Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

EO StudioEO Studio

48 Hours Inside the Startup Building the World’s First On-Device AGI

Wispr Flow's Notetaker captures every call and gives you a clean summary and to-do list the moment it ends.

SteveguestDiv GargguestNamanguest
Nov 9, 202625mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

At a glance

WHAT IT’S REALLY ABOUT

Building an on-device AI agent that controls your phone privately

  1. Div Garg and the AGI Inc. team describe building a truly personal AI that can control a phone and use apps on a user’s behalf, with the long-term aim of running across multiple device types.
  2. Steve explains that beyond basic action-taking, the main constraints become memory, context, and correctly inferring user intent—especially for ambiguous requests and proactive assistance.
  3. Naman outlines a technical loop where the system interprets screenshots and predicts UI actions (click/type/scroll) repeatedly until the task is completed, and says the hardest bottleneck is moving inference fully on-device without losing capability or speed.
  4. The team stresses that real-world deployment is much harder than a demo because edge cases and compounding errors across many steps demand production-grade testing and reliability.
  5. Multiple speakers describe a culture built around autonomy and intensity, plus recruiting via in-person events to build trust and serendipity in an environment they describe as filled with low-quality digital information.

IDEAS WORTH REMEMBERING

5 ideas

AGI Inc. is aiming for a “personal AI” that acts for you across devices, not a chatbot.

Div Garg frames the product as an AI you own that can take actions across any app on your phone, with a longer-term vision spanning other devices like smart glasses and watches.

The hardest problems shift from “taking actions” to understanding intent, context, and memory.

Steve says that as agents get more capable, the limiting factors become memory, context, and intention-understanding—especially when user requests are ambiguous or when the agent should become proactive.

Trust is positioned as a prerequisite to capability—and on-device operation is the trust strategy.

Steve describes a feedback loop where increased trust leads users to share more context, which improves capability; he says the ultimate trust target is running everything on-device so there’s “no cloud to trust.”

Their approach is a screenshot-driven action loop that predicts and executes UI actions iteratively.

Naman explains their agent loop: receive a task and a screenshot, use an action model to choose click/type/scroll, execute on the screen, and repeat until reaching an end state, with the stack split across application layer, agent harness, and inference.

“Demo success” isn’t the goal; compounding errors make production reliability the real bar.

Naman emphasizes that prototypes can work inconsistently, but production requires rigorous testing because errors compound across multi-step tasks; he argues that even high single-step accuracy can collapse over long sequences.

WORDS WORTH SAVING

5 quotes

Most people's experience with AI is still chatbots. The future is that you just talk to your device and it does things for you.

— Steve

If you love what you're doing, it doesn't matter. I dropped out of my Stanford PhD and, uh, turned down a billion-dollar offer from OpenAI to build AGI Inc.

— Div Garg

There's a relationship between trust and capability. If you trust the AI more, if you trust the tool more, then you'll give it more context, more understanding about you.

— Steve

Getting the agent to be fully on device.

— Naman

At the end of the day, you don't, you don't go to an everyday person and say, "Agent. I made a great agent for you." One thing it can mean for people is that you look at your screen less, and you spend less time on the screen and more time in real life.

— Steve

Read the available transcript
Jarvis-like personal AI visionOn-device privacy and trustPhone control via UI action loopMemory, context, and intent understandingDemo vs production reliabilityCompounding error across multi-step tasksHiring culture and in-person events

SPEAKERS

  • Steve

    guest

    Co-founder at AGI Inc., focused on building an on-device AI agent with privacy-first execution and edge partnerships.

  • Div Garg

    guest

    Founder and CEO of AGI Inc., building on-device personal AI agents; background includes Cornell and Stanford research/PhD track.

  • Naman

    guest

    Co-founder at AGI Inc. leading engineering and research on the agent stack and on-device inference.

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.

iOSAndroidClaudeChatGPTChrome