Dario Amodei, CEO and founder of Anthropic, just published an online essay that starts out by comparing AI to the Hobbits in the Lord of the Rings. And that may not be the only piece of fiction in it, if the system card from Anthropic’s own Claude Mythos model is to be believed.
“The intersection of AI and our political institutions feels a bit like the Hobbits and Treebeard,” Amodei writes — that is, Treebeard the Ent moves so slow that he can’t even understand the speedy Hobbits. But then Amodei pivots to a controversial assertion — one that, to continue the Lord of the Rings analogy, would mean that Hobbits are moving exponentially faster all the time.
“AI’s scaling laws, which predict an exponential increase in general cognitive capabilities with increasing computing power, now have over a decade of empirical evidence behind them,” Amodei claims. “If these scaling laws continue for only a year or two longer, we are likely to get what I’ve called Powerful AI.”
It’s hardly a one-off reference; Amodei, who is about to cash in on a bonanza Anthropic IPO, uses the word “exponential” six times in the essay. That includes the title, “Policy on the AI exponential.” But is it true?
Let’s leave for one moment the dubious 2020 OpenAI paper Amodei links to prove his assertion, or the abundant evidence from other AI experts that the “cognitive capabilities” of most Large Language Models are not growing that fast, if at all. Amodei’s essay appears to contradict the word of … well, Anthropic itself.
On the system card for the preview of Claude Mythos [PDF], the model that Amodei hypes up in the essay for the cybersecurity concerns it has caused, you’ll find the following statement: “The [intelligence] gains we can identify are confidently attributable to human research, not AI assistance … early claims of large AI-attributable wins have not held up.” [Emphasis ours.]
Mashable Light Speed
You might think it can’t get more definitive than that — unless you read the system card for Anthropic’s other new frontier model, Fable 5. Using a test called the Epoch Capabilities Index, Anthropic researchers specifically set out to see if there was evidence of a feedback loop that would lead to what AI experts have variously called AGI or Digital Superintelligence. And the result couldn’t be clearer.
“We do not observe a sustained, AI-attributable 2× acceleration in the pace of our AI progress,” the Fable system card [PDF] says.
So, where is Amodei getting his exponential information from? We’ve reached out to Anthropic for clarification, but the citation the CEO uses is for a 2020 paper called Scaling Laws for Neural Language models, co-authored by Jared Kaplan (then with OpenAI, now a co-founder of Anthropic). The conclusion of that paper has been called into question by another leading AI researcher, Gary Marcus, for the past four years.
“There are serious holes in the scaling argument,” Marcus wrote in 2022. “Indeed, we may already be running into scaling limits in deep learning, perhaps already approaching a point of diminishing returns.” He cited research on OpenAI’s GPT-3 model, which has “shown that scaling starts to falter on some measures, such as toxicity, truthfulness, reasoning, and common sense.”
Marcus was pilloried by AI true believers at the time, but has since been vindicated — especially since the release of GPT-5, which was not the Superintelligence some of its users hoped for.
Finding evidence for AI exponential growth since then may be harder than simply walking into Mordor.
Additional reporting provided by Timothy Werth
Topics
Artificial Intelligence
Anthropic
