Dario Amodei, CEO and co-founder of Anthropic, discusses AI progress three years after his previous interview. He maintains his 2017 'Big Blob of Compute Hypothesis' that raw compute, data quality, and scalable objective functions matter more than clever techniques - a view later echoed in Richard Sutton's The Bitter Lesson.

The conversation covers Anthropic's explosive revenue growth from zero to projected $10 billion annually, the company's approach to responsible scaling, and Dario's predictions for achieving 'country of geniuses in a data center' capabilities within 1-3 years. He references his essay Machines of Loving Grace extensively when discussing economic implications and policy considerations.

Key topics include the current state of AI capabilities, from coding agents to computer use, the challenges of economic diffusion versus technical progress, and geopolitical considerations around AI development. Dario also addresses questions about business models, constitutional AI, and the balance between AI safety and ensuring broad access to AI benefits.

The Big Blob of Compute Hypothesis Continues to Hold

Dario's 2017 hypothesis remains unchanged: only seven factors matter for AI progress - raw compute, data quantity, data quality/distribution, training duration, scalable objective functions, and numerical stability, echoing themes later formalized in The Bitter Lesson.

Pre-training scaling continues delivering gains, with RL scaling now showing the same log-linear returns: 'we train the model on math contests, AIME or other things. And how well the model does is log linear and how long we've trained it' - Dario.

The technology exponential has progressed as expected, moving models from 'smart high school student to smart college student to like, you know, beginning to do PhD and professional stuff' - Dario.

Sample Efficiency Puzzle: Evolution vs Human Learning

AI models require vastly more data than humans but show strong in-context learning with long contexts, suggesting pre-training exists 'somewhere between human learning and human evolution' - Dario.

Unlike humans who aren't blank slates (as explored in The Blank Slate), AI models 'literally start as random weights, whereas the human brain starts with all these regions, connected to all these inputs and outputs' - Dario.

Continual learning may not be necessary: 'I think continual learning, as I've said before, might not be a barrier at all' since pre-training and RL generalization may suffice - Dario.

Anthropic's 10x Annual Revenue Growth Trajectory

Revenue has grown 10x annually: 'zero to 100 million in 2023, 100 million to a billion in 2024, billion to like nine or 10 billion in 2025' with expectations this curve will bend but remain fast - Dario.

Claude Code drives adoption despite enterprise diffusion delays: 'any given product will get adopted by individual developers who are on Twitter all the time, by Series A startups many months faster than by big financial companies' - Dario.

Internal productivity gains are 'really unambiguous' with 15-20% total factor speedup currently, up from 5% six months ago, creating competitive advantages - Dario.

Computer Use and End-to-End Task Automation

Computer use benchmarks have climbed from ~15% to 65-70% on OS World in just over a year, approaching reliability thresholds for deployment.

Dario predicts models will do 'SWE end-to-end' within 1-2 years, including 'setting technical direction and understanding the context of the problem' - not just code generation.

Video editing and similar creative tasks requiring context learning over months will be achievable 'when we have the country of geniuses in a data center' in 1-3 years - Dario.

Economic Diffusion: Fast But Not Infinitely Fast

As outlined in Machines of Loving Grace, even with powerful AI, 'how long does it take to cure all the diseases?' involves manufacturing, regulatory approval, and global distribution challenges - Dario.

Diffusion will be 'much faster than any previous technology, but it has its limits' due to enterprise procurement, security compliance, and change management requirements - Dario.

Clinical trials will accelerate because 'most clinical trials fail because the drug doesn't work' but still require time for safety validation and manufacturing scale-up - Dario.

Responsible Scaling and Compute Investment Strategy

Anthropic's 'responsible' approach means careful analysis rather than absolute spending limits: 'some of the other companies have not written down the spreadsheet, that they don't really' think through the economics - Dario.

The industry will scale to 'multiple trillions a year by 2028 or 2029' with roughly 3x annual compute growth reaching 300+ gigawatts - Dario.

Profitability timing depends on demand prediction accuracy: 'if you're off by only a year, you destroy yourselves' when buying compute infrastructure ahead of revenue - Dario.

Geopolitical Implications and Democratic Values

Export controls on AI chips to China remain difficult despite bipartisan support: 'there's so much money riding on it. And, you know, that money wants to be made' - Dario.

Dario worries about 'offense-dominant' scenarios and hopes 'dictatorships become morally obsolete' as AI makes authoritarian control more dangerous and less sustainable.

Initial conditions matter for setting 'rules of the road' - democratic nations should 'hold the stronger hand' during critical negotiation windows as AI capabilities advance - Dario.