David Pantera

Personal update: I've joined @a16z as a Scout! Excited to be stepping into this new role as I depart @Google for @Stanford this fall. If you're an early-stage startup (pre-seed/seed), I can write you an angel check! Let's chat if you're building a solution to an important problem. My experience is in consumer AI, but I'm happy to talk to anyone building anything cool! (b2b, AI infra, ML tooling, etc). Brownie points if we’ve worked together or have mutuals!

@venturetwins Ready for Demo Day next week! Excited to see what theyre up to

I've been in venture for almost a decade, and I've never seen anything like the current YC batch. Truly blown away by what small teams are able to ship with AI. What a remarkable time to start a company!

@XRarchitect This turned out awesome

Testing off-axis projection demo. Basic scene layout in Blender glb -> detailed splat via World Labs Three.js for engine MediaPipe for face tracking Next step is to add a character you can move around in the environment

@theworldlabs @XRarchitect What a great time! Already looking forward to the next one 🚀🚀

3D is becoming an interface for creating and interacting with worlds 🌎 Earlier this month, we brought builders together in SF to explore that idea at our first World Labs Hackathon. Here’s a glimpse of what they built. Interested in joining the next one? Let us know!

I think we’re too early to be overindexing on the shape of the robot and instead should be focused on the spatial intelligence driving it. the companies that capture the most value in the robotics race will be the teams working with AI research labs to build the foundational world models that allow any physical form factor to understand, predict, and manipulate three dimensional space.  no matter which form factor they build, optimization at the model layer will lead to the best results, fastest

The biggest constraint in embodied AI / robotics right now is the cognitive architecture. Whether a robotic system has two legs, four wheels, or a single mechanical arm, it requires a foundational world model to fundamentally understand spatial geometry, gravity, and force dynamics.  The physical robot is ultimately just a hardware wrapper for the software's ability to simulate and predict the physical world around it

If the goal is to move heavy pallets across a warehouse, a low profile robot on wheels is infinitely more stable, power efficient, and cost effective than a humanoid robot constantly burning compute just to keep itself balanced on two legs. Form should follow function, and for most industrial tasks, the human body is a terribly inefficient design.

If you look at the robotics companies making headlines, the industry seems convinced that the future of embodied AI looks exactly like us: humanoids. The thesis driving the humanoid robotics boom is that our entire physical environment was built for the human form factor, so building a bipedal, general purpose robot is the fastest way to drop AI into existing workflows without needing to reengineer our factories, homes, cities, etc. It’s a compelling narrative but ignores the reality of unit economics. Right now, non humanoid robots are the ones actually generating commercial revenue and solving tangible problems at essentially every application of robotics. If the goal is to move heavy pallets across a warehouse, a low profile robot on wheels is infinitely more stable, power efficient, and cost effective than a humanoid machine constantly burning compute just to keep itself balanced on two legs. Form should follow function, and for most industrial tasks, the human body is a terribly inefficient design. However, the actual constraint in embodied AI is the cognitive architecture. Whether a robotic system has two legs, four wheels, or a single mechanical arm, it requires a foundational world model to fundamentally understand spatial geometry, gravity, and force dynamics.  The physical robot is ultimately just a hardware wrapper for the software's ability to simulate and predict the physical world around it. I think we’re too early to be overindexing on the shape of the robot and instead should be focused on the spatial intelligence driving it. My prediction for the next five years is that the companies that capture the most value in the embodied AI race will be the teams working with AI research labs to build the foundational world models that allow any physical form factor to understand, predict, and manipulate three dimensional space.  No matter which form factor they build, optimization at the model layer will lead to the best results, fastest.

@venturetwins @krea_ai Can you choose which model it uses for generative fill like this?

This is my new favorite way to edit images ✨ If you haven't checked out @krea_ai's mobile app lately, it's worth a look. I find it particularly convenient to draw on an image, prompt my edits, and apply Nano Banana!

Even a static 3DGS contains a learned representation of 3D space. It understands occlusion and view-dependent lighting (eg how reflections change as you move). This is a form of spatial logic that 2D-only models lack. If a 3DGS accurately represents the physical bounds and surfaces of a room, is it not a functional model of that world?

Many 3D generators output Gaussian Splats (3DGS) for fast rendering, flexible deployment, and high visual fidelity. Static 3DGS aren't world models (no dynamics/semantics) but a true world model must allow distilling 3D-consistent representations for any given time step (3DGS/meshes). This post-distillation serves a dual purpose: 1) validates physical consistency of the model. 2) extracting explicit representations avoids continuously running a heavy generator, thus saves compute and facilitates real-time interaction.

Totally agree. storytelling is the most underrated competitive advantage. too many founders treat their pitch like a data dump when it should be a narrative arc. if you can’t describe the problem, the insight, and the vision in plain English (& in a way that gets ur users excited), the product is probably a feature in search of a mission x.com/davidpantera_/…

founders at the earliest stages consistently underestimate this: describing their business in just 3 super clear, plain English sentences it’s harder than it looks, but soo important! potential customers who never thought about your solution get it instantly. Operators you want to hire see the mission. VCs understand without needing translation you’re the expert BUT distilling your deep knowledge is the #1 founder skill this remains one of the most common first pitfalls I see in pitches same principle applies to your 2-minute version and the long-form pitch

@AnishKattukaran These are great! Home deals with some of our most sensitive data. it's tech we've welcomed into our homes & private lives. would be great to see some privacy initiatives on the roadmap...eg gemini nano (on device models), Private AI Compute, etc

Back again with another round of Google Home updates, based directly on the feedback we’ve been getting from you. Today, we’re starting to ship a mix of country and language expansions, popular feature requests, and bug fixes. Here’s what’s new. 🧵👇

bringing codex functionality to both consumer use cases (whatever that looks like) AND enterprise is super challenging. completely different pain points ur addressing enterprise users want reliability, data governance, specialized coding environments everyday consumer users want low-friction, speed, conversational ease. theyll eventually probably have to split the product entirely to satisfy the differing demands

if you’re designing the product roadmap for openai, it’s likely to be centered around bringing codex driven functionality to everyday users in chatgpt (consumer use cases) & obviously enterprise. thus focused development of codex takes hold as the primary mechanic for company progress. everything branches off of codex just like for anthropic now every functionality branches off of claude code.

"Your customer knows even less about these models than you do." Another example of how storytelling is a skill that separates a great founder from a good one. Building a great product is only half the battle. Your actual job is drawing a bright, undeniable line between your AI tool and the actual value it creates for your users. You have to actively communicate that value and help your customer visualize the most ambitious version of their own future. They have to see the vision before they will ever care about the product.

@bilawalsidhu My thought exactly x.com/davidpantera_/…

it's all over when google realizes the treasure trove that is street view + aerial data and launches a version of genie grounded in the real world...

Congrats, @ruilong_li! 15.2 fps for a closed-loop sim is a big milestone for making generative world models actually playable in real time. does the model struggle with identity persistence? eg for dynamic objects if the retrieved reference images has different traffic or lighting than the current frame

Special moment to see something I’ve worked on so closely come to life! Today we announce Alpadreams — a world model that lets you explore ♾endlessly♾️in ⚡real time⚡. Video: me (left) and Alpamayo policy (right) driving in Alpadreams at #GTC26. research.nvidia.com/labs/sil/proje…

@thefox AI mode with personal intelligence is probably my favorite AI product x.com/davidpantera_/…

We're shipping Personal Intelligence in AI Mode in Search to everyone in the US! 🚀 What's blown me away as a user: even for already specific questions (like my recent one here ⬇️), Personal Intelligence brings helpful nuance to the response -- it kinda feels like subtle magic! 🪄 Here’s a direct link to opt in now: myactivity.google.com/search-service…

This is awesome @jyseo_cv ! Google Maps has spent two decades indexing the physical world via street view...very cool to think about how we could use that data to create simulations of the real world like this. Using training pairs from different timestamps was a pretty smart way to mitigate hallucinations

What if a world model could render not an imagined place, but the actual city? We introduce Seoul World Model, the first world simulation model grounded in a real-world metropolis. TL;DR: We made a world model RAG over millions of street-views. proj: seoul-world-model.github.io

@jongranskog We’ll hit a ceiling of what rasterization can do for realism. World models feel like the first time we’re actually simulating reality instead of just drawing + upscaling triangles.

Naturally there will be a spectrum of models depending on how much control you want. From world models without any geometric 3d representation to a refinement pass over your game renders.

@XRarchitect I suspect step 1 is the vast majority of the time ur spending on a project like this? Automated generation of step 1 would be a huge unlock in this workflow. Even if you could generate a first draft of the 3D env in blender that you could edit/perfect.

Pipeline I've been experimenting with for building out structured worlds: → Blender to block out the environment with simple shapes → Chisel (World Labs) to generate detailed worlds while preserving the layout → Cursor + Three.js to add a character controller and physics

@XRarchitect @theworldlabs Incredible, thank you! 🙏

@davidpantera_ @theworldlabs I got you 🫡 youtube.com/watch?v=fvZBjq… the mini map on top right is a bit more external extra work. but if you are curious about that as well, can help explain!

Exploring cozy worlds 🍀 24 million splats Take a walk through it right in your browser ↓