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

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.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

From chatbots to a phone that acts for you: “your own Jarvis”
The host and Div Garg contrast today’s chatbot-centric AI with a future where you talk to your device and it completes tasks for you. Div frames the vision as a personal AI you own, supported by a broader ecosystem of devices like wearables.
- •Steve says most people still experience AI as chatbots, but the future is talking to your device to get things done
- •Div Garg describes a Jarvis-like personal AI that you own and carry everywhere
- •Div says they want “an army of agents” that can do tasks on your behalf
- •The vision includes new devices such as smart wearables alongside the phone experience
Founding bet and ambition: dropping out, saying no to OpenAI, building AGI Inc.
Div Garg introduces himself and the company’s core product: on-device models that can take actions across apps on your phone. He also shares major personal stakes—leaving a Stanford PhD and turning down an offer from OpenAI—to pursue this mission.
- •Div Garg says he dropped out of his Stanford PhD to build AGI Inc.
- •Div says he turned down a billion-dollar offer from OpenAI
- •Div says AGI Inc. built on-device models that can take actions on your phone and control apps
- •Div claims they are number one on benchmarks like AndroidWorld
Day one inside the company: onboarding with agents (and culture signals)
A new team member, Sharik, narrates his first day and onboarding experience. The episode shows how the team uses agents internally, including a playful test where an agent is instructed to open a coffee link and brew coffee.
- •Sharik introduces himself and says it’s his day one onboarding at AGI Inc.
- •Sharik says he previously worked on voice AI and voice LLMs
- •The team jokes that interns don’t pick up coffee—agents do, demonstrated with an instruction to brew coffee
- •Sharik describes talking with Div at an AGI Inc. party/co-working session and liking the founders’ outlook on competition
Origins story: creating Stanford’s transformers course and cold-emailing Geoffrey Hinton
Div recounts his academic background and how he helped start a transformers course at Stanford after noticing a gap. He describes writing a proposal, emailing Geoffrey Hinton, and quickly assembling high-profile guest lecturers for the course series.
- •Div says he studied at Cornell (double major in physics and CS) before joining Stanford for graduate studies
- •Div says Stanford didn’t have a transformers course, so he and two friends created one via a proposal to the dean
- •Div says he emailed Geoffrey Hinton using the public email on Hinton’s website and got a reply within 10 minutes
- •Div says they brought in notable AI speakers, including Andrej Karpathy and people from Anthropic and OpenAI
Why bet everything: productivity, “an army of agents,” and equalizing access
Div explains why he sees a personal AI butler as a high-impact direction. He argues that wealthy people benefit from teams who do tasks for them, and that agents could narrow this gap by giving everyone similar leverage.
- •Div says this is the most impactful thing he can do for humanity right now
- •Div uses the Jarvis analogy to describe an AI butler that knows you and can do things for you
- •Div argues a key difference between a typical person and a billionaire is having an “army of people” to do tasks
- •Div says an “army of agents” could remove gaps and “equalize” people on the same basis
Steve’s focus: what makes agents hard—memory, context, intent, and proactivity
Steve describes how building action-taking agents goes beyond executing commands. He emphasizes that key bottlenecks are memory and context, understanding ambiguous requests, and ultimately becoming proactive in ways aligned with user intent.
- •Steve says he initially focused on action execution, but the main limiters became memory and context
- •Steve highlights the challenge of understanding ambiguous instructions and recognizing underlying intent
- •Steve frames the end goal as proactivity: the agent knows what you’d want and does it
- •Steve describes a feedback loop where more trust leads to sharing more context, enabling more capability
Trust before capability: the on-device goal and “no cloud to trust”
Steve connects trust directly to capability, arguing that users will only provide enough context if they trust the system. He says running everything on-device is the ultimate target because data never leaves the device, removing the need to trust a cloud service.
- •Steve says trust and capability are linked: more trust enables more context and understanding
- •Steve says more context makes the agent more capable and ultimately proactive
- •Steve says the team is working toward running everything on-device
- •Steve says if data never leaves the device, “there’s no cloud to trust”
How the phone agent works: screenshot → action loop (click/type/scroll)
Naman explains the technical stack in an accessible way. He describes how tasks are paired with screenshots, then an action model predicts the next interaction, looping through actions until the end state is reached.
- •Naman says the system takes in a user task plus a screenshot
- •Naman says the agent looks at the screenshot to decide what action to perform next
- •Naman describes actions as click, type, and scroll, executed in a loop until completion
- •Naman outlines the stack as application layer, agent harness, and inference engine
The hardest bottleneck: making it fully on-device without losing speed or accuracy
Naman identifies on-device inference as the hardest part because reducing model size can reduce capability. He says the challenge is keeping accuracy comparable to larger setups while staying fast, and that this is what unlocks privacy and security.
- •Naman says the biggest bottleneck is getting the agent to run fully on-device rather than in the cloud
- •Naman says shrinking models tends to decrease capability, creating a research challenge
- •Naman says they need the model to remain accurate while also being fast enough
- •Naman says fully on-device operation “truly unlocks privacy and security” for users
Why frontier labs may not ship it: ownership vs renting a cloud model
Div addresses what happens if major labs release similar phone-control capabilities. He argues AGI Inc. is differentiated by on-device execution and a business vision where users own the model and data, unlike cloud-rental models that depend on ongoing usage to monetize.
- •Div says their work will run on the device and differs from frontier labs’ business models
- •Div frames the alternative as “renting a model running on someone else’s cloud” versus owning it on your device
- •Div says users would own the model and the data in their approach
- •Div argues frontier labs don’t want to ship fully on-device models because it complicates how they make money
Accessibility and consumer use cases + live demo: ordering Chipotle hands-free with guardrails
A team member describes focusing on consumer and accessibility scenarios, like helping a grandmother navigate a phone. The episode then shows a hands-free ordering demo through DoorDash, including a purchase confirmation step as a safety guardrail.
- •A team member describes accessibility use cases, like ordering food or starting a FaceTime call for someone who struggles with phones
- •The interface shows presets across different channels and supports a custom prompt
- •In the demo, the agent opens DoorDash hands-free and clicks through the order step-by-step
- •The system prompts for confirmation before placing a paid order as a guardrail
From viral demo to real business: customers, pain points, and early scrappy validation
Naman shifts from demo excitement to the practical need for customer value. He emphasizes finding users who love the product and spending time identifying real friction in daily life; another speaker shares scrappy tactics like going to a music festival to gather feedback and iterating fast under time pressure.
- •Naman asks how to turn a demo into a business and answers: find customers and users who love it
- •Naman says the key is understanding why someone would buy/use it and what value it creates
- •A speaker describes going to Golden Gate Park during a music festival and giving out drinks to engage potential customers and get feedback
- •A speaker describes intense crunch and rapid iteration to build an Android agent
Partners, edge-device positioning, and the “demo vs production” gap (error compounding)
Steve describes how strong demos helped attract interest from device and chip companies, positioning the company in an under-served edge-device lane. Naman explains why production reliability is much harder than a prototype, comparing it to self-driving and highlighting how small per-step error rates compound across multi-step tasks.
- •Steve says partnerships came from seeing tech demos that made people say they’d never seen a phone do that before
- •Steve says device-makers especially wanted edge-device automation and few others were pursuing it
- •Naman says a prototype can work inconsistently, but shipping requires rigorous testing and handling edge cases
- •Naman explains error compounding across many steps and argues high single-turn accuracy still fails over long task sequences
How work happens: production code on day one, trial hiring, late nights, and in-person events
The episode shows internal engineering coordination, including a new hire working directly with production systems and planning changes. Naman describes using trial periods to assess independent thinking and iteration, explains why many work late for focus, and the team discusses why in-person events matter when online content feels less trustworthy.
- •A speaker notes it’s uncommon to work on production-grade code on the very first day, but that’s what’s happening here
- •Naman describes trial-to-hire: giving ambiguous feature outcomes and evaluating solution quality and product UX thinking
- •Naman says late-night work helps focus due to fewer distractions and autonomy as long as performance meets a bar
- •A speaker explains events build connection and serendipity in an era where digital outreach and information can feel discounted or unreliable
Hardware experiments, culture philosophy, and the long-term vision: billions of personal models and less screen time
A team member highlights creative freedom through hardware experiments like a mini phone that you speak into, designed to spark curiosity and enable voice-driven actions. Naman describes hiring slowly but firing fast to protect culture fit, and outlines a future of billions of personalized, self-learning models that users can own; Steve closes with a vision of an agent-driven operating system that reduces screen time and supports real life.
- •A team member describes a mini phone experiment: a small device you speak into that takes actions for you
- •Naman says culture is shaped by each person; they hire slow but fire fast due to culture-fit importance
- •Naman envisions billions of personalized models that self-learn preferences and are owned/customizable by individuals
- •Steve says the goal is a system/operating system that lets people look at screens less and live more while the AI acts in the background

Publisher timestamps

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.