Anthropic’s CEO says the growth of AI is accelerating. Anthropic research says otherwise.

Dario Amodei, CEO and founder of Anthropic, recently published an online article that begins by comparing AI to Hobbits in The Lord of the Rings. And that may not be the only piece of mythology in it, if the system card from Anthropic’s Claude Mythos model is to be believed.

“The intersection of AI and our political institutions feels like Hobbits and Treebeard,” writes Amodei – that is, Treebeard the Ent is so slow that he can’t even understand the fast Hobbits. But then Amodei insists on a controversial claim – that is, continuity The Lord of the Rings For example, it would mean that the Hobbits are leaving very fast all the time.

“The scaling laws of AI, which predict a massive increase in general cognitive abilities with increasing computing power, now have more than a decade of empirical evidence behind them,” Amodei said. “If these scaling laws continue for a year or two only, we may get what I called for Powerful AI.”

It is never a single reference; Amodei, who is about to cash in on the bonanza Anthropic IPO, uses the word “exponential” six times in the story. That includes the title, “Policy on the AI ​​exponential.” But is it true?

Let’s leave for a moment the dubious OpenAI paper of Amodei of 2020 that links to prove his assertion, or the many evidences from other AI experts that the “cognitive abilities” of many Big Language models are not growing that fast, if at all. Amodei’s story seems to defy the word … well, Anthropic itself.

On the Claude Mythos preview system card [PDF]model that Amodei recommends in the case of cybersecurity concerns that he created, you will find the following statement: ” [intelligence] benefits we can identify confidently created by human research, not AI help… Early claims of AI-induced big wins are still unproven.” [Emphasis ours.]

You might think you won’t get any clearer than that – unless you read the system card of another of Anthropic’s new frontier models, Fable 5. Using a test called the Epoch Capabilities Index, Anthropic researchers are particularly keen to see if there is evidence of a feedback loop that could lead to what AI experts have variously called AGI or Digital Superintelligence. And the result couldn’t be clearer.

“We don’t see a sustained, AI-attributable 2× acceleration in the pace of our AI progress,” Fable’s system card [PDF] he says.

So, where does Amodei get his descriptive knowledge? We reached out to Anthropic for clarification, but the citation the CEO is using is from a 2020 paper called Scaling Rules for Neural Language Models, co-authored by Jared Kaplan (formerly of OpenAI, now an Anthropic co-founder). The conclusion of that paper was questioned by another leading AI researcher, Gary Marcus, four years ago.

“There are big holes in the measurement argument,” Marcus wrote in 2022. “Sure, we might already run into the limits of scaling in deep learning, perhaps approaching the point of diminishing returns.” He cited OpenAI’s GPT-3 model study, “which showed that scaling begins to falter on certain measures, such as toxicity, realism, reasoning, and general understanding.”

Marcus was named by the true believers of AI at the time, but has since been confirmed – especially since the release of GPT-5, which was not the Superintelligence that some of its users expected.

Finding evidence of AI growth from there may be more difficult than simply entering Mordor.

Additional reporting provided by Timothy Werth