We’re 96% to AGI—Google’s New “Nested Brain” Model Explains Why

Quick Overview

The video discusses Alan's conservative countdown to AGI currently standing at 96%, followed by a review of recent AI advancements, including Google DeepMind's self-modifying language model architecture (TiDAR) that blends diffusion and autoregression, impressive humanoid robot agility demonstrations from China, and a New York law regulating personalized AI pricing, all of which suggest the rapid acceleration toward AGI and ASI.

Key Points: Alan's conservative countdown to AGI increased to 96%, up from the 60s/low 70s previously seen, indicating accelerating progress. Google DeepMind introduced TiDAR, a sequence-hybrid architecture combining diffusion models (for parallel generation) and autoregressive models (for quality), which they claim performs better than large language models like GPT-6. A Chinese humanoid robot, MagicBot Z1, demonstrated advanced agility by side-flipping to dodge an arrow and performing martial arts-style kicks. A New York law targets personalized/surveillance pricing, requiring retailers using AI to disclose that prices are set by an algorithm using personal data. The speaker notes that the complexity of AI models, like the consciousness theory suggesting consciousness is fundamental rather than a byproduct of the brain, is growing rapidly. The video also shows various autonomous cleaning robots in a competition in Shenzhen, China, highlighting practical service robot applications.

Context: The video presents a rapid-fire review of several recent developments across artificial intelligence, robotics, and technology policy, framed around the speaker's ongoing 'conservative countdown to AGI' metric, which suggests the timeline for achieving Artificial General Intelligence is rapidly approaching. Key topics include new AI model architectures, advanced robotics, and emerging regulatory challenges.

Detailed Analysis

The video begins by showing Alan's conservative AGI countdown gauge at 96%, noting an increase from previous levels. The speaker then reviews several recent developments. First, he highlights a paper from Google DeepMind introducing TiDAR: Think in Diffusion, Talk in Autoregression, a hybrid architecture that combines the parallel generation strength of diffusion models with the quality output of autoregressive models, allowing it to draft tokens sequentially while thinking in diffusion, which is claimed to achieve high throughput and quality. Next, the speaker shows footage of a Chinese humanoid robot, MagicBot Z1, demonstrating impressive agility by side-flipping to dodge an arrow and executing high kicks, noting its advanced motor control. The discussion shifts to policy, referencing a new New York law requiring retailers to disclose when personalized pricing (surveillance pricing) is set by an AI algorithm using personal data. The speaker then reviews a theoretical physics paper suggesting consciousness is the foundation of reality, rather than a byproduct of the brain, which he finds interesting but potentially complex. Finally, the speaker shows clips from a sanitation robot competition in Shenzhen, China, demonstrating various autonomous cleaning bots and noting that while these are practical applications, the rapid advancements in AI (like the self-modifying architecture of TiDAR) suggest an accelerating timeline toward AGI, perhaps even earlier than expected.