QUESTION · SATURDAY Q&A

What are you building with large datasets and agentic AI?

ASKED BY · Unknown · 19 SEP 2026 · ▶ 00:40:25

THE PANEL'S ANSWER
Guy CochranPANELIST

Feeding 35 million words of Office Hours transcripts to Claude; Claude suggested local models or batch API ($80–300) for cost efficiency; agentic features like 216 simultaneous agents demonstrate compute power at scale.

Summarised by AI from what the panel said. Watch the moment for the full answer.

IN THEIR OWN WORDS

From the transcript.

Machine-transcribed, so it may contain errors. Every timestamp opens the episode at that second.

40:24all the bots have their own computer, so you could pull up the screen of the bot. It's pretty wild. So you could see what it's doing. And to watch these things just take off and go, especially if you send five things. And yesterday when I asked Claude to do this thing, it sent 216 agents off at once. So, I mean, that's the level of stuff that you can see when you have a lot of data in a big project.

40:49But yeah, I'm anxious to see what happens with this new release of Apple and all this on-device stuff because it's going to get exciting when every home has its own Mac Mini or whatever, and it has its own model on it. That's going to be cool because then ideally that cost is zero. You know, like we just download it and we just ask it stuff and it's all on-prem.

41:09So that would be cool. Something to look forward to. So I've only been using ChatGPT. I'm only paying 20 bucks a month. I feel weird saying that now, seeing you guys spend like $100, $200 a month. I just can't imagine spending that much on an AI. I'm sorry. Well, one thing, try doing this, Alexander. or try downloading the ChatGPT app on your Mac

41:31and then go into where it says Codex and then actually have it build something or do something agentic where it takes over your computer and you'll watch your tokens just burn. I haven't done that yet. Well, a lot of people don't realize that you could do it, but when you ask it to actually go through something and do something pretty hardcore, it's crazy.

READ THE FULL TRANSCRIPT →
MENTIONED IN THE ANSWER
Claude→agentic→
ALSO IN THIS EPISODE

26 more questions
from CoreVideo Pro Zoom App Uses GPU-Native Encoding.

The whole episode →
▶ 00:00:02I hear you've been working on some production-related tools. Do you have anything you can share with the class this morning?▶ 00:04:14If someone is an OBS user, how do you get Zoom sources into this app? Where does OBS come into play?▶ 00:05:00Is that performance meter of some kind showing in the toolbar?▶ 00:05:23Do you have individual telemetry for participants joining, like frame rate and resolution they're sending?▶ 00:06:09If someone has joined and their audio is fluctuating, can you remotely turn their audio up or down on a per-user basis?▶ 00:09:08I see you have a Mac version of this app now. What can you speak to about the differences in video performance, engine, and graphic subsystems between Windows PC and Mac? Are you running OpenGL on GPU? Is it GPU-driven on Mac as well?▶ 00:10:45How do you live stream from this? What about recording? Where does it go from there?▶ 00:11:57Is there multiview and aux output capability?▶ 00:12:44Can you pop out program and preview to different screens or move the entire interface?▶ 00:13:28I know the Zoom API has been getting a lot of updates. Have you heard about the new talkback feature that allows communication? Is that a comms solution? Are you working on it with Core Video?▶ 00:14:56What's a low-budget underwater video and audio setup versus a high-budget setup?▶ 00:17:11What are professional hydrophones for underwater recording?▶ 00:20:15What's your AI project of the week? Anything to share?▶ 00:22:32When you develop a big project like Show Hub, do you write a spec doc or product requirements doc outside of AI tools and then upload it? How do you start?▶ 00:26:34What episode-search tools are you working on to help find specific content in the archive?▶ 00:30:03How much does a talker's distance to the wall behind them matter compared to the wall behind the mic? Which wall should you treat first?▶ 00:32:41What's everyone paying for AI plans?▶ 00:35:24What's the relationship between Cursor and X? Is Cursor run by X?▶ 00:41:52How are you integrating AI with DaVinci Resolve for automated clip generation?▶ 00:44:14What's an MCP server and how do you use it in Resolve?▶ 00:50:15How much faster is AI development compared to traditional team-based development?▶ 00:53:21The iPhone 18 Pro has variable aperture—what can't the iPhone still do? What does Apple need to add next?▶ 00:54:17What's the sensor size on iPhone 18 Pro and what's still relying on computational photography?▶ 00:55:51Didn't Samsung make a camera phone like that years ago?▶ 00:57:46Does Apple claim to be the first with these camera features?▶ 00:58:32What's the historical relationship between Newton patents and iPhone patents?