Dario Amodei

AI summary of what Dario Amodei said

Dario Amodei says his predictions about AI writing code have been repeatedly misunderstood, and lays out a spectrum of milestones from models writing 90% of code to handling end-to-end software engineering tasks. He argues that even when models write all the code, software engineers will move to new higher-level work rather than lose their jobs. He acknowledges AI's effects take time to diffuse through the economy but insists the impact will arrive very fast, citing Anthropic's roughly tenfold annual revenue growth.

• He says his prediction that AI models would write 90% of code was misunderstood as a prediction that 90% of software engineers would no longer be needed, when these are very different benchmarks.
• He lays out a spectrum of milestones, from models writing 90% of code, to 100% of code, to handling end-to-end software engineering tasks including compiling, setting up clusters and testing, and says even then engineers move to higher-level work.
• He says he agrees the effects take time to diffuse through the economy, but argues the effect will be very fast, pointing to Anthropic's roughly 10x annual revenue growth from 2023 to 2025.

AI summary of what Dario Amodei said

Speaking at an AI summit in India, Dario Amodei says AI is approaching a "country of geniuses in a data center" and will surpass most humans' cognitive capabilities within a small number of years. He describes both vast opportunities, like curing disease and lifting billions out of poverty, and risks including autonomous AI behavior, misuse, and economic displacement. He announces Anthropic's new Bengaluru office and partnerships in India, and calls for cooperation with the Indian government on safety testing and managing economic disruption.

• He says AI has advanced on an exponential curve for the last ten years and that models will surpass most humans' cognitive capabilities for most things within a small number of years.
• He outlines both opportunities, such as curing diseases and lifting billions out of poverty, and risks, including autonomous AI behavior, misuse, and economic displacement.
• He announces Anthropic's new Bengaluru office, its India managing director, and partnerships with Indian enterprises and nonprofits to spread AI's benefits across the global south.

AI summary of what Dario Amodei said

Dario Amodei says the underlying exponential progress of AI technology has gone roughly as he expected over the past three years, with models moving from smart high school student level toward PhD and professional level. He says the biggest surprise has been the lack of public recognition of how close we are to the end of the exponential. He also explains that he still holds his 2017 'big blob of compute' hypothesis, and that RL is now scaling the same way pre-training did.

• He says the exponential progress of the technology has gone about as he expected, with plus or minus a year or two in places.
• He says he still holds his 2017 'big blob of compute' hypothesis, which lists a few things that matter, like raw compute, quantity of data, quality and distribution of data, training duration, and a scalable objective function.
• He says RL is now showing the same log-linear scaling behavior that pre-training showed, across a wide variety of RL tasks.

AI summary of what Dario Amodei said

Dario Amodei predicts it is hard for him to see that there won't be trillions of dollars in AI revenue before 2030, describing a plausible path where a 'country of geniuses in a data center' emerges around 2028 and accelerates revenue to trillions. He explains why AI companies currently lose money despite high gross margins: each model is profitable, but exponential compute scale-up spending on the next model exceeds those profits. He argues that with a small number of firms, the industry does not equilibrate to perfect competition with zero margins.

• He predicts trillions of dollars in AI revenue before 2030, with a plausible slower path reaching a 'country of geniuses in a data center' in 2028 and trillions by 2030.
• He explains that each trained model can be highly profitable (citing stylized figures of 75% gross margins), yet companies lose money because they spend far more training the next model during the exponential compute scale-up phase.
• He argues that an industry with a small number of firms behaving rationally does not equilibrate to zero margins, and that the profit equilibrium he describes arrives once model training scale-up levels off.

AI summary of what Dario Amodei said

Dario Amodei discusses the financial risk of buying compute ahead of AI-driven revenue, explaining why Anthropic cannot simply assume 10x annual revenue growth will continue. He argues that even if powerful AI arrives as he expects around 2026 or 2027, the diffusion of its economic value, such as curing diseases, takes time. He contrasts Anthropic's thoughtful approach to data center spending with what he sees as riskier behavior by other companies.

• He says that even if AI technology progresses as fast as he expects, the timing of revenue growth is uncertain, and being off by a couple of years when buying data centers can be ruinous.
• He says that even a 'country of genius' AI arriving in 2026 or 2027 would take time to deliver economic value, using the example of how long it takes to develop, manufacture, and distribute cures for diseases.
• He says Anthropic cannot buy compute as if 10x annual revenue growth were guaranteed, because if revenue falls short even to 800 billion dollars, no hedge could prevent bankruptcy.

AI summary of what Dario Amodei said

Dario Amodei says that long context windows and related capabilities within the existing paradigm may be enough to achieve powerful AI in a data center and generate trillions of dollars of revenue. He says continual learning is being worked on and may be solved within a year or two, though he believes most of the way can be gotten without it. He argues that extending context length is an engineering and inference problem rather than a research problem.

• He says a million tokens of context can represent days of human learning, and that this plus related capabilities may already be enough to generate trillions of dollars of revenue.
• He says Anthropic is working on continual learning and thinks there is a good chance it is solved within the next year or two, though he believes you get most of the way there without it.
• He says extending context length is an engineering and inference problem, and suggests degradation at long context may result from training at a short context length while serving at a longer one.

AI summary of what Dario Amodei said

Dario Amodei says the exponential progress of AI has gone roughly as he expected, moving from smart high school student to PhD-level capabilities, and that the most surprising thing is the public's lack of recognition of how close the technology is to the end of that exponential. He predicts a "country of geniuses in a data center" within one to three years with high confidence it arrives within ten, while arguing that economic diffusion will be extremely fast but not instant. He also defends Anthropic's compute spending as carefully balanced against bankruptcy risk, discusses AI governance, export controls on chips to China, and how he communicates with his company.

• He says he has held the same scaling hypothesis since 2017, which he called the "Big Blob of Compute Hypothesis", and that RL scaling now shows the same log-linear gains that pre-training scaling did.
• He predicts a "country of geniuses in a data center" with 90% confidence within ten years, and as a roughly 50/50 hunch within one to three years, with coding reaching end-to-end capability in one or two years.
• He explains that Anthropic cannot buy compute as if revenue will keep growing 10x a year, because being off by even a year in demand prediction could bankrupt the company.

AI summary of what Dario Amodei said

Dario Amodei, guest on the show, lays out the optimistic vision from his essay "Machines of Loving Grace": AI acting as a country of geniuses could cure cancer and other diseases and dramatically boost GDP growth. He then discusses the disruption of entry-level white collar and software jobs, the geopolitical race with China, Anthropic's AI constitution, and the open question of whether AI models could be conscious. He says he is not a doomer but worries adaptation may be too slow.

• He argues AI should be seen as doing the whole job of the biologist, and that a 'country of geniuses' — a hundred million AI minds at peak human level — could cure cancer, Alzheimer's and other diseases rather than requiring a godlike superintelligence.
• He predicts rapid disruption of entry-level white collar work and software engineering, saying the technology is moving far faster than past economic transitions and may overwhelm society's normal adaptive mechanisms.
• He describes Anthropic's 75-page AI constitution used in training, the precautionary measures around possible model consciousness, and his view that the US and democracies should stay ahead in AI to shape global liberty.