LLMs consume more text than any human could in a lifetime yet remain worse at language than a 5-year-old, evidencing that human language learning is fundamentally different from LLM training.

Ethereum's real economic security is only $5 million per day, not the $100 billion figure commonly cited.

The US must keep real rates significantly negative to avoid a debt death spiral, and current Treasury buyback actions are a down payment toward that goal.

A method that worked once is not necessarily the right way; it is just one way that happened to work in that instance.

Corporate income as a share of gross domestic income is at an all-time high, comparable to dot-com era extremes in 1999.

AI is democratizing access to technology globally while simultaneously concentrating the epicenter of innovation in the Bay Area.

Neuroscience researchers at OpenAI and Anthropic have noted it is easier to do neuroscience on AI models than on biological brains, suggesting AI models are now the most productive substrate for neuroscience research.

LLMs are inadequate for taste-based product search because they regress to the mean and return the most probable result, not the personally relevant one.

Consumers do not evaluate live shopping as a marketplace or category; they only care about whether they will enjoy the experience.

AI creates a 'tragedy of the cognitive commons': checking AI output requires deep expertise, deep expertise comes from doing grunt work for years, and grunt work is the first thing AI automates away, so organizations are building systems that need expert supervision while dismantling the process that creates experts.

Dario Amodei predicts that AI will make it possible to cure most human diseases within 5 to 10 years.

When the team that defines risk is separate from the team building and running the systems that created it, misalignment is inherent to the design.