Musk and Altman agree we are at the singularity

Elon Musk and Sam Altman rarely agree publicly. In recent weeks they did: Musk tweeted that artificial intelligence is already superhuman in many domains and declared “we are at the singularity.” The tweet was a response to similar remarks by Altman, and unlike their usual barbs, no rebuttal followed.

They are not alone. Demis Hassabis of Google DeepMind and Jensen Huang of Nvidia have begun invoking the term in interviews and conferences over the past months. Each comes from a different angle—research, hardware infrastructure, product—but the emerging consensus is unusual in a landscape where “singularity” was until recently treated as a slur or a distant speculation.

The more interesting reason lies behind the scenes. In the labs where the leading models are trained, the picture differs markedly from what the public encounters. Researchers work with early-stage versions, before the alignment step and before the RLHF layers that curb output. In those settings the models display raw, sometimes unpredictable capabilities that have not passed the “limiting switch” intended to make them safe and useful for end users.

Outside, the models released to the public have undergone rigorous fine-tuning, often to the point of stripping certain abilities. Inside, the unfiltered, more powerful, stranger—and, according to researchers, more unsettling—versions are observed. This gap explains why the engineers building the systems talk about singularity while the public sees a chatbot that refuses to answer sensitive questions.

The agreement between Musk and Altman does not guarantee the claim’s correctness, but it is an indicator. Those who live with the unaligned versions experience a reality different from users of the final product. The question is not whether we are at the singularity—a vague notion anyway—but when the gap between what models can do and what they are allowed to do will close, and who will decide what that looks like.