SAI: adaptation speed instead of AGI
On 27 February 2026 Goldfeder, Wyder, LeCun and Shwartz-Ziv posted a position paper on arXiv: “general” AI is a flawed goal because human intelligence is specialised too; in place of AGI they propose “Superhuman Adaptable Intelligence” (SAI), measured by the speed of acquiring new tasks.
Why it matters
One of the field’s best-known figures publicly asks to replace “AGI” with another goal and another yardstick. An editorial assessment: this is an argument, not a result; nobody has turned the paper’s yardstick into a test, and the authors’ preference for self-supervised learning and world models is their own research line.
The paper states five positions: human intelligence is not general; generality is not needed to be useful; there is no agreed definition of AGI; existing definitions are insufficient; the focus should be SAI. It lays existing definitions (Hendrycks et al., Morris et al., Hassabis, Chollet, Legg and Hutter, OpenAI) on two axes and judges them by three requirements, feasibility, internal consistency and assessability, and shows in a table that each fails at least one. SAI is defined as the ability to adapt to whatever humans can do and to useful tasks outside the human range; the metric is the speed with which an agent acquires new skills. As a route the authors name self-supervised learning, prediction in latent space (JEPA, Dreamer 4, Genie 2) and diversity of architectures against the homogeneity of autoregressive models; one claim is that such models’ errors diverge exponentially with prediction length. The paper answers objections it attributes to Elon Musk and Demis Hassabis: that the brain is “general” in the Turing-machine sense. What the record does not state. The paper has no experiment, dataset or measurement: the argument rests on citations and reasoning. The authors themselves write that they have not yet precisely defined “utility”, and they give no concrete test of adaptation speed. That humans are not “general” is the paper’s thesis, not an established fact. The date is the submission of v1; the PDF says March 2. There are no later versions.