Google DeepMind Is Starting to Ask What Comes After AI

Morning, {{first name | folks}}! Today, we’re looking at Google DeepMind’s new AGI institute, OpenAI’s legal-focused Astra, Anthropic’s expanded Claude access for life sciences, Figure’s humanoid testing across unfamiliar homes, and Mark Zuckerberg’s take on the AI safety race.

Google DeepMind has launched a new institute focused on the technical and societal questions around increasingly capable AI. The DeepMind Institute brings together researchers from Google DeepMind, Google, and the wider research community to explore AGI, AI safety, economic disruption, model transparency, and how frontier systems should be evaluated.

The institute says its contributors won’t necessarily agree with each other, with the goal of bringing different perspectives into the same discussion. Its first collection of essays covers questions around AGI’s development, safety, economic impact, and how society should respond as these systems become more capable.

OpenAI Builds a Version of Astra Specifically for Lawyers

OpenAI has launched Astra for Law, a version of GPT-6 Astra built specifically for legal work. It combines the model with a legal search index covering U.S. case law, statutes, regulations and court rules, along with tools for legal research, analysis and drafting. The system is designed around the legal information and workflows lawyers already rely on.

Selected law firms can use Astra for Law through ChatGPT and Codex, while legal AI companies including Harvey and Legora will be able to access it through the API. The move shows how AI companies are starting to build more specialised systems around specific industries, rather than relying on one general-purpose model for every type of work.

Anthropic is launching a new program that gives verified life science teams broader access to Claude for biology-related work. The Life Sciences Verification Program is aimed at researchers working across drug discovery, biology, clinical development, and manufacturing. Access will depend on research credentials, security standards, and the type of work being done.

The important part for builders is what this looks like at the product level. A biology AI tool may need to verify who is using it, limit certain capabilities, and route higher-risk requests for additional approval. Instead of treating safety as a single model setting, the product itself becomes part of how access is controlled.

Figure has tested Helix 2.5 across 30 homes the humanoid had never seen before. The robot completed three whole-body tasks across different layouts and environments, including homes with no prior training data. Figure ran the same model across all 30 locations without collecting new data or adapting it for each home.

Figure says Helix 2.5 was able to transfer the same skills across the different environments without additional training. That’s an important step for robots moving from controlled demos into real homes, where every environment is different.

Meta is making the case that AI labs can keep pushing capabilities while putting more weight on safety and alignment. Mark Zuckerberg argues that labs already have strong reasons to build models people can trust, from user demand for aligned agents to legal liability when systems cause harm.

Zuckerberg also points to Meta delaying Muse to work on safety and security, using independent evaluators, and putting most of its compute toward serving users rather than recursive self-improvement. His broader argument is that keeping the balance of power between AI labs, rather than collectively slowing progress, is part of building a safer future.