Beyond Superintelligence: What Is AI’s Ultimate Destination?

Four former Google researchers — Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals — left Google in August 2026 to launch Discovery Loop, a company betting that the fastest route to superintelligence is a closed loop where AI improves the machinery that produces AI. The report cites Sakana AI's Darwin Gödel Machine improving SWE-bench performance from 20% to 50%, Google DeepMind's AlphaEvolve designing algorithms for Gemini, and Anthropic reporting engineers shipping roughly 8× more code per quarter than in 2021–2025.

For most of the AI era, we have treated AGI as the finish line. A new class of company is betting it is a roadside marker. In August 2026, four of the most consequential people in computing — Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals — left Google to launch Discovery Loop https://radical.vc/our-investment-in-discovery-loop , a company organized around a single wager: that the fastest route to superintelligence is not a smarter model, but a closed loop in which AI improves the machinery that produces AI. Recursive intelligence is not another rung on the same ladder as AGI and superintelligence. It is a process — a system that helps improve the machinery that produces intelligence, uses those improvements to create a more capable successor, and applies that greater capability to the next round. If the loop closes, AGI flashes past as a marker on the road, and superintelligence may be the last threshold humans can recognize before intelligence moves beyond any scale based on us. The idea is old. I. J. Good described the intelligence explosion in 1965. Vernor Vinge named the singularity in 1993. Jürgen Schmidhuber formalized the Gödel Machine https://arxiv.org/abs/cs/0309048 in 2003. What is new is that the loop is being built. Sakana AI’s Darwin Gödel Machine https://sakana.ai/dgm moved SWE-bench performance from 20% to 50% by having agents modify their own code. Google DeepMind’s AlphaEvolve https://deepmind.google/blog/alphaevolve is now designing algorithms that ship back into Gemini. And Anthropic reports https://anthropic.com/institute/recursive-self-improvement its engineers now ship roughly 8× the code per quarter they shipped in 2021–2025, with the model itself doing most of the writing. This report traces recursion from Turing’s child machine to today’s self-modifying agents, separates genuine recursive self-improvement from the marketing that now surrounds the word, maps six possible destinations beyond superintelligence — from an S-curve stall to a runaway ascent — and sets out the observable gates by which the rest of us can tell the difference. It also names what could slow the loop: compute limits, alignment failure, regulation, or the simple possibility that the loop degrades instead of compounding. AGI is a milestone. Superintelligence is a threshold. Recursive intelligence is a trajectory. The question is no longer simply whether humanity can build intelligence greater than its own — it is whether we can give a direction and a reason to an intelligence that may never stop becoming something more.