
Building multimodal agents for businesses & public sectors @GoogleDeepMind. Views are my own.
New York, NY
Joined November 2021
- Tweets46
- Following322
- Followers94
- Likes83
One-step world models are fast, but keeping them accurate is hard. Excited to share our approach - Beckmann World Models, which learn to predict the next frames with a Beckmann transport objective 👇
1/8 Introducing Beckmann World Models (BWM): adapting Beckmann transport into fast, one-step action-conditioned world models.
Predicting the next video state from recent observations, proposed actions in one denoising step function call.
Our final RGB model predicts four PushT frames in 24.7 ms on one A800 GPU.
2
7
1,086
Wrapped up #GoogleCloudNext ‘26 by demoing Sky — a real-time digital avatar concierge built with the Gemini Live API, in partnership with @Citi Wealth, @GoogleDeepMind, and @googlecloud.
It’s exciting to see AI products move beyond pure utility. By pairing fluid dialogue with a nuanced avatar, we're beginning to support interactions that feel human.
Proud of our progress! Sharing some demo clips and photos from the past week.
10
479
JP Suh retweeted
Excited to share Kiki, an open source chrome extension that lets you control your browser with your voice and via chat!
There is a ton of exciting work in the headless browser automation space and in new ai-native browsers. @junepyosuh and I wanted to try a different angle to bring browser automation directly into the browser you already use.
We think that there are a ton of cool use cases, and we would love to hear your thoughts and to collaborate directly on the project here: github.com/kiki-group/kiki-o…. Really looking forward to seeing how we can evolve Kiki together!
JP Suh retweeted
Replying to @hanlee4g
@hanlee4g and I worked on Kiki, an open-source Chrome extension that lets you control your browser with voice and text.
It’s a lightweight assistant that makes the web AI-guided. You can use it to pull a specific detail from a page you've never seen before, or hand off a boring multi-step task like filling out a form, searching Amazon, adding items to cart, and placing an order.
With a lot of progress happening through browser agents and layers built for agentic workflows, we wanted to try a simpler angle. This is for people who just want to keep the browser they already have, install one extension, and talk to it.
It’s an early attempt, but we think this direction could evolve into meaningful use cases together. Welcome any thoughts/comments! github.com/kiki-group/kiki-o…
Spent the weekend judging the @CerebralValley @GoogleDeepMind Gemini 3 Hackathon. Lots of cool projects solving unique problems in various modalities. Check out the gallery - cerebralvalley.ai/e/gemini-3…
Huge thanks to Cerebral Valley, judges, and all the teams for the energy!
3
239
Gemini 3.1 just landed and it's cool to see some mindblowing vibe coded examples already. Perfect timing for the Gemini 3 hackathon with @Cerebral_Valley in NYC next weekend 😏
I'll be there - come say hi if you're building! (link in reply)
We used Gemini 3.1 Pro to build a realistic city planner app. 🏙️
Watch how the model tackles complex terrain, maps out infrastructure, and simulates traffic to generate a high-quality visualization.
1
5
410
JP Suh retweeted
The UK is an amazing place for science & innovation. Thrilled to deepen our partnership with the UK Government to turbocharge scientific discovery with AI - giving scientists here priority access to our frontier models like AlphaEvolve, AI Co-Scientist, AlphaGenome, WeatherNext & more. We’re also building our first automated lab here in the UK for materials science!
JP Suh retweeted
We just dropped Nano Banana Pro, built on Gemini 3. 🍌
With state-of-the-art text rendering, vast world knowledge and studio-quality creative controls, Gemini 3 Pro Image can create and edit more complex visuals, infographics and more. Here’s what’s under the hood. 🧵
JP Suh retweeted
This is Gemini 3: our most intelligent model that helps you learn, build and plan anything.
It comes with state-of-the-art reasoning capabilities, world-leading multimodal understanding, and enables new agentic coding experiences. 🧵
JP Suh retweeted
Today we introduced Gemini Enterprise, built with our most advanced Gemini models. It allows you to chat with your company’s documents, data and apps as well as build and deploy AI agents, all grounded in your information and context.
Have a look at how it helps you build an agent for almost anything:
JP Suh retweeted
Gemini 2.5 Flash and 2.5 Flash-Lite got a significant update🔥
- Better instruction following
- Better agentic tool use
- More efficient, with 50% reduction in output tokens (hence costs) for Gemini 2.5 Flash-Lite and 24% reduction for Gemini 2.5 Flash
JP Suh retweeted
Gemini Robotics 1.5 from @GoogleDeepMind marks the official introduction of agentic capabilities to robots, allowing them to complete complex, multi-step tasks.
But... what does that mean? 🧐
Before, robots were able to accomplish a single task, like picking up fruit or zipping a bag. Now, we're moving beyond that with models that can truly reason, plan, and generalize. For example, now you could ask a robot, “Can you sort these objects into the correct compost, recycling and trash bins?" and the robot would be able to search the local recycling guidelines, understand the objects in front of it, and figure out how to sort them based on the local guidelines.
Gemini Robotics-ER 1.5 is the first thinking model optimized for embodied reasoning, achieving state-of-the-art performance on both academic and internal benchmarks.
Starting today, it's available in Preview in @GoogleAIStudio.















