2929s) frustrated with it. And so when this [00:48:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2931s) system one and Jev paradigm started [00:48:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2933s) coming on board, we were like this makes [00:48:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2935s) so much sense for us to adopt, right? We [00:48:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2937s) don't have to send every request to an [00:48:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2939s) LLM. We can send some of it to a system [00:49:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2941s) one model. And so the system one models [00:49:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2944s) actually end up taking work off of the [00:49:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2946s) giant LLMs and you give it to the small [00:49:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2948s) fast models and the two are supposed to [00:49:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2949s) work in tandem. They're supposed to [00:49:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2951s) complement each other and get to the end [00:49:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2953s) goal faster. And so if you look we are [00:49:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2956s) back at these small and fast models [00:49:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2958s) right this was the original workers AI [00:49:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2961s) hypothesis [00:49:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2963s) and so Cloudflare was not wrong here we [00:49:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2966s) actually were just early with these [00:49:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2968s) small and fast models the network [00:49:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2970s) latency became important again we want [00:49:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2972s) the model to be responding fast but also [00:49:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2974s) the network latency to be even faster [00:49:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2976s) and so we decided to do something here [00:49:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2979s) but Jev was closed source cloudflare [00:49:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2983s) couldn't actually host the Jev models [00:49:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2984s) and take advantage of our [00:49:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2985s) infrastructure. So what did we decide to [00:49:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2988s) do? We built our own version of Jev. So [00:49:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2991s) Jev launched on September 15th. Um on [00:49:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2994s) September 18th, just three days later, [00:49:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2995s) we started experimenting and we launched [00:49:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2997s) our version of Jev. Uh our first [00:49:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=2999s) experiment of it. And then it took us a [00:50:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3002s) week of deliberation to decide do we [00:50:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3004s) actually want to be going into this [00:50:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3005s) Neolab territory. It's completely new to [00:50:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3008s) us. We don't know what we need to know [00:50:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3010s) it. you know, are we going to set the [00:50:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3012s) expectation that we're want to be a [00:50:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3013s) Neolab and are we going to live up to [00:50:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3015s) those expectations? But we decided to do [00:50:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3017s) it. I remember on a Friday evening, we [00:50:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3019s) were like, let's just go for it. And we [00:50:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3021s) did it. And we ended up training the [00:50:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3023s) model over 48 hours over the weekend. [00:50:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3025s) And then we launched it the following [00:50:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3026s) week. And so we ended up with a model [00:50:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3028s) called KFF. [00:50:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3030s) KFF gave power back to the agents to be [00:50:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3033s) able to make decisions on their own. So [00:50:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3034s) agents could ask act more autonomously. [00:50:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3037s) You didn't need to have a LLM reason [00:50:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3039s) through for minutes on if it should make [00:50:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3041s) this decision or not. You also didn't [00:50:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3043s) need to put a human in the loop and be [00:50:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3044s) interrupted by like a human's sleep [00:50:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3046s) schedule when you want your agents to be [00:50:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3048s) running 24/7. So you can now [00:50:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3049s) programmatically in code decide make [00:50:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3052s) decisions for your agent to take action [00:50:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3054s) on. [00:50:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3057s) So these decision models countered LLM [00:50:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3059s) models because they were a lot faster [00:51:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3061s) and they were also a lot cheaper as [00:51:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3062s) well. And so it actually as an inference [00:51:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3064s) provider takes a lot of pressure off the [00:51:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3066s) LLMs as well because the LLMs were [00:51:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3068s) reserved for work for meaningful work [00:51:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3070s) whereas the cleft and uh decision models [00:51:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3073s) could be used for fast um intuitive [00:51:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3075s) thinking and work right [00:51:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3078s) and so we also went back and thought to [00:51:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3080s) ourselves you know what else do agents [00:51:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3083s) and people use for LLMs for that are [00:51:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3086s) overkill. So we came up with decisions [00:51:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3088s) and um there must be other workloads out [00:51:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3091s) there that are unsuitable for an LLM [00:51:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3093s) both because they are either too [00:51:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3095s) expensive or take too long or maybe that [00:51:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3097s) it just would get the answer wrong [00:51:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3099s) because the auto reggressive nature of [00:51:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3100s) an LLM means that it would hallucinate. [00:51:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3103s) So we also launched a model 3 weeks [00:51:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3106s) later called cleric. We actually ended [00:51:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3107s) up launching four models in the span of [00:51:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3109s) three weeks. So I think we did pretty [00:51:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3111s) good job as a new lab. Um but cleric is [00:51:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3113s) a new model paradigm. So it's completely [00:51:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3116s) new. We're calling it an extraction [00:51:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3117s) model, but it's a system one model in [00:51:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3119s) that in the sense that it is just as [00:52:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3121s) fast and just as cheap, but it is also [00:52:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3123s) doing like one forward pass through the [00:52:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3125s) user. And the purpose of cleric is when [00:52:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3127s) you're interacting with the world, [00:52:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3128s) you're not interacting in perfectly [00:52:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3130s) typed JSON objects, right? Sometimes [00:52:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3132s) you're getting an email, a tweet, a [00:52:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3134s) slack message, but those people are not [00:52:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3136s) filling out perfect forms for you to [00:52:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3137s) extract inputs in out of. So what cleric [00:52:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3140s) does is it can take unstructured input [00:52:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3142s) and it can extract, compute, estimate [00:52:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3145s) and forecast values from those inputs [00:52:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3147s) and it gives you back a perfectly typed [00:52:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3149s) JSON objects which means the agents are [00:52:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3151s) able to do things with it. So we're kind [00:52:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3153s) of giving agents the ability to read [00:52:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3155s) here and take action with cleft as well. [00:52:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3160s) And so in code again you're giving [00:52:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3162s) cleric you know typed values from [00:52:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3164s) unstructured data. You can pass it in [00:52:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3166s) whatever kind of unstructured data. It [00:52:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3167s) can also take image input as well. And [00:52:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3169s) so you can just tell it to read [00:52:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3171s) basically what's in text here, pass it [00:52:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3173s) to KFT, decide what to do with this [00:52:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3175s) text, and then pass it to an LLM to [00:52:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3177s) decide what to do with the agent. [00:53:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3181s) And so Cloudflare had to become a neolab [00:53:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3183s) here. This was our missing bit. We had [00:53:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3185s) the infrastructure. We had the compute [00:53:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3187s) primitives. We had the data primitives. [00:53:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3188s) We had the large models powering the [00:53:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3190s) agents. But in order to take advantage [00:53:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3192s) of all of that and our infrastructure [00:53:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3194s) that we were built on, we had to be a [00:53:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3196s) new lab and ship these system one models [00:53:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3198s) that fit perfectly with our [00:53:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3199s) infrastructure. And I think one of the [00:53:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3200s) more important things is that we're not [00:53:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3202s) limited to this. Being a new lab means [00:53:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3204s) that we can create what we are missing, [00:53:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3207s) right? It means that we're not uh if [00:53:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3209s) we're missing a system one model because [00:53:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3211s) we're um at risk of it things being [00:53:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3213s) closed sourced or regulatory risks or [00:53:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3215s) things like that, we can actually create [00:53:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3217s) and invent things. So I think being a [00:53:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3218s) Neolab has actually opened a lot of [00:53:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3220s) doors for Cloudflare to do more in the [00:53:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3222s) AI world. [00:53:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3225s) So we went through backwards. We went [00:53:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3226s) through a little bit of a history [00:53:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3227s) lesson. I want to go forwards as well. I [00:53:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3229s) want to teach you guys what you guys can [00:53:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3231s) be doing with KF and Cleric. And I'm [00:53:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3233s) cognizant that I am at AI engineer code [00:53:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3235s) summit. So in a in a room full of [00:53:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3237s) developers and engineers, how can you [00:53:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3239s) guys leverage KFF and cleric and system [00:54:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3241s) one models in order to better use your [00:54:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3243s) workflows? [00:54:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3245s) So we'll do a quick demo here. Um it is [00:54:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3248s) an let's think you have an agent you're [00:54:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3250s) triaging like a CI build that went wrong [00:54:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3252s) right um let's say you have 40 tool [00:54:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3255s) tools that are available to this agent [00:54:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3257s) you have your PR that went wrong um and [00:54:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3259s) then you have your agent maybe it's [00:54:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3260s) running PI durable on a durable object [00:54:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3262s) says all of this is built on Cloudflare [00:54:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3265s) so this is the setup here tech typically [00:54:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3267s) what you would do is you would take an [00:54:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3269s) LLM and in the LLM would take a tool [00:54:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3271s) make a tool call the tool call would go [00:54:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3272s) back to the LLM the LM would make [00:54:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3274s) another tool call etc etc until the [00:54:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3276s) issue is resolved So this is like five [00:54:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3278s) back and forths with GLM 5.3. You're [00:54:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3281s) sending 40 tools every single time and [00:54:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3284s) maybe you're blowing up the context, [00:54:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3285s) right? Because you're sending the same [00:54:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3287s) kind of inputs back and forth to the [00:54:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3288s) LLM. And so what if we augment this with [00:54:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3291s) system one models [00:54:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3294s) with system one models? Uh this could [00:54:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3296s) look a lot different. And so let's say [00:54:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3298s) we were we want speed. So we pick cleft [00:55:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3300s) flash. Cle can actually help you prune [00:55:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3302s) your tool list. Maybe it only needs 10 [00:55:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3304s) out of the 40 tools that are available. [00:55:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3306s) So you only need to pick 10 of those. [00:55:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3308s) Then let's say you take cleric and you [00:55:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3310s) use it to extract the values from the [00:55:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3312s) input. So it can get the PR number [00:55:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3314s) without reasoning and thinking about it [00:55:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3315s) too long. Um and then let's say you just [00:55:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3318s) you know orchestrate some code here to [00:55:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3319s) execute these API calls and uh you can [00:55:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3322s) actually just you know get those answers [00:55:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3324s) back. Then you ask cleric again um to [00:55:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3326s) pull the data from the logs. So normally [00:55:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3329s) when you're using with working with an [00:55:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3330s) LLM you probably just paste in all of [00:55:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3332s) the logs right like all 300 600 lines of [00:55:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3334s) logs. But Cleric can actually just go [00:55:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3337s) through and read those and say, "Here's [00:55:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3338s) the exact error. Here's the error [00:55:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3340s) message. Here are the lines of code it's [00:55:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3342s) affecting." And you're basically [00:55:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3343s) shrinking your context a lot. And then [00:55:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3345s) finally, you use an LLM to actually do [00:55:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3348s) take action and um make decisions and [00:55:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3351s) take action and extract values [00:55:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3353s) altogether. This same agent can run the [00:55:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3355s) system one. Let's do a little bit of a [00:55:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3357s) live demo. Let's see how it goes. We're [00:55:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3359s) going to have to click run both here. [00:56:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3361s) And this is a live demo. And uh I will [00:56:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3363s) say that [00:56:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3365s) Cloudflare our workers AI is uh has a [00:56:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3369s) very dal usage pattern. So we're in the [00:56:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3371s) middle of the workday. So we'll see how [00:56:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3372s) this how this goes. But it'll give you [00:56:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3374s) some illustrative example here of how it [00:56:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3376s) goes. Uh and we're also dealing with [00:56:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3378s) conference Wi-Fi. So so bear with me [00:56:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3380s) here. All right. The system one model [00:56:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3382s) side finished already and the all LLM [00:56:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3385s) one is is going. So I think even [00:56:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3387s) illustratively you can see that the uh [00:56:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3390s) system one set of models is a lot [00:56:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3392s) faster. Even by just this simple [00:56:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3394s) waterfall chart, you can see that the [00:56:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3395s) system one part takes so much less time [00:56:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3398s) than the uh actual generation which is [00:56:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3401s) the purple bars here. And I think they [00:56:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3403s) both ended up getting to the answer at [00:56:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3404s) the end. So the quality was not [00:56:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3406s) affected, but it took a lot less time. [00:56:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3409s) It was a lot cheaper. Six LLM calls [00:56:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3411s) versus one. 40,000 LLM tokens versus [00:56:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3414s) 1,800. And so you can see how you can [00:56:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3417s) use these kind of system one and system [00:56:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3419s) two models even in like developer [00:57:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3420s) workflows in order to get you to your [00:57:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3423s) answer faster. [00:57:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3426s) So the same answer but it took a [00:57:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3428s) fraction of the LLM time and cost. [00:57:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3430s) Right? I ran this when we were not doing [00:57:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3433s) conference Wi-Fi but we had about 9.2 [00:57:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3435s) seconds and 17.2 seconds on the LLM. It [00:57:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3438s) was almost two times faster. It was 72% [00:57:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3442s) less expensive. We shrunk our LLM calls [00:57:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3445s) from five down to one and it was about [00:57:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3447s) 94% less tokens. And so this is an [00:57:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3451s) illustrative example of how you could be [00:57:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3454s) using these system one and system 2 [00:57:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3455s) models in tandem. It's not intended to [00:57:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3458s) completely replace your LLM usage, but [00:57:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3460s) it can augment it and make it better by [00:57:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3462s) making allowing you to free up time and [00:57:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3465s) compute for your other more important [00:57:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3467s) tasks here. And you can h now do you [00:57:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3470s) know higher concurrency tasks because [00:57:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3471s) these models are able and capable to do [00:57:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3474s) so much more. [00:57:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3476s) And so that was one illustrative [00:57:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3477s) example. I just wanted to flash on the [00:57:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3479s) screen really quickly as well. What else [00:58:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3480s) are people building with cleft and [00:58:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3482s) cleric right? There's a lot of um [00:58:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3484s) routers that you can build and in making [00:58:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3486s) decisions there. Um there's fraud that [00:58:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3488s) we've been seeing a lot like spam or [00:58:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3490s) fishing. It can make decisions on there. [00:58:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3492s) With Cleric, you can use it to extract [00:58:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3494s) data from like web pages or support [00:58:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3496s) tickets and things like that. So, I'm [00:58:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3498s) excited to see uh we actually have a new [00:58:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3500s) launch coming out for Cleric. Uh we [00:58:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3501s) released an early preview, but it's [00:58:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3503s) coming out. So, I think these use cases [00:58:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3504s) will grow as uh people are trying it out [00:58:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3507s) more and more. But the idea is that [00:58:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3509s) there's so much in the world. We're not [00:58:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3511s) limited to just LLMs. Now, these new [00:58:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3513s) paradigms have been introduced and I [00:58:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3515s) think we should all take advantage of [00:58:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3516s) that and try building with it. [00:58:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3519s) And so we went backwards, we went [00:58:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3521s) forwards. But I think the most important [00:58:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3522s) thing here is that we go onwards, right? [00:58:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3525s) We have to be very reactive in this [00:58:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3528s) industry. Love it or hate it. Things [00:58:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3530s) just change on a whim. You know, if we [00:58:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3533s) as Cloudflare um back in the day said [00:58:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3535s) our hypothesis was wrong, uh we have to [00:58:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3538s) go and actually pivot and start serving [00:59:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3541s) these giant LLMs in order to take [00:59:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3543s) advantage of the fact that we want to be [00:59:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3544s) an agent cloud and that it makes sense [00:59:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3546s) for our story. when an east wing wind [00:59:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3549s) comes and we need to um have decision [00:59:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3552s) models because it fit perfectly with our [00:59:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3554s) infrastructure story, we had to pivot [00:59:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3556s) and be a neolab as well. So it's not [00:59:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3558s) it's one thing to sit here and look back [00:59:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3560s) but the more important thing is that we [00:59:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3562s) took what we had in stride and moved [00:59:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3565s) forwards. [00:59:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3566s) So I'll leave you with one last quote [00:59:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3567s) here [00:59:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3570s) which is a story from tool story. Uh [00:59:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3572s) remember then there's only one time that [00:59:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3574s) is important now. It is most important [00:59:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3576s) time because the only time when we have [00:59:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3578s) any power and so I'll leave you with [00:59:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3580s) this quote and a question of what are [00:59:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=3582s) you going to build today?