Keep pulling the thread on Jensen Huang.
TSMC and SK Hynix are building numerous new chip fabrication plants to meet the demand created by the AI industry.
In NVIDIA's Grace Blackwell supercomputer architecture, an entire server rack is designed to function as a single, unified GPU.
The cost per million tokens for models in the class of GPT-4 decreased by more than 100x in 2024.
The combined innovation in models, algorithms, and computing hardware is causing the cost of AI to decrease by more than 10x annually.
Hanson predicts that the cost-performance of AI computing will improve by a factor of 100,000 to 1,000,000 over the next 10 years.
Hanson predicts that a combination of multi-modality, very long context models, and breakthroughs in synthetic data generation will create a 'ChatGPT moment' for the field of digital biology.
Hanson predicts major breakthroughs in humanoid and multi-embodiment robots, driven by generative AI, multimodal models, and reasoning systems.
The end of Moore's Law is driving a fundamental shift in the computing industry from general-purpose CPUs to accelerated computing architectures.
AI tools like OpenEvidence are now being used as trusted resources by doctors for medical information.
AI tools like Harvey are emerging as trusted tools for legal professionals.
The token generation rate for AI inference, particularly for reasoning-intensive tokens, is growing at a rate of "several exponentials at the same time."
The AI medical information company OpenEvidence reportedly operates with 90% gross margins.