Transcription: OpenAgents Episode 195 - Designing 10x Better

Source: https://x.com/OpenAgentsInc/status/1988293182942228779 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 195: Designing 10x Better How can our product become '10x be... Upload date: 20251111 Transcription model: gpt-4o-transcribe-diarize Generated at: 2026-06-01T03:31:45Z

Machine-generated transcript. Review speaker labels and wording before using this as quote-grade source material.

[00:00] Christopher David: The end of cloud inference?

[00:03] Christopher David: Maybe.

[00:05] Christopher David: People just remember,

[00:07] Christopher David: just remember that when people around Teapot, AI Twitter,

[00:12] Christopher David: days or weeks from now,

[00:13] Christopher David: talk about how cool the foundation models API is from Apple,

[00:17] Christopher David: that you heard it here first,

[00:19] Christopher David: okay?

[00:19] Christopher David: For some reason, all of you are sleeping on it.

[00:23] Christopher David: Fine,

[00:23] Christopher David: good for me.

[00:25] Christopher David: But Apple has put out amazing new primitives for app devs, rivaling and in some cases exceeding the Vercel AI SDK in my own experience and what I can tell of its potential.

[00:39] Christopher David: Massively, massively impressed by it.

[00:41] Christopher David: And it seems to be working fine.

[00:43] Christopher David: This is the first ever that I've seen agentic search.

[00:48] Christopher David: through a code base,

[00:49] Christopher David: grep, and all of that stuff using tool calls.

[00:53] Christopher David: You can run it on device.

[00:56] Christopher David: I have it hooked up because we're doing a desktop mobile sync. I just run it on the M2 chip and then stream it over WebSockets to my mobile device for controlling QuadCode. But we're going to be supplementing QuadCode and Codec with the on-device.

[01:15] Christopher David: Which got me thinking,

[01:18] Christopher David: hey Chat GPT, what percentage of AI workloads in the next five years can or will run on Apple Silicon give me a range of possibilities and explore the ramifications for OpenAI,

[01:32] Christopher David: NVIDIA, and other major players if certain percentages of AI workloads move from cloud to the edge.

[01:38] Christopher David: So I've had this at Chat GPT Pro.

[01:40] Christopher David: which can search the internet for sources and stuff like that.

[01:43] Christopher David: I also fed in recent tweets and analysis,

[01:46] Christopher David: a little bit of this, a little bit of that,

[01:48] Christopher David: Chris Paik's essay,

[01:50] Christopher David: and I'm going to share this link in the tweet below this.

[01:54] Christopher David: You'll see it.

[01:55] Christopher David: Summaries, blah, blah, blah.

[01:56] Christopher David: I've got a one-liner down here. If the next five years play out as current vectors suggest,

[02:01] Christopher David: Apple Silicon could end up running 15 to 25% of the world's AI inference by 2030 with a credible 7 to 31% total range.

[02:10] Christopher David: It gave like a stretch amount earlier of 35% also.

[02:14] Christopher David: And that swing is large enough to reshape GPU demand,

[02:17] Christopher David: API business models,

[02:18] Christopher David: and who captures the economics of everyday AI.

[02:26] Christopher David: And then I have to do a little bit of math about, like, what is the actual market cap implications of all that.

[02:32] Christopher David: Is it correct to call this a trillion dollar question?

[02:35] Christopher David: Yeah,

[02:35] Christopher David: yeah, easily a trillion dollar swing.

[02:38] Christopher David: Oops.

[02:39] Christopher David: Thumbs up.

[02:42] Christopher David: And what does it say up here for being able to tell if it's on track to being on the high end of that 7 to 30 percent?

[02:54] Christopher David: What to watch for.

[02:55] Christopher David: What would have to be true for the high scenario?

[02:57] Christopher David: Small capable models, it's already got one of these, the one on there is, you know, like equivalent of 4B.

[03:05] Christopher David: I'll skip to for a second. PCC scale out,

[03:07] Christopher David: the sort of like central cloud that things could elevate to if you want a more powerful model. Hardware cadence,

[03:13] Christopher David: okay.

[03:14] Christopher David: Now I don't have much control over 1, 3,

[03:16] Christopher David: and 4 there.

[03:17] Christopher David: Number two, however.

[03:19] Christopher David: Developer adoption of Apple's Foundation Models plus MLX is broad.

[03:25] Christopher David: None of you people have been talking about this at all.

[03:28] Christopher David: You didn't know, I guess.

[03:30] Christopher David: You're waiting for me to make a video about it. Is that right?

[03:33] Christopher David: Okay.

[03:35] Christopher David: Well,

[03:35] Christopher David: we're going to continue doing what we're doing, which is putting out open source releases on our GitHub.

[03:42] Christopher David: We're going to continue building in public,

[03:44] Christopher David: we're going to continue making videos,

[03:46] Christopher David: and I'm going to chronicle for you over the next few videos over hopefully the next week or so what percentage of our own workload as I'm focusing on coding

[04:01] Christopher David: agents,

[04:02] Christopher David: specifically letting you manage the top coding agents,

[04:07] Christopher David: currently cloud code and codecs from your phone.

[04:11] Christopher David: It's like what percentage of our own coding agent workload is in the cloud?

[04:16] Christopher David: Well, right now it's 100%, but with version 0.3 of our Tricoder app in a few days,

[04:22] Christopher David: that's going to drop from 100% to,

[04:24] Christopher David: I don't know,

[04:24] Christopher David: 95%. And in the beginning we're just going to be using the local model for a little bit of orchestration, we're generating title summaries, but like 100 is going to drop to 95 and then it might keep dropping. What other things can we strip out?

[04:41] Christopher David: and go local.

[04:43] Christopher David: Can we bring that down to 50%?

[04:47] Christopher David: 20%?

[04:49] Christopher David: At what point does that start having broader ramifications?

[04:54] Christopher David: So stay tuned.

[04:56] Christopher David: If you're a developer who wants to leverage the best coding agents from your phone,

[05:02] Christopher David: go sign up for our test flight.

[05:03] Christopher David: This is going to be iOS only for probably the foreseeable future,

[05:06] Christopher David: next month or so at least.

[05:08] Christopher David: Go sign up. If you're an angel investor or an early stage VC fund and you want exposure to this thesis,

[05:14] Christopher David: my DMs are open.

[05:17] Christopher David: Message me sooner than later,

[05:18] Christopher David: I recommend.

[05:19] Christopher David: See ya!

[05:20] Christopher David: They say new products should be a 10x improvement over what came before.

[05:24] Christopher David: So what does a 10x improvement over Claude code over Codex look like over Cursor look like?

[05:30] Christopher David: Here's 10 areas that we think make sense.

[05:33] Christopher David: Let's go through each of them.

[05:35] Christopher David: Ditch the TUI.

[05:36] Christopher David: Okay, these janky fake terminals that everyone is just copying off of Claude Code because that's what they did.

[05:42] Christopher David: Where you're like pretending you're a developer and cool in the terminal.

[05:45] Christopher David: Where there's no reason to be in the terminal because you're not using like actual terminal commands.

[05:50] Christopher David: There's nothing about the TUI that you couldn't also do from a desktop app.

[05:53] Christopher David: Give me a sidebar. Get rid of this screen flicker. I can't copy paste properly.

[05:58] Christopher David: There's all this stuff that's already been solved.

[05:59] Christopher David: even solved by like actual good ui give me a chat gpt style desktop app with like history on the side some widgets down here if i want to manage long-running agents like why does that not exist keep it simple number two go mobile okay i want to have a desktop app that syncs to my mobile phone so

[06:26] Christopher David: Give me the exact same flow,

[06:27] Christopher David: the same UI components,

[06:29] Christopher David: the same controls.

[06:30] Christopher David: If I'm at home doing my usual workflow,

[06:33] Christopher David: let me have the desktop app.

[06:36] Christopher David: But if I want to go to the store,

[06:37] Christopher David: let me be able to do all of the same stuff from my mobile phone.

[06:40] Christopher David: Okay, we figured out how to do that. See the previous episodes.

[06:45] Christopher David: Some people say like, oh use, you know, what about OpenAI's Codex? You can use it on the web. You can.

[06:52] Christopher David: Have you tried it?

[06:53] Christopher David: Codex. Watch this.

[06:54] Christopher David: Oh,

[06:55] Christopher David: get started.

[06:56] Christopher David: Codex web.

[06:57] Christopher David: takes you to the mobile app you can't use it from the mobile app there's no codex option here so are these 500 billion dollar companies or what no they need a startup i guess to make an actual freaking app let's do that code overnight um i want to be able to let stuff run overnight and not in this hilarious horrible way that you have to do with codex oh yeah it fixes thousands overnight how do you actually do that right now you have to just queue up continue continue continue continue like and it's it's so dumb

[07:27] Christopher David: So yeah,

[07:28] Christopher David: but hey,

[07:29] Christopher David: this is an idea for a new feature,

[07:30] Christopher David: scheduled prompts.

[07:31] Christopher David: In three hours do this, in four hours do this, in six hours after a limit resets do this.

[07:36] Christopher David: Or make it even more intelligent, like give the agents a little bit of discretion about like checking to see what's been done or what's not.

[07:42] Christopher David: to advance your kind of higher level objective of adding more tests or whatever. I did like a basic version of this last night. You can see here I like set up an orchestration to run overnight.

[07:51] Christopher David: It made like a few, it's basically like a glorified cron with different kind of like sub agents. It's kind of orchestrating different codex things.

[07:58] Christopher David: You can read the audit of what went well and not with that, but I'm going to iterate on this and in a very soon version we'll have this out, okay?

[08:06] Christopher David: CLI agents is sub agents.

[08:08] Christopher David: You know how Claude will like delegate to a sub agent and it'll kind of like have its own context window when it like issues a task.

[08:15] Christopher David: That's what it's doing.

[08:16] Christopher David: Why can't I just have my main open agents chat delegate to codex, delegate to Claude code like in the context of a single conversation with all of that context available to agents who want it. So here is like, you know, a tool call that my Apple Foundation.

[08:32] Christopher David: model chat model,

[08:35] Christopher David: which is basically a little glorified router that we built with a few basic tool calls where it can delegate to codex.

[08:41] Christopher David: It's going to delegate to cloud code. It can have them like talk with each other.

[08:44] Christopher David: Whole lot of design space to explore there.

[08:47] Christopher David: history and memory like yeah just give me the sidebar with the chat so i can actually see like what's already happened and like let me easily reference past chats like in this other chat or let it retrieve it itself there should be a local sqlite database where it's able to be like oh yeah like you know three days ago we discussed this i saw that we came to this conclusion right now all of the like clawed code and codex conversations are just in these like giant json blobs not optimized for discovery at all

[09:10] Christopher David: hassle-free interrupt all the debate and like hoops you have to jump through to get mcp working people oh yeah no you have the so the tool registry definitions are too long so now you need to like create an api from it but then you need to like create a script that'll consume that or rewrite like stop it stop stop making me think in those terms like if you're in a app you should be able to like easily pull in integrations

[09:36] Christopher David: i'll figure all this shit out but like make it available to other people that they don't have to like go through all these insane hoops um embrace open source the the leader here in my view on the app side for coding agents is open code they've built a massive community of contributors they've got like very healthy community people actually building like props to codex for being open source i think that's cool but like embracing it and like having a whole bunch of people contribute and add to it is great um i just want to do that kind of thing but for our more opinionated

[10:05] Christopher David: Agentic flow here with open agents.

[10:08] Christopher David: Okay local swarm and cloud inference if I've got a bunch of compute

[10:13] Christopher David: On my computer,

[10:13] Christopher David: if I've got stuff that my foundation models like M2 chip is able to do for summarization, for updating title summaries, like use your local compute where it's needed.

[10:24] Christopher David: Let's say that there's a whole bunch of other people that have free compute.

[10:28] Christopher David: Hey,

[10:29] Christopher David: maybe I'm able to like pay them and like use that swarm compute.

[10:32] Christopher David: More on that in a second. And then cloud like,

[10:34] Christopher David: yeah, obviously the we're steering Codex and Cloud Code,

[10:38] Christopher David: which are cloud based agents.

[10:39] Christopher David: agents but like maybe i do want to like go hit grok groq for they've got let's say they put kimmy's new thinking model up with that and i want to like use that like there should be a mix of inference possibilities people should be able to like mix and match based on their their cost preferences or have that done automatically after setting some preferences um compute fracking okay

[11:02] Christopher David: you you have compute available to you that you are not doing anything with why are you not doing anything with it okay you've got an m2 chip that's just sitting there are you using that compute right now to run agents no then you're wasting electricity well there just hasn't been any software for you to run that on your computer was just able to like have agents run around the clock or if you don't have any economically sensible use cases

[11:28] Christopher David: Why can't you sell that to your neighbor who will pay you a tiny fraction of a dollar,

[11:33] Christopher David: tiny fraction of a Bitcoin,

[11:35] Christopher David: to use your compute?

[11:36] Christopher David: See episode,

[11:37] Christopher David: I don't know, 140, we did some compute marketplace stuff. We're going to be folding that back into this.

[11:43] Christopher David: Sam Altman and all these people talking about,

[11:44] Christopher David: oh, we're bottlenecked by compute.

[11:46] Christopher David: Okay, you've got every freaking device,

[11:50] Christopher David: you know, the 2 billion Apple devices,

[11:52] Christopher David: what percentage of those are iOS 16 and can now run or 26 and can now run the...

[11:59] Christopher David: On device foundation models.

[12:01] Christopher David: Where is it? See our recent tweets about, yeah, yeah.

[12:04] Christopher David: So what percentage of these global AI workloads are going to run on Apple Silicon?

[12:08] Christopher David: The big labs are not going to innovate in this direction because they are now like in the NVIDIA drinking the NVIDIA Kool-Aid, but we have no such limitations.

[12:19] Christopher David: So just remember,

[12:20] Christopher David: compute fracking,

[12:21] Christopher David: you heard it here first.

[12:23] Christopher David: Revenue sharing.

[12:24] Christopher David: Now, to anyone who's like,

[12:25] Christopher David: Chris,

[12:25] Christopher David: why are you spilling all the juice here,

[12:27] Christopher David: the sauce? You're giving it all away,

[12:29] Christopher David: the alpha.

[12:29] Christopher David: You're just leaking it out here.

[12:31] Christopher David: No, no, no.

[12:31] Christopher David: We're here to build network effects.

[12:33] Christopher David: So we're going to be open source.

[12:34] Christopher David: We're going to be building in public.

[12:36] Christopher David: We're going to monetize this.

[12:38] Christopher David: And how we're going to be building the biggest network is because we're going to be getting developers paid.

[12:45] Christopher David: By building an actual marketplace hey developer and this is going to solve so many of the coordination problems around like why MCPs suck in implementation because no one has an actual financial incentive to maintain them or make them good hey developers you wrote an agent plug-in would you like to put that in a registry and get paid a tiny fraction of a cent every time someone uses it

[13:06] Christopher David: How do you do that?

[13:07] Christopher David: Built-in Bitcoin wallet.

[13:08] Christopher David: Oh, have you built this already?

[13:10] Christopher David: We built this already.

[13:11] Christopher David: Like, that's the benefit of having been around for two years and on episode 190.

[13:14] Christopher David: We've built all of this stuff in prototype form,

[13:16] Christopher David: and now we're just mushing it into, you know, a single product now that we've got the sort of agentic coding go-to-market piece figured out.

[13:23] Christopher David: Okay, what would you add to this list?

[13:26] Christopher David: Please tell us.

[13:27] Christopher David: We are aiming to get at least basic versions of every one of these 10 upgrades.

[13:32] Christopher David: in place by the end of this year,

[13:35] Christopher David: six weeks from now,

[13:35] Christopher David: what would you add to it?

[13:39] Christopher David: See you soon.