I think reading https://red.anthropic.com/2026/mythos-preview/, along with some reasonable assumptions on what else Mythos is being used for, and it’s clear that we may as well date the Singularity to April 2026. That is, Anthropic is there or there-abouts on exponential self-improvement of AI coding models.

I guess of anyone (that we know of at least!) to be in the lead, Anthropic is probably best case amongst the pool of single-actor, private-sector players. Which is not, to be clear, the best of all possible cases, but it could be worse.

Now we see just how accurate the sci-fi models of the Singularity are, versus my view that real society-wide change has a built in speed limit, or equivalently that, “The future is already here, it’s just not very evenly distributed,” as per William Gibson.

Wild times. Hard to focus on every day activities.

---

EDIT: let me be a little clearer what I actually think. By “Singularity” I guess I really mean “the start of the process of exponential self-improvement that, because it’s exponential, has a singularity (infinity) of rate-of-improvement in the future.” That infinity is not yet, but when talking about Singularity (capital S) I guess the start of exponential self improvement is relevant.

I’m a little skeptical of the direct Singularity claim, though, because I think:

1. We’ve been in the space of exponential improvement for quite some time, see e.g., Moore’s law. The reason now feels special is because we’re passing an especially interesting point in that trajectory: human-level capabilities.
2. But, it’s only human level capabilities in some areas where LLMs are really strong (text-driven logic, such as computer programming); we’re quite a way off still in other areas (e.g., physical control of bodies/robots, pursuit of long-running motivations). Though exponentials do compress time periods.
3. Most importantly, the impact this has on the “real” world depends on many things. If a process has a few steps, then when one step tends to zero time required, the other steps are more relevant. Things like speed of hardware rollout, power supply, regulatory requirements, social inertia are still just as relevant as always, so I don’t think that everything will suddenly change at once.