🚨 China called out a Robot Bubble

Good morning, AI enthusiast. China just dropped a rare economic warning, a humanoid robot bubble may already be forming as money floods into the sector with very few real-world use cases.

In today’s AI newsletter:

• China warns of a humanoid robot bubble

• A New AI Can Play 3D Games

• Quantum leap in biotech modeling

• How to generate editable design layouts from a single prompt

• AI tools & more

ROBOTICS

🤖 China Warns of a Humanoid Robot Bubble

China’s top economic planning agency says investment in humanoid robotics is growing too fast with too little substance, raising fears the industry could mirror the instability seen in the broader AI boom.

• Over 150 humanoid robot companies now operate in China

• Many models are “highly similar” due to shrinking R&D budgets

• More than half the companies are startups or firms pivoting into robotics

• Beijing designated embodied intelligence (tech behind humanoid robots) as a national priority earlier this year

• Officials urge balancing rapid growth with the rising risk of a bubble

Humanoid robots are becoming a prestige race, but not a proven business. With massive investment, little differentiation, and few practical applications, China’s caution signals a reality check for global robotics hype, and a reminder that not every AI-adjacent boom is built on solid ground.

AI RESEARCH

🧠 A New AI Can Play 3D Games

Researchers trained a VLA (Vision-Language-Action) agent, basically a chatbot that predicts actions instead of words, to handle complex, hours-long missions inside Genshin Impact. The result? An AI that completes the full five-hour Mondstadt storyline and generalizes to brand-new regions, puzzles, bosses, and even entirely different games.

• Trained Qwen2VL-7B to play Genshin using 2,424 hours of human gameplay

• Zero reinforcement learning, only supervised fine-tuning

• Runs at 5 Hz controls with sub-200 ms action generation

• Generalizes to unseen locations, missions, characters, items

• Shock result: plays Honkai: Star Rail for 7 hours zero-shot

• Also plays Wuthering Waves and even Wukong (navigation decent, combat rough)

Lumine is the clearest proof yet that large-scale action data + lightweight reasoning can produce generalist agents. Instead of training a model per game or task, we might be entering the era of universal AI players that navigate any digital world with human-level control.

BIOTECH

🧬 Quantum Leap in Biotech Modeling

A Japanese research team has shown that a quantum algorithm can solve a core metabolic modeling problem, one of the very first real applications of quantum computing to a biological system.

• Researchers used quantum interior-point methods to model how cells move energy and resources

• The quantum approach successfully reproduced classical results for glycolysis + the TCA cycle

• The demo ran on a simulator using 6 qubits, aligned with what early fault-tolerant machines could handle

• This could help scale metabolic simulations to whole cells or even microbial communities

Modeling metabolism at large scale is brutally hard for classical computers, the matrices get huge, complex, and slow. If quantum machines can eventually speed up these simulations, they could unlock faster biotech research: designing engineered microbes, understanding diseases, or even mapping emerging therapies. This is still early, but it’s a big signal that quantum computational biology is starting to get real.

HOW TO AI

💻 How to Generate Editable Design Layouts from a Single Prompt

In this tutorial, you’ll learn how to generate fully editable design layouts using Google Mixboard, an AI tool that transforms simple text prompts into complete posters, ads, and web sections, giving you multiple professional compositions you can tweak, refine, and export in minutes.

🧰 Who is This For

• Designers who need fast layout drafts

• Founders creating ads, landing pages, and brand assets

• Marketers testing visual concepts

• Anyone who wants high-quality design without opening Figma

STEP 1: Access Google Experiments

Mixboard is part of Google’s experimental AI suite. Head over to the Mixboard page and you’ll land on the main dashboard. Here, look for the “New Project” button and start a new project. You’ll see built-in examples and a prompt box at the bottom of the screen where you’ll type your instructions.

STEP 2: Describe Your Vision

In the prompt box at the bottom, write a natural language description of the asset you want to create. Mixboard doesn’t work like a typical image generator; it understands layout intent, structure, spacing, and hierarchy.

For this tutorial, you can try something like:

• “A minimalist ad poster for a new noise-canceling headphone brand, using dark blue tones and modern typography.”

• “A vibrant 3-column website layout for a summer music festival.”

Hit Enter or press Generate to continue.

STEP 3: Review the Layout Options

Give Mixboard a moment to process. Instead of giving you one image, it will pitch multiple concepts, each with a different composition and design direction. One variation might center the headline, another might move the product to the bottom-right, another might use illustration instead of photography.

Click the version that best matches the structure and aesthetic you’re aiming for. This becomes your editable canvas.

STEP 4: Tweak and Finalize

The real power of Mixboard is that everything it generates is editable. Click on any element and you’ll see customization options.

You can adjust typography with one click, remix color palettes, replace background visuals, or swap out product photos without breaking the layout. Once you’re satisfied, export the design and use it immediately.

Stanford researchers develop a web-based tool that uses an LLM to downrank X posts with antagonistic language in a user's feed, to reduce “partisan animosity”.

The USPTO issues new guidelines clarifying that generative AI is considered a tool that assists human inventors and cannot itself be considered an inventor.

OpenAI projects ChatGPT will reach 220M paid subscribers by 2030, which would place it among the world's largest subscription totals alongside Netflix and Spotify.

Perplexity just launched a virtual try-on feature that lets you see clothes on a digital avatar of yourself, along with a cleaner, updated shopping experience.

🤖 Claude Opus 4.5: Anthropic’s new flagship model for coding, agents, and real computer use

🍌 Nano Banana 2: Google’s Gemini 3 image model with perfect text and character consistency

🎆 FLUX.2: Black Forest Labs’ new smart vision model

✨ Gemini 3: Google’s next-gen engine for multimodal reasoning

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas