science + applied ai, office of the CTO, google cloud, ex cto stability ai, creator of dreamstudio, working on science!
Joined February 2022
- Tweets596
- Following795
- Followers4.5K
- Likes1.1K
nin retweeted
We're advancing AI benchmarking systems by letting them play games 🕹️
Games like Poker & Werewolf in the @Kaggle Game Arena allow us to test AI capabilities and "soft skills" in controlled sandbox environments before they’re deployed 🧵 (1/4) ↓
The @Kaggle Game Arena has added our results for Werewolf, Poker and Chess.♟️✨
These challenges test the skills AI models need for the real world – like contextual communication, building consensus, and navigating ambiguity.
Find out how Gemini did → goo.gle/4qU4qbS
nin retweeted
Introducing Google Workspace Studio, where anyone can build a custom AI agent in minutes to delegate the daily grind. Automate daily tasks and focus on the work that matters instead. → goo.gle/4p9owy5
nin retweeted
Excited to share our new paper on training agents inside world models with @danijarh !
Excited to introduce Dreamer 4, an agent that learns to solve complex control tasks entirely inside of its scalable world model! 🌎🤖
Dreamer 4 pushes the frontier of world model accuracy, speed, and learning complex tasks from offline datasets.
co-led with @wilson1yan
Big update here to the Gemini Flash models. Flash is chosen for many business applications purely due to latency needs so the increases here raise the baseline across a huge set of tasks. This release in addition to the new Gemini Live model earlier this week, and the Robotics 1.5 preview... epic week of releases by the Gemini Team.
Updates to Gemini 2.5 Flash & Flash-Lite are here, with improvements in tool use, reliability and efficiency.
Preview models get our newest updates into your hands, quickly. Access them via the -𝚕𝚊𝚝𝚎𝚜𝚝 alias to skip updating code for each release ⚡
developers.googleblog.com/en…
Today we are releasing Gemini Robotics-ER 1.5 — the first embodied reasoning model for robotics and stream reasoning that’s now generally available. Very excited about this leap forward in Google Gemini and video understanding for Robotics.
Since I started working a lot with Gemini Live earlier this year, I've held strong conviction that video reasoning is going to unlock not just robotics applications but many other things, for instance, working with streams with the hierarchical, general and complex reasoning that proper vision models can do.
Some thoughts:
✨ this is the visual cortex and frontal lobe "reasoning brain” for robots — it can understand natural language, reason through multi-step tasks, and figure out what to do next.
✨ It can do spatial + temporal reasoning, so understand over time what is going on in spatial terms
✨ It can use grounding, so Google Search to custom APIs etc
✨ It has thinking budget controls like existing Gemini APIs
✨ Built-in safety upgrades, like avoiding plans that break physical constraints.
As an example, it can act as the route to do things like, “sort this stuff into waste and recycling, according to my local rules.” This is definitely a new frontier and pumped this is now available as another huge leap from the Gemini team to push forward the industry.
You can try it out now in Google AI Studio and through the Gemini API (preview).
🔗 Full announcement here: developers.googleblog.com/en…
🔗 Tech report: storage.googleapis.com/deepm…
Gemini API Grounding with Google Maps is now generally available, so devs can build apps connected to fresh information from Google Maps. This is now globally GA and supports combined Grounding with Google Search.
This is a big deal for ensuring factual and reliable model outputs with spatial real-world information especially for industries like travel, real estate, and social media. Let's build :))
Docs: cloud.google.com/vertex-ai/g…
We (Google) just open-sourced Agents to Payments (AP2) — a standardized protocol enabling LLM-based agents to securely interact with payment processors.
This unlocks agentic commerce → agents moving from info queries to real transactions.
🔗 Blog: lnkd.in/enf9XhqS
🔗 Spec: lnkd.in/eZuNAQwD
nin retweeted
Gemini and Grok ended April with a ~10 million gap in store downloads, increasing to ~136 million by May.
nin retweeted
MINIMAX IS UNHINGED → NAUGHTY PIKACHU
This tool is too much fun.
MIDJOURNEY PROMPT:
flash photo, Batman and real life pikachu having fun at a house party, 1980s --ar 3:2 --stylize 75 --v 6.1
MINMAX PROMPT STRUCTURES:
01. The camera [motion/direction] as Pikachu and Batman [action]
02. The camera [motion/direction] revealing [outcome]
MINIMAX PROMPT EXAMPLES:
01. The camera pans down to reveal empty beer bottles everywhere
02. The camera pulls out revealing a man walking into the scene and offering Pikachu and batman comical mushrooms
03. The camera pans up to reveal Spiderman crawling on the ceiling upside down
04. The camera pans down to reveal the joker crawling on the ground, before he looks at the camera and waves awkwardly
05. The camera pulls out revealing a Wonder Woman walking into the scene, blowing smoke into Pikachus face comedically
06. The camera pulls out revealing Deadpool throwing Pikachu out of the party
#minimax #hailuo_ai #midjourney @Hailuo_AI
50% barrier beaten on ARC prize :)
Wow! 50% intelligence barrier is broken by AI, that’s a major milestone!
For those not familiar with the ARC-AGI (brain child of @fchollet ) here is an intro and why it’s important towards developing AGI (via ChatGPT) :
The ARC Prize is a public competition offering over $1 million to encourage the development of artificial general intelligence (AGI). AGI refers to machines capable of understanding, learning, and applying knowledge across a wide range of tasks, similar to human intelligence.
The competition centers on the ARC-AGI benchmark, a collection of 800 logic puzzles designed to test an AI’s ability to learn and adapt to new problems. These puzzles are straightforward for humans but challenging for current AI systems, making them an effective measure of progress toward AGI.
The ARC Prize is significant because it aims to advance AI research beyond specialized tasks, pushing toward machines that can think and learn more like humans. By offering substantial rewards, the competition motivates researchers to develop innovative solutions and share them openly, promoting transparency and collaboration in the AI community.
The team at @heygibsonai is working on generation of complete back-end systems, through to middleware and APIs. Doing some cool work.
More code is not always better! GibsonAI is purpose built to generate production ready backend code in minutes. The focus is on creating high quality, maintainable and efficient code, avoiding the pitfalls of AI-induced tech debt that can come from current corp of AI coding assistants. Learn more at bit.ly/47NVY5z
nin retweeted
Today, we’re excited to share an early preview of our latest video model @hotshotco. And it’s available for you to use today.
Link to try + more results in the thread below 👇
SD3 paper released today. Well done team!
Congrats to the Baseten team!
We're excited to announce that we've raised a $40M Series B to help power the next generation of AI-native products with performant, reliable and scalable inference infrastructure.
baseten.co/blog/announcing-o…
nin retweeted
thanks a ton @StabilityAI raising the absolute bar
it has rotary embedding kernels
digits in the tokenizer are split into individual ones
bias terms removed
and it retains the neox throughput optimizations
i am going to push this thing to the absolute limits
local gpt4? lets hit phone gpt4 by q3.
Stability AI presents Stable LM 2 1.6B
demo: huggingface.co/spaces/stabil…
model base: huggingface.co/stabilityai/s…
instruction-tuned: huggingface.co/stabilityai/s…
introducing first language model from the new Stable LM 2 series: the 1.6 billion parameter base model and an instruction-tuned version. The base model is trained on approximately 2 trillion tokens for two epochs, incorporating multilingual data in English, Spanish, German, Italian, French, Portuguese, and Dutch. Leveraged recent algorithmic advancements in language modeling to strike a favorable balance between speed and performance, enabling fast experimentation and iteration with moderate resources.
Proud of the team today on releasing the Video API! Well done team @StabilityAI
Today, we are adding Stable Video Diffusion, our foundation model for generative video to the Stability AI Developer Platform API.
The model can generate 2 seconds of video, comprising of 25 generated frames and 24 frames of FILM interpolation, within an average time of 41 seconds.
Developers interested in utilizing Stable Video Diffusion through an API can access it now on the Stability AI Developer Platform.
Learn more here: bit.ly/3Rymnw9
nin retweeted
Introducing Stable Zero123: Quality 3D object generation from single images.
Stable Zero123, a new in-house generative AI model is capable of creating 3D objects from any input image.
Due to improved training datasets and elevation conditioning, the model demonstrates 3D understanding of the object’s appearance from various angles with notably improved quality over Zero1-to-3 or Zero123-XL
Learn more here: bit.ly/48a5crq