Amodei and colleagues found regular scaling patterns and predicted that raising investment from about $10,000 to hundreds of millions of dollars would bring huge change and economic value.

The claim has two parts: Amodei and colleagues (a) found regular scaling patterns and (b) at the time predicted that scaling investment from about $10,000 to the hundreds of millions of dollars would produce dramatic change and large economic value. TRUE: dated records (papers, internal or public writings, or contemporaneous accounts) show his group documented regular scaling patterns while experiments cost around $10,000, and at that time anticipated that scaling to hundreds of millions of dollars would bring transformative change or economic value. MOSTLY TRUE: both parts hold, but the dollar figures are off by a modest amount (for example the starting cost was around $100,000), or the prediction was framed somewhat less explicitly. MIXED: (a) is documented but (b) is not shown to have been made at the time, so it appears only in later retrospective accounts. MOSTLY FALSE: scaling patterns were found, but contemporaneous evidence shows no such prediction, or the investment scale was very different. FALSE: there is no evidence that his group found regular scaling patterns.

Set before the research began.

1. E1 Scaling Laws for Neural Language Models arXiv (Cornell University preprint server) · 23 January 2020

   “We study empirical scaling laws for language model performance on the cross-entropy loss. The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude.”

2. E2 CEO Speaker Series With Dario Amodei of Anthropic | Council on Foreign Relations Council on Foreign Relations · 10 March 2025 Own record

   “And we were measuring these trends back when models cost $1,000 or $10,000 to train. So that’s a kind of an academic grant budget level. And we forecast that these trends would continue, even when models cost 100 million, a billion, $10 billion to train, which now we’re getting to.”

3. E3 The Man Who Counted the Future bretkerr.substack.com (Substack) · 15 April 2026

   “In the spring of 2017, in a nondescript office that could have been mistaken for a graduate school common room, a physicist with a very specific obsession sat down to write a memo that would, in retrospect, read like a detonation timer.”

4. E4 The Man Who Counted the Future bretkerr.substack.com (Substack) · 15 April 2026

   “He argued that OpenAI would need to secure a hundred million dollars of compute to reach the next fundamental level.”

E1 is a dated 2020 preprint that documents regular power-law scaling trends in language models, but the quoted page does not name Amodei or his group. It also gives no experiment costs and makes no prediction about investment. E2 is Amodei's own retrospective remark from 2025, saying the trends were measured when models cost $1,000–$10,000 and were forecast to continue at $100 million and beyond; it is his ordinary statement, not a contemporaneous record. E4 and E5 are a 2026 Substack account of a 2017 memo that argued for $100 million in compute, but the quoted text does not name the author. With no evidence independently tying the scaling findings and the early prediction to Amodei, the evidence does not settle either part of the claim.