This week, we dive into SIMA 2, the new agent from DeepMind/Google that learns by acting in video games and virtual worlds. Then we tackle the controversial venture backed by Sam Altman which aims to edit human embryos to prevent inherited disease. Let’s dive in ⬇️

In today’s newsletter ↓
🎮 The rise of virtual-world agents
🧬 Embryo editing enters Silicon Valley
🛰️ Big platform data rule shakeup
🤦 Russian humanoid robot face plants on stage
🏆 Weekly Challenge: Use a new multimodal assistant

SIMA 2 is the latest from DeepMind, an AI agent built to reason and act inside full 3D virtual worlds. Unlike earlier systems that simply followed instructions, SIMA 2 integrates the advanced reasoning of the Gemini model with embodied interaction (moving, choosing, tool use) across unfamiliar game environments.

It is not restricted to any single engine or map; instead, it demonstrates the ability to generalize across multiple commercial game titles, each with different physics, textures, objectives, and visual styles. Early testers note that the system also interprets more abstract instructions, such as navigating to a landmark based on metaphorical descriptions or performing a sequence of subtasks with minimal prompting.

DeepMind says SIMA 2 can pick up tasks in games it has not seen before, such as identifying a red house when asked “go to the tomato colored house” or navigating newly generated terrain.

The agent uses Gemini to create and evaluate new data: when it enters a new world it asks a model to generate tasks, scores itself, then fine tunes accordingly. This loop lets the agent expand its skill set without relying on hand crafted missions or dense human supervision.

Virtual worlds serve as scalable, safe, and controllable environments for training embodied intelligence — the kind of intelligence necessary for robotics, AR and VR, and assistants that move beyond chat. DeepMind sees this as a fundamental step toward generalist agents.

These environments simulate many of the same challenges real robots face: partial observability, complex navigation, object manipulation, and unpredictable sequences of events.

Despite the leap, SIMA 2 is still research only. It struggles with very long horizon goals, precise low level motor control, and maintaining large memory of interactions. These remain bottlenecks before moving into the physical world. DeepMind acknowledges that transferring skills from games to real sensors and actuators is still an open research problem.

If the approach scales, we could see agents that assist humans across dynamic environments, from factory floors to virtual workspaces. Because the same underlying reasoning and action skills apply whether you are in a “game” or real life, SIMA 2 may represent one of the more concrete bridges from simulated environments to real world applications of general purpose AI.

A startup named Preventive, backed by Sam Altman, Brian Armstrong, and several other high-profile tech investors, is making waves by pursuing embryo-level gene editing aimed at preventing severe inherited disease before a child is born. The firm has raised roughly US $30 million and is incorporated as a public-benefit corporation, positioning itself as pursuing broad social value rather than pure profitability.

Preventive’s goal is straightforward in concept and extremely complex in practice: edit human embryos at the earliest stage so that children born from them no longer carry known disease-causing genetic variants. The company focuses on severe monogenic disorders — conditions where a single mutation reliably causes a serious disease. In theory, correcting those variants could eliminate the risk entirely for that child and all future generations.

Because embryo editing for reproduction is prohibited in the United States and the United Kingdom, the startup is reportedly exploring regulatory pathways in countries with more permissive frameworks, such as the UAE. Public filings and reporting suggest Preventive is still in the preclinical R&D stage, testing editing accuracy, off-target effects, and embryo viability while preparing for eventual regulatory submissions.

Supporters argue the technology could eliminate devastating conditions that affect millions of families, shifting medicine from treatment to true prevention. They compare embryo editing to preimplantation genetic testing, but with the added ability to correct harmful variants rather than merely select among embryos that do or do not carry them.

Critics warn that editing embryos is inherently heritable, meaning changes pass to all descendants, and today’s tools cannot guarantee long-term safety. Off-target mutations, mosaicism, or unintended trait interactions could create new problems that would be impossible to reverse.

This moment sits at the intersection of biotech, ethics, policy, and Silicon Valley scale. If embryo editing becomes commercially viable, it could redefine how society thinks about disease prevention, reproductive autonomy, and inherited traits.

And because tech capital tends to accelerate every field it touches, the debate may move far faster than regulators, ethicists, or the public are prepared for — setting up a generational question about what kinds of children future parents might be allowed, or even encouraged, to create.

🎮 Apps must spell out when they send your data to outside AI tools

🪟 Windows 11 gets a Click To Do helper that reacts to your screen

🦿 Russia’s flagship humanoid robot face plants on stage

🫆 U.S. Senator warns of ‘destruction’ via AI cyberattacks

💸 The $1 billion bet against U.S. AI companies

📸 Google Messages adds an AI remix mode for group photo experiments

Here’s what to do:

🗂️ Pick a realistic workflow (reviewing a slide deck, reading a long article, processing multiple screenshots).

🖱️ Use Click To Do to select an element (text, screenshot, UI widget) and ask it to perform three different actions:

🖼️ Then switch and ask it to process a non text element (image, design mock up, or UI layout) and request insights or suggestions.

👍 Record how many smart actions actually helped you and how many were off target or irrelevant.

⚖️ Give yourself a helpful to wasted ratio. If fewer than half of the actions were useful, iterate and ask: what context or framing would have made the tool respond more accurately?

That’s it for this week! Is SIMA 2 the next logical step towards the Matrix? And, do you think gene editing embryos is good or bad for humanity? Hit reply and let us know your thoughts.

Zoe from Overclocked