Keep pulling the thread on Mike Schreppfer.
In many deep tech applications like materials science, AI-driven discovery addresses less than 10% of the total time and cost required to bring a new product to market, with manufacturing scale-up being the primary challenge.
The United States needs to increase its power grid capacity by approximately 5 times by 2050 to meet its goals, even without accounting for AI-driven demand.
In 2024, utility-grade solar power is expected to account for 80% of all new energy generation added to the grid in the United States.
Meta has issued a Request for Proposal (RFP) for the development of next-generation nuclear fission power plants.
The primary demand for AI compute is shifting from model training to inference, particularly for advanced reasoning models.
Meta's foundational AI strategy, including for projects like PyTorch and Llama, was to open-source the technology to ensure broad access and accelerate industry-wide progress.
Implementing a known software algorithm in specialized hardware typically yields a 10x advantage in performance per watt or price.
The strategy of creating custom silicon (ASICs) carries the significant risk that the chip could become worthless if the underlying algorithm it was designed for changes before the chip is released.
The AI industry is hitting diminishing returns from simply scaling up large language models with more data and pre-training.
Current LLMs still lack a key human-like capability: associative long-term memory.
Pantolosa is developing a 200-meter tall offshore compute platform that generates energy from ocean waves and uses seawater for immersive cooling.
The market demand for AI inference and reasoning compute is effectively infinite if the price can be made low enough.