chatgpt

Google’s Genie: Is It LeCun’s Way or Yours?

Posted on by clementon
Google's Genie: Is It LeCun's Way or Yours?

The demo started innocently enough: a digital paintbrush sketching a forest onto the screen. Seconds later, the trees swayed in a nonexistent breeze, and a blocky avatar began to explore. But as the presenter zoomed out, revealing the entire “world” fit neatly inside a smartphone screen, a chill snaked down my spine – was this a glimpse into the future or just a sophisticated illusion?

Google is dabbling in something new, but the velvet rope is up. This week, they released Project Genie, calling it a “general-purpose world model” capable of generating interactive environments. After a small, invite-only test group played with it last August, Genie 3 is now available to Google AI Ultra subscribers in the US…for a mere $250 (€230) per month.

The Allure of “World Models”

Consider the difference between memorizing a map and actually driving the route. Large language models (LLMs) like Google’s own Gemini recognize patterns, but world models simulate environments, learning physics and spatial relationships. It’s a different approach to AI.

Yann LeCun, formerly the chief scientist at Meta AI, champions world models. LeCun believes LLMs won’t achieve true artificial general intelligence. He recently joined a startup betting big on world models. Think of it this way: LLMs spot trends, while world models run simulations to deduce how the world functions and predict outcomes.

Google’s entry validates the potential of world models. Preview videos from Project Genie’s early stages show visual promise, despite being brief. Google limits users to 60-second world generations that “might not look completely true-to-life or always adhere closely to prompts or images, or real-world physics”—in other words, results vary. Outputs are 720p videos at 24 frames per second, according to Ars Technica, and some users report lag issues.

That’s expected in beta. However, it hints at limitations: the world might be smaller than expected. Claims that this will eliminate video game developers may be premature. For now, treat Project Genie like a prototype, not a product.

How does Project Genie relate to Yann LeCun’s world model vision?

Google’s Genie 3 diverges from LeCun’s vision. Project Genie creates continuous, video-based worlds navigable like video games, enabling AI agents to learn by exploring. LeCun’s Meta concept, Joint Embedding Predictive Architecture (JEPA), embeds a world model within the AI agent itself.

Copyright Concerns and the Limits of Scale

The path isn’t without its thorns. Google will face the same copyright issues plaguing other image and video generation models like OpenAI’s Sora 2. Early Project Genie outputs frequently replicate Nintendo worlds, which will raise red flags, fast. Still, Google’s investment suggests that even AI leaders see potential limits to LLMs. The company is planting seeds for a future beyond pattern recognition.

What are the limitations of Google’s Project Genie?

The project has a cap for a reason. If training a text-based model is expensive, imagine simulating an entire world. It requires massive data to understand visuals and physics, plus enormous processing power. Scaling it presents significant hurdles. So, for now, the worlds may seem expansive, but they’re practically tiny. Project Genie is like a goldfish in a bowl–it looks bigger than it is.

What is the Difference Between LLMs and World Models?

LLMs are like encyclopedias; they contain vast amounts of information. World Models are more like physics engines. The former can repeat and reassemble information; the latter understands how that information actually interacts. This approach may be what’s needed to truly mimic human intelligence.

Will world models like Project Genie escape their constraints, or remain a costly curiosity?