From Joint Warning to Formal Paper: RSI Gets Its First Full Modeling

On September 28, "godfathers of AI" including Geoffrey Hinton and Yoshua Bengio co-signed a public warning that the regulatory window for self-improving AI is closing. One week later, on October 4, the discussion escalated into a harder academic deliverable: Hinton's first paper on recursive self-improvement (RSI), "What if automating AI R&D triggers an intelligence explosion?", whose 22 authors include Turing Award winners Hinton and Yoshua Bengio, reinforcement-learning pioneer Andrew Barto, and OpenAI chief scientist Jakub Pachocki.

The paper asks a single question: if AI starts massively participating in building the next generation of AI, will an intelligence explosion actually arrive?

Lab Data: AI Is Already Inside the R&D Pipeline

The paper cites internal data from Anthropic and other frontier labs: in January 2025, the share of AI-generated approved code was still in the low single digits; by May 2026 it exceeded 80%. In March 2026, Claude could autonomously complete about 1% of internal AI R&D work under high-level human supervision; by August the figure reached 26% — a more-than-20-fold increase in half a year. As of September 2026, OpenAI's internal AI systems routinely finish R&D tasks that used to take human staff days.

Key Numbers at a Glance

· Share of approved AI-written code: low single digits (Jan 2025) → over 80% (May 2026, Anthropic)

· Claude's autonomous share of internal AI R&D: ~1% (Mar 2026) → 26% (Aug 2026)

· Central estimate of returns to research effort r: ~1.2–1.9 across three AI research subfields

· Model result: with full automation, AI progress could speed up 10x in ~1.5 years — a year's progress compressed into ~5 weeks

A "Software-Driven" Explosion: Researchers Become Software

The paper calls this path a software-driven intelligence explosion: unlike chipmaking, validated algorithms, training methods and agent workflows can be redeployed into the next R&D cycle almost immediately. Crucially, AI researchers can be copied. Introducing the notion of an effective R&D workforce, the authors estimate that a leading AI company's existing compute could in theory support millions of top-researcher-equivalent AI workers — versus a few thousand humans at a frontier lab today.

Yet the paper's tone stays measured: current evidence is far from proving an intelligence explosion is underway — but it is already enough to turn the question from science fiction into a real one that demands preparation.

Four Hard Walls: Compute, Data, Experiment Time, Diminishing Returns

The paper also lists the constraints: GPUs cannot be copied out of thin air, and frontier training runs take months; high-quality natural data does not grow with the number of AI researchers; experiment windows are finite; and science itself keeps getting harder. If diminishing returns outrun AI R&D capacity, the explosion loop never takes off.

The takeaway for enterprises: monitoring, auditing and circuit-breaker capabilities for self-improvement loops are shifting from nice-to-have to must-have. NineZenith's Zenith-Safety (Tiandun) product line delivers exactly this kind of engineered AI-pipeline audit and compliance capability for government and enterprise clients.

(Compiled from public reports by QbitAI, Geoffrey Hinton's X account and the paper itself, and other outlets)