Help make better ads

Did you recently see an ad for Roku Ads Manager in a newsletter? We’re running a short brand lift survey to understand what’s actually breaking through (and what’s not).

It takes about 20 seconds, the questions are super easy, and your feedback directly helps us improve how we show up in the newsletters you read and love.

If you’ve got a few moments, we’d really appreciate your insight.

Take the survey.

Today:

• Google Unveils Instant AI Avatars for Workspace 

• DeepMind's Cybersecurity Blueprint for AI Agents 

• OpenAI's Autonomous Chemist Boosts Drug Discovery 

• DeepMind Loses Nobel Winner to Anthropic 

• Amazon Cancels Sam Altman Movie 'Artificial'

🎬 Google Workspace Launches Free AI Video Avatars

Google Workspace just introduced a massive update for content creators and marketers: the ability to turn slide decks directly into fully voiced video presentations without a studio budget.

Key Details:

• Instant Generation: Available starting today at vids.new, the platform allows anyone to generate product demos and presentations using digital human avatars.

• Multilingual Support: Users can direct the avatars and generate voiceovers across 24 different languages.

• Accessibility: The AI avatar generation features are rolling out at no cost.

The Takeaway: For rapid content turnarounds, this shifts the production workflow away from heavy staging and traditional editing, allowing you to generate globalized video presentations directly from your slides.

🛡️ DeepMind's "Defense-in-Depth" AI Control Roadmap

As agentic AI becomes more capable and integrates deeper into internal workflows, Google DeepMind (GDM) has released its "AI Control Roadmap (v0.1)" to safeguard against potential adversarial behavior by highly autonomous models.

Key Details:

• Insider Threat Model: GDM is treating future highly autonomous AI under a cybersecurity "insider threat" framework. Using their new TRAIT&R taxonomy, they categorize risks into three main areas: rogue deployments, work sabotage, and direct asset harm.

• Core Defensive Invariants: The system relies on two unbreakable rules: reliably detecting misaligned intent, and effectively preventing or intercepting attacks before they can be executed.

• Escalating Tiers: Defenses are scaled to match model capabilities, broken down into four Detection tiers and three Prevention/Response tiers.

• Concrete Mitigations: The roadmap outlines 15 specific safeguards, ranging from simple chain-of-thought monitoring for current models to advanced real-time access controls and "shutdown infrastructure" for future, more powerful iterations.

The Takeaway: DeepMind is acknowledging that as models become more autonomous, pure alignment isn't enough. We need robust, automated, system-level safety nets to catch misaligned actions in real time.

🧪 OpenAI & Molecule.one Build a "Near-Autonomous" AI Chemist

In a major milestone for medicinal chemistry, OpenAI partnered with Molecule.one to use GPT-5.4 to successfully optimize a notoriously difficult drug-making reaction.

Key Details:

• The Target: The system was tasked with improving the Chan-Lam coupling reaction—specifically for primary sulfonamides, which historically produce low yields and limit the types of carbon-nitrogen bonds researchers can test for new medicines.

• The Workflow: GPT-5.4 operated as the lead researcher. It reviewed existing literature, generated thousands of hypotheses, ranked the best experiments, and analyzed the results.

• Physical Execution: Human chemists validated the top proposals, which were then executed by "Maria Lab," an automated high-throughput laboratory.

• The Results: Across 10,080 reactions over three months, the AI's top proposal (adding a chemical called TEMPO, later optimized to a much cheaper variant) successfully improved yields for 88% of tested boronic acids and 83% of tested sulfonamides. The average yield jumped significantly from 16.6% to 25.2%.

The Takeaway: Instead of just generating text or pixels, frontier models like GPT-5.4 are now successfully acting as reasoning engines for physical robotics—drastically accelerating the trial-and-error bottleneck in pharmaceutical research.

🧠RESEARCH

Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance

Moebius is a new AI tool for inpainting—smartly filling in missing parts of an image. It matches the quality of massive industry AI programs but is built at a fraction of their size. This smaller design allows it to run 15 times faster, making rapid visual editing much more practical.

DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects

DragMesh-2 helps robotic hands skillfully grasp and move articulated objects—items with moving parts like hinges or handles. Usually, robots struggle if how heavy something feels changes during work. This clever new system uses built-in physics rules so robots can reliably adapt to changing physical forces without needing extra touch sensors.

ENPIRE: Agentic Robot Policy Self-Improvement in the Real World

ENPIRE is a system that lets AI independently write code to train new physical robots. It creates an automated loop where the AI tries a physical task, checks the result, resets the room, and actively improves its own instructions—called a policy—achieving a 99% success rate without requiring any human supervision.

📲SOCIAL MEDIA

🗞️MORE NEWS

John Jumper Leaves Google DeepMind for Anthropic Nobel Prize winner John Jumper, the co-creator of the groundbreaking biology AI AlphaFold, is leaving Google DeepMind to join rival company Anthropic. The exit is a major loss for Google's scientific team but strongly boosts Anthropic's push into medical and biological research.

Amazon Drops Sam Altman Movie 'Artificial' Amazon MGM Studios has canceled the release of Artificial, a nearly finished movie about OpenAI's 2023 company leadership crisis starring Andrew Garfield as CEO Sam Altman. The studio abandoned the project despite positive audience test screenings, a move that closely follows Amazon's new $50 billion business partnership with OpenAI.

Hyundai Takes Full Control of Boston Dynamics Hyundai has bought out SoftBank's remaining share of Boston Dynamics for $325 million to take complete ownership of the robotics company. The deal paves the way for Hyundai to put human-shaped robots to work in its own car factories by 2028, while SoftBank shifts its investments toward building physical data centers.

The Atlantic Uncovers Massive AI Music Database A new investigation revealed four massive databases containing tens of millions of songs that developers secretly used to teach AI systems how to generate new music. Programmers used automated tools to copy protected tracks from sites like YouTube and Spotify without permission, fueling ongoing lawsuits from major record labels.