🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Current robot learning methods are good at imitating tasks seen during training, but struggle to compose behaviors in new ways. When training imitation policies, we found something surprising—using temporally-aligned task representations enabled compositional generalization. 1/

What does temporal alignment mean? When training, our policy imitates the human actions that lead to the end goal 𝑔 of a trajectory. Rather than training on the raw goals, we use a representation 𝜓(𝑔) that aligns with the preceding state “successor features” 𝜙(𝑠). 2/

New release! Great for a short, high-level overview of a variety of different areas within AGI safety that we're excited about. x.com/vkrakovna/stat…

We start with a 4-minute introduction and outline of the course youtube.com/watch?v=rHDAzE…

Learn more about AGI safety in a new short course from Google DeepMind alignment team! Check out the short 5-minute videos in thread, or see full playlist here youtube.com/playlist?list=…

we're releasing a short course on agi safety & alignment -- hope it helps! thanks team for pulling it together, special thanks to Victoria Krakovna deepmindsafetyresearch.medium.com/introducing-ou…

We're hiring

Founders who were PhD or post-doc in my lab at Berkeley, **largely funded by NSF / DoD grants**, start-up, market cap (collected by OpenAI Deep Research)

The native image generation launch was a lot of work from a safety POV. But I'm so happy we got this functionality out, check this out:

We built an AI assistant that plays Minecraft with you. Start building a house—it figures out what you’re doing and jumps in to help. This assistant *wasn't* trained with RLHF. Instead, it's powered by *assistance games*, a better path forward for building AI assistants. 🧵

RLHF is great but it encourages short-term optimization: trying to solve the user's entire problem in a single response. For example, if you ask ChatGPT to "clean up some disk space," it will immediately give you a program to run without asking which files are okay to delete!

A better assistant would maintain *uncertainty* about its goal and ask clarification questions until it really understood, leading to a better solution. Assistance games can enable this.

Unlike RLHF, assistance games explicitly treat the user-assistant interaction as a two player game, where the user knows their goal but the assistant doesn't. AGs model *communication* about the goal from the user to the assistant and *collaboration* between them to achieve it.

Our new RL algorithm, AssistanceZero, trains an assistant that displays emergent helpful behaviors like *active learning* and *learning from corrections*.

Check out the full paper for more details. And download our code to play with the assistant! arxiv.org/abs/2504.07091 github.com/cassidylaidlaw… Joint work with Eli Bronstein Timothy Guo Dylan Feng Lukas Berglund Justin Svegliato Anca Dragan

Per our Frontier Safety Framework, we continue to test our models for critical capabilities. Here’s the updated model card for Gemini 2.5Pro with frontier safety evaluations + explanation of how our safety buffer / alert thresholds approach applies to 2.0, 2.5, and what’s coming.

A really fun project led by Joey Hong, with Anca Dragan Website: jxihong.github.io/pnlc_website/ Paper: arxiv.org/abs/2505.18098