The Myth That Superintelligence Is Almost Here

Listen to the loudest voices in the industry and you might think a machine smarter than any human is a year or two away. Meta has named its flagship research group Superintelligence Labs. Anthropic released its most capable model, Fable 5, only days after warning that AI was becoming dangerously powerful. The word "superintelligence" now turns up in launches and funding decks as if it described something already sitting on the workbench. It does not, and the gap between the marketing and the evidence is worth a careful look.

The boldest timelines come from the people building the systems. Anthropic's Dario Amodei has suggested AI broadly smarter than humans could arrive as soon as 2026 or 2027. OpenAI's Sam Altman has talked about reaching artificial general intelligence within a few years. Elon Musk has floated superintelligence by 2030. These are striking claims from serious people. They are also forecasts made by those with the most to gain if the rest of us believe them.

What the wider field actually thinks

Step outside the executive suite and the picture shifts. As Gizmodo reported, when AI researchers themselves were surveyed, a clear majority doubted that simply scaling up today's methods would ever produce general intelligence. Aggregated forecasts tell a more measured story too. By early 2026, the combined view of expert predictions and prediction markets put the odds of AGI by 2029 at around 25 percent, and reached a coin-flip only around 2033.

That is still a dramatic compression. In 2020 the median expert guess for human-level AI sat somewhere near 2060. The point is not that progress has stalled, it plainly has not. The point is that "within a couple of years" is nowhere near the consensus the headlines imply.

Benchmarks are not the same as understanding

Skeptics push the argument further. Researchers including Yann LeCun and Gary Marcus contend that current architectures cannot reach general intelligence at all, no matter how much data and computing power get thrown at them. Their case rests on a distinction that tends to get lost in the excitement: scoring well on a benchmark is not the same as understanding. A model that matches humans on a desktop-task test has learned to handle that test. It has not necessarily learned to reason its way through a genuinely new problem the way a person can.

This is where the myth does real damage. When a system clears a benchmark, the leap to "it is nearly conscious" or "it will soon surpass us at everything" is a leap of faith, not a result. We looked at a close cousin of this confusion in our piece on why today's humanoid robots do not think like people.

Why it matters

None of this argues for complacency. The models are powerful, they are improving quickly, and the safety questions are real, which is exactly why Anthropic's own caution around the release of Fable 5 deserves to be taken seriously. But "superintelligence is almost here" is a prediction wearing the costume of a fact, and even the insiders making it disagree with one another by decades. The honest answer to when, or whether, machines will outthink us across the board is that nobody knows. That uncertainty deserves more respect than either the hype or the panic tends to give it.