AGI is coming soon, wake up
// essay
This is what I sent out to peers at Harvard via email:
---
In June of 2024, podcaster Dwarkesh Patel released an episode with Leopold Aschenbrenner. Listening to this now in March of 2026, we find Leopold to be quite prescient. Despite many at the time calling Aschenbrenner a lunatic, he was markedly conservative in his views.
We hit 1GW in January 2026. Prometheus and clusters in Abilene will come online later this year. It is increasingly likely that we will actually get a $500B, 10GW buildout in the coming years. To add to this, the US security apparatus and federal government have begun making moves earlier than many projected. Claude for war. Buying 10% of Intel. Project Jupiter.
There are major implications here. When we get superintelligence, will states use it to install an authoritarian regime? Will there be a World War fought over energy supply? Is all white collar work cooked — and how do we think about a post-abundance society?
There is no time to waste. These questions need to be discussed now so that we can plan and shape the future. The majority of Harvard students still believe AI is a “bubble” and that large language models are stochastic parrots. Professors at Harvard regularly tout these claims.
Every institution that mattered in the last great technological transition had a small group who saw it coming and acted. Harvard produced many of them. Right now, most of Harvard is asleep.
Scaling laws have held for 6+ orders of magnitude. Critics have called the ceiling at every log-step and been wrong every time. The null hypothesis should be that scaling continues, not that it stops — and yet the default Harvard prior is the opposite. When will our community learn The Bitter Lesson?
But those of us who used GPT-2 on Huggingface in 2020 know the truth. We were using text-davinci-003 in 2022 to write essays, and so when ChatGPT was released, it was not a big deal. But, it was a big deal when o3 solved ARC-AGI. We feel and hear the AGI — whether in Opus 4.6 or from our friends at the labs.
Verifiable domains are cooked. And solving these verifiable domains will lead to a system capable of architecting Agent-1. The army of 1,000,000 automated AI research engineers is coming.
While EA circles provide a necessary ethical framework, the urgency of 2026 demands a pragmatic, high-stakes coalition. We are no longer debating if the curve remains exponential; we are literally in the midst of takeoff.
The real questions are no longer technical. They’re about power, war, labor, and what human purpose looks like on the other side.
We are looking for the realists who recognize that the next 24 months will reorder the global hierarchy.
So, if you already believe in the singularity, come find the others.
https://singularityharvard.com