Google’s Titans Just Solved AI’s Biggest Weakness, But...

Google's AI Innovations: Titans and Myraz [00:02]

• Google is significantly advancing its AI ecosystem with updates to long-term memory systems, introducing a new "nano banana 2 flash" model, experimenting with AI-generated headlines, and boosting Gemini's growth.
• The "Titans" research aims to overcome transformer limitations with long inputs by blending short-term memory (window attention) and a dynamic long-term memory module that updates during use.
• This long-term memory system stores information based on surprise and forgets in a smart way, building three types: memory as context, memory as gate, and memory as layer.
• The "Mac version" of Titans has demonstrated exceptional performance, handling context windows over 2 million tokens and outperforming models like GPT-4 and Llama 3.170B on benchmarks like Babylon.
• "Myraz" is presented as a framework for understanding sequence models, breaking down memory into shape, storage decision, eviction rate, and update mechanisms, leading to new attention-free models like Moneta, Yad, and Mamora.

  "Titans is Google trying to blend both worlds."

Lux: The Next-Generation AI Agent [04:24]

• The Open AGI Foundation has released Lux, a computer vision model that operates directly on user interfaces, outputting clicks, scrolls, and key presses, rather than relying solely on APIs.
• Lux excels at handling real-world UI interactions across desktops, browsers, and applications, achieving an 83.6 score on the Mind 2 Web benchmark, significantly higher than competitors like Gemini CUA (69.0) and OpenAI Operator (61.3).
• It features three levels of autonomy: Actor mode for straightforward tasks, Thinker mode for broader goals, and Tasker mode for deterministic execution with error handling.
• Lux's training utilizes "agentic active pre-training," allowing it to learn through direct interaction within digital environments, leading to a more intuitive understanding of real interfaces.
• The model is also reportedly 10 times cheaper per token than OpenAI's Operator, making long, multi-step automation more economically viable.

  "Lux isn't a chat model with some browser plugin slapped onto it. It actually looks at the screen, reads the UI, and outputs clicks, scrolls, key presses, everything."

Nano Banana 2 Flash and AI Headline Experiments [07:11]

• Google is preparing to release "Nano Banana 2 Flash," a more cost-effective version of its image generation model that offers performance comparable to the Pro version.
• This strategy allows Google to deploy flash models for high-volume scenarios where cost efficiency is paramount, especially for users engaging in high-frequency image generation or large batch workloads.
• In a separate, controversial experiment, Google Discover was observed rewriting headlines for Android users, sometimes drastically altering the original meaning and causing concern among newsrooms about trust erosion.
• Google stated this was a "small UI experiment" to make topics easier to scan, but did not fully address why many rewritten titles deviated significantly from the articles' content.

  "The public announcement is probably coming sometime in December."

Gemini's Rapid Growth and Competitive Landscape [10:13]

• New user data reveals Gemini is experiencing substantial growth, with monthly active users (MAUs) rising 30% in a recent period, significantly outperforming ChatGPT's 6% growth during the same timeframe.
• While ChatGPT still leads in global downloads (around 50%) and MAUs (around 55% or 810 million users), its growth has slowed considerably compared to Gemini's rapid ascent.
• Gemini's MAU share has climbed by 3 percentage points, while ChatGPT's has dropped by 3 percentage points, indicating a shift in market momentum.
• The surge in Gemini's popularity is largely driven by the "nano banana" image generator and its integration within the Android ecosystem, making it the default AI for many users.
• In response to Google's acceleration, OpenAI is prioritizing upgrades and has put advertising experiments on hold, focusing on developing a new model named "Garlic" to compete with Gemini 3.

  "The timing shows how seriously they're taking Google's acceleration."