Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Dario Amodei says he wrote a document in 2017 called The Big Blob of Compute Hypothesis, arguing that scale, not clever techniques, drives AI progress.
“in 2017, I think I talked about it last time, but I wrote a doc called The Big Blob of Compute Hypothesis.”
0:01:58 · Dario Amodei · Scaling hypothesis · 2017
Amodei says his hypothesis lists seven factors that matter, including raw compute, data quantity, data quality and distribution, training duration, a scalable objective function, and numerical stability.
“I think I listed seven of them. One is like how much raw compute you have. The other is the quantity of data that you have.”
0:02:54 · Dario Amodei · Scaling factors · 2017
Amodei recalls that GPT-1 was trained on a narrow body of literary text, which generalized poorly, whereas GPT-2 used a broad internet scrape.
“GPT-1 itself was trained on a bunch of, I think it was fan fiction actually.”
0:06:47 · Dario Amodei · GPT-1 training · 2017
Amodei says Anthropic's revenue grew from roughly zero to $100 million in 2023, to about $1 billion in 2024, and to about $9 or 10 billion in 2025.
“you know, in 2023, it was like zero to 100 million. 2024, it was 100 million to a billion. 2025, it was a billion to like nine or 10 billion.”
0:21:34 · Dario Amodei · Anthropic revenue · 2023-2025
Amodei says his earlier prediction that AI would write 90% of code within three to six months came true at some places, including Anthropic.
“the AI model will be writing 90 % of the lines of code in like, you know, three to six months, which happened at least at some places, right?”
0:17:49 · Dario Amodei · AI coding · 2025
Amodei says performance on the OSWorld computer-use benchmark rose from about 15% at the first release of computer use to roughly 65 to 70%.
“OS World went from like 5% – like I think when we first released computer use like a year and a quarter ago, it was like maybe 15%. I don't remember exactly. But we've climbed from that to like 65 or 70%.”
0:32:15 · Dario Amodei · Computer use · 2025
Dwarkesh Patel cites a METR study in which experienced developers felt more productive with AI tools but merged about 20% less work.
“There was a meter study I'm sure you saw last year where they had experienced developers try to close pull requests in repositories that they were familiar with.”
0:35:07 · Dwarkesh Patel · Developer productivity · 2025
Amodei estimates coding models currently give about a 15 to 20% total-factor speedup to AI development, up from about 5% six months earlier.
“I would say right now the coding models give maybe, I don't know, a like 15, maybe 20% total factor speed up. Like that's my view.”
0:37:36 · Dario Amodei · AI productivity · 2026
Claude Code began as an internal tool called Claude CLI, and Amodei decided to launch it externally after seeing fast adoption within Anthropic.
“It had been originally called Claude CLI, and then the name eventually got changed to Claude Code internally”
1:28:25 · Dario Amodei · Claude Code origin · 2025
Amodei says longer context windows are mainly an engineering and inference problem, because serving them requires storing a large key-value (KV) cache in memory.
“If you want to serve long context, you have to store your entire KV cache.”
0:43:54 · Dario Amodei · Long context
Amodei guesses that AI able to learn on the job like a six-month human editor could arrive in one to three years.
“I mean, you know, my guess for that is, you know, there's a lot of problems that are basically like we can do this when we have the country of geniuses in a data center.”
0:45:31 · Dario Amodei · AI timelines · 2026-2029
Amodei says his essay Machines of Loving Grace suggested powerful AI could arrive around 2026 or 2027, with an error margin of a year or two.
“I said, well, get that 2026, maybe 2027 again. That is my hunch. Wouldn't be surprised if I'm off by a year or two”
0:48:51 · Dario Amodei · AI timelines · 2026-2027
Amodei says buying about a trillion dollars a year of compute starting in late 2027 would bankrupt Anthropic if revenue fell even modestly short of that.
“I could buy $1 1 trillion of compute that starts at the end of 2027. And if my revenue is not $1 1 trillion, if it's even 800 billion, there's no force on earth.”
0:51:37 · Dario Amodei · Compute purchasing · 2027
Amodei estimates the industry is building roughly 10 to 15 gigawatts of compute this year, tripling each year, with a gigawatt costing about $10 to 15 billion a year.
“the amount of compute the industry is building this year is probably in the, you know, I don't know, very low tens of, you know, call it 10, 15 gigawatts next year.”
0:56:24 · Dario Amodei · Industry compute · 2026
Dwarkesh Patel says Anthropic has told investors it plans to be profitable starting in 2028, which Amodei says is a hard-to-predict estimate that could shift to 2026 if revenue grows fast enough.
“You've told investors that you plan to be profitable starting in 28.”
0:58:39 · Dwarkesh Patel · Profitability · 2028
Amodei gives a stylized example in which a $1 billion model earns $4 billion of revenue at 75% gross margin, while the company still loses money by spending $10 billion on the next model.
“Let's say a model got trained that costs $1 billion last year. And then this year it produced $4 billion of revenue and cost $1 billion to inference from.”
1:09:31 · Dario Amodei · AI economics
Amodei argues a few-firm AI industry would resemble a Cournot equilibrium, in which margins stay positive rather than falling to zero, similar to cloud computing with three or four big players.
“The Cournot equilibrium I think is what the small number of firm equilibrium is.”
1:08:46 · Dario Amodei · Industry structure
Amodei says he expects AI to produce roughly 10 to 20% annual economic growth, as he wrote in Machines of Loving Grace, rather than 300%.
“I think we may get 10 or 20 % per year growth in the economy, but we're not going to get 300 % growth in the economy.”
1:11:43 · Dario Amodei · Economic growth
Amodei says the FDA approval process should be sped up because AI-accelerated drug discovery could jam the pipeline, and he calls chatbot-restriction bills stupid.
“I do think we should work to speed the approval process in the FDA.”
1:46:09 · Dario Amodei · Drug regulation
Dwarkesh Patel says a Tennessee bill introduced on December 26th would make it an offense to knowingly train AI to provide emotional support.
“on December 26th, the Tennessee legislature introduced a bill which said, quote, it would be an offense for a person to knowingly train artificial intelligence to provide emotional support”
1:36:20 · Dwarkesh Patel · State AI laws · 2025
Amodei says Anthropic opposed a proposed ten-year federal moratorium on state AI laws because there was no federal plan to replace them.
“we're going to ban all state regulation of AI for 10 years with no apparent plan to do any federal regulation of AI”
1:37:47 · Dario Amodei · AI regulation
Amodei says Anthropic advocates export controls on advanced chips to China, which he says are in US national security interests but have not held because of the money at stake.
“a thing I've been trying to fight for is export controls on chips to China, right?”
1:45:09 · Dario Amodei · Chip export controls
Amodei says democracies should hold a strong hand when post-AI rules of the road are negotiated, but that no one country or coalition should unilaterally set them.
“I don't think a coalition of democracies or certainly one country should just say, these are the rules of the road, there's going to be some negotiation, right?”
1:49:53 · Dario Amodei · AI geopolitics