Superintelligence

Superintelligence is an artificial intelligence that outperforms humans in all areas. It is sometimes used interchangeably with AGI (artificial general intelligence), and Anthropic calls this “powerful AI.”¹

It is not really clear to me how this will be quantified and strikes me as something like the Turing test—easy to define, hard to concretely assess.

Dario Amodei wrote an essay called Machines of Loving Grace which has a comprehensive and hyperbolic description of what a world with superintelligence could be like. From my own experience and observations, small as they are, I tend to agree with most of what he said.

Architecture

In practice, will not be a singular LLM but instead a reasoning model or a collection of interconnected models that applies a bunch of techniques and chews on a problem for a period of time.¹

Dario Amodei predicted that it would be millions of times cheaper to inference than train, allowing the datacenter that trained such a model to be repurposed to act as a “country of geniuses.”¹

He then claimed in January 2025 that “Making AI that is smarter than almost all humans at almost all things will require millions of chips, tens of billions of dollars (at least), and is most likely to happen in 2026-2027.”²

When?

In May 2023, OpenAI said that superintelligence might be achievable within ten years.³

In September 2024, Sam Altman’s estimate shortened to “a few thousand days,” or the shorter end of a decade,⁴ using current computer hardware.⁵ In November 2024, Sam Altman clarified further:

In a recent interview with Y Combinator CEO Garry Tan, Altman said, “We basically know what to go do” to achieve artificial general intelligence—technology that is on par with human abilities—and part of it involves “using current models in creative ways.”

The same month, Dario Amodei stated he expects superintelligence “as early as 2026”¹ and 2026-2027⁶ which is effectively the same thing. He repeated this timeframe in January 2025.²

However, Demis Hassabis predicted AGI would “start to emerge” between 2030 and 2035,⁷ while the CEO of Baidu predicted that AGI is “more than 10 years away.”⁸

Footnotes

1. Machines of Loving Grace ↩ ↩² ↩³ ↩⁴

2. Dario Amodei — On DeepSeek and Export Controls (darioamodei.com) ↩ ↩²

3. Governance of superintelligence ↩

4. The Intelligence Age ↩

5. “we believe it is achievable with current hardware” (reddit.com) ↩

6. This Week in AI: Anthropic’s CEO talks scaling up AI and Google predicts floods | TechCrunch ↩

7. AI that can match humans at any task will be here in five to 10 years, Google DeepMind CEO says (cnbc.com) ↩

8. Elon Musk predicts smarter-than-humans AI in 2 years. The CEO of China’s Baidu says it’s 10 years away (cnbc.com) ↩