Dwarkesh Podcast: Jeff Dean & Noam Shazeer – 25 years at Google: from PageRank to AGI

Zyphra AI announces a beta release of Zonos, a highly expressive open source TTS model under the Apache lincense

Luma AI - "Image to Video with Ray2 is here. Drop any image into Dream Machine to make it a video and bring it to life"

Sam Altman - "There will be some services as we get deep into the agent world that we have to charge much more than $200 a month for, but for ChatGPT I'd like to just offer pay-per-use or packs of some sort beyond some level."

Ethan Mollick - "Interesting data point on OpenAI's Deep Research: I have been getting a steady stream of messages from very senior people in a variety of fields who have been, unsolicited, sharing their chats and how much it is going to change their jobs. Never happened with other AI products."

Pika introduces Pikadditions. You can now add anyone or anything to any video, whether that’s a video you shoot yourself, or a favorite clip.

Topaz Labs - "We're launching Project Starlight: the first-ever diffusion model for video restoration. Enhance old, low-quality videos to stunning high-resolution. This is our biggest leap since Video AI was first launched."

Leading AI researcher and OpenAI co-founder John Schulman has left Anthropic after around five months working at the company

OpenAI employee - "Using Deep Research over the past few weeks has been my own personal AGI moment. It takes 10 mins to generate accurate and thorough competitive and market research (with sources) that previously used to take me at least 3 hours."

Google DeepMind is hiring a Research Scientist/Engineer for automated AI research

Demis Hassabis says that agentic systems this year will act in the world, which will be useful for robotics, but it doesn’t seem necessary for acquiring intelligence. A true understanding of the physical nature of the world can be learned through passive observation of videos and text.

BridgeDP Robotics, a Chinese startup specializing in bipedal control solutions, showcased their system on AGIBOT's A2 humanoid.

Wojciech Zaremba from OpenAI - "Reasoning models are transforming AI safety. Our research shows that increasing compute at test time boosts adversarial robustness—making some attacks fail completely. Scaling model size alone couldn’t achieve this. More thinking = better performance & robustness."

China has opened its first humanoid robot training center in Shanghai's Pudong District. The Kylin Training Ground focuses on robotics, AI, and machine learning, with capacity for 100 robots now and plans to scale to 1,000 by 2027

Imagen 3 is now ranking #1 on the lmarena text-to-image leaderboard, and will be available in the Gemini API and AI Studio soon

Dozens of humanoid robots from over 20 Chinese companies will join a half-marathon in Beijing’s Daxing district this April, running alongside 12,000 human participants on the 21 km (13 miles) route.

Microsoft researchers introduce MatterGen, a model that can discover new materials tailored to specific needs—like efficient solar cells or CO2 recycling—advancing progress beyond trial-and-error experiments.

EngineAI's SE01 humanoid robot walking around among people at the company's office campus in Shenzhen.

OpenAI employee - "too bad the narrow domains the best reasoning models excel at — coding and mathematics — aren't useful for expediting the creation of AGI" "oh wait"

Microsoft paused some construction on a Wisconsin AI data center that OpenAI is slated to use. The company said it needs to evaluate “scope and recent changes in technology,” as well as how “this might impact the design of our facilities.“

Foxconn (aka Hon Hai), the world's largest contract manufacturer of electronics, has partnered with NVIDIA to develop humanoid robots in Taiwan.

The agreement between Microsoft and OpenAI says that AGI would be achieved only when OpenAI has developed systems that have the ability to generate the maximum total profits to which its earliest investors, including Microsoft, are entitled. Those profits total about $100 billion.

New paper presents Exbody2, a whole-body tracking framework enabling humanoid robots to mimic dynamic, human-like motions (e.g., running, dancing) with stability. Trained via RL in simulation and transferred to real robots, it outperforms prior methods.

Stereo4D from Google DeepMind is a clever approach to a persistent problem in computer vision: Getting good training data for how things move in 3D. The key insight is using stereo VR180 videos from the internet. They have now created a dataset of over 100,000 real-world 4D scenes with metric scale

WebDev Arena Leaderboard is now live with 10K+ votes. There is currently a huge gap between Claude 3.5 Sonnet and the rest in WebDev Arena.