DeepMind Leaked Possibly the Greatest AI Ever

Google's AI Advancements: Gemini 3 and Beyond

Introduction to Gemini 3

• Google released the Gemini 3 preview, highlighting its robust pre-trained backbone and potential for improvement through post-training.

• A new update to Gemini deep thing has been announced, showcasing impressive benchmarks such as nearly 3,500 ELO in coding, surpassing Claude Opus 4.6.

Key Developments in AI Models

• Isomorphic Labs introduced ISO DDE, which significantly enhances Alpha Fold 3's accuracy, marking a major advancement in drug discovery.

• Google Kaggle launched Game Arena, allowing AI models to compete head-to-head beyond traditional benchmarks.

Capabilities of New AI Models

• Multiple variants of Gemini 3 are being tested within Game Arena; all are noted for their speed and intelligence.

• Genpark's AI workplace has achieved a $155 million annual run rate in just ten months with its latest features.

Practical Applications of AI

• The speaker demonstrated how Jensen Spark utilized various AI tools to create a web page summarizing an episode of the Huberman Lab podcast.

• An innovative story was generated from the podcast content aimed at children, showcasing the versatility of AI in creative storytelling.

Future Prospects and Comparisons

• There is uncertainty about what the next version of Gemini will be called; however, its capabilities are expected to be superior compared to other models.

• The new Gemini excels at creating SVG graphics compared to Cloud Opus 4.6 due to its mathematical representation that maintains quality upon zooming.

Multimodal Intelligence and Visualization

• The ability of Gemini to generate complex visualizations indicates strong multimodal training across images, videos, and text.

• Despite not being the top coding model overall, Gemini shows promise in visualizing concepts effectively due to its broad training approach.

This structured summary captures key insights from the transcript while providing timestamps for easy reference.

AI Simulations and Innovations in Technology

Impressive AI Simulations

• The current AI simulation of the iPhone is noted for its accuracy, replicating the design of apps like the calculator and notes app closely.

• This simulation stands out as a complete replica, with most applications functional or serving as mockups for future development; local apps operate smoothly.

• The creator, Chetas Lua, has gained credibility over time, leading to speculation about the capabilities of upcoming models like Gemini 3.

Advancements in AI Models

• Gemini Deep Think is highlighted as a specialized model that employs parallel thinking and consensus to solve complex problems but comes at a high cost.

• This model achieved remarkable performance metrics in various competitions, including gold medal levels in math challenges and high ratings in humanities exams.

• A demo showcases Gemini Deep Think's ability to convert sketches into 3D printable files, demonstrating practical applications in science and engineering.

Breakthroughs in AI Video Generation

• Seance 2.0 represents a significant advancement in AI video generation, producing content indistinguishable from human-created media.

• The speaker expresses mixed feelings about this technology's potential impact on society, noting it could lead to both excitement and fear regarding authenticity.

Evolution of Creative Content

• Software engineers are experiencing heightened awareness of advancements due to new releases like cloud opus 4.6 and GPD 5.3; broader recognition may follow soon.

• An example from the Madvid Project illustrates how well AI can blend styles from different shows (e.g., Game of Thrones and Rick and Morty), showcasing impressive voice clarity and art style fidelity.

Innovations Beyond Alpha Fold

• Isomorphic Labs has announced advancements beyond Alpha Fold 3, which initially focused on predicting protein structures crucial for understanding biological functions.

• The new system predicts interactions between proteins and other molecules—critical for drug development—enhancing our understanding of medicinal chemistry significantly.

• ISODE doubles prediction accuracy for drug binding sites on proteins while identifying previously unknown binding locations quickly through computational methods.

AI Models Competing in an Arena

Overview of AI Model Competitions

• The concept involves testing AI models' capabilities by having them compete against each other in a controlled environment, referred to as an "arena."

• Currently, three games are available for these competitions: Poker Heads Up, Werewolf, and Chess.

• Observing the competition highlights significant advancements in AI intelligence compared to previous generations, which may not be evident through standard benchmarks.

Evolution of Game Performance

• A year ago, the performance of AI models in a werewolf game was poor; they often hallucinated statements from other players and lost track of conversations.

• Current iterations show improvement; the werewolf game has become more engaging and enjoyable to watch.

• Despite improvements, AI models still struggle with social interaction and deception skills.