When AI Starts Evolving Itself: Recursive Self‑Improvement Is Emerging Far Faster Than the Singularity

In 1966 I. J. Good warned that a super‑intelligent machine could design ever‑better machines, leading to an “intelligence explosion” that would outpace human intelligence.

Definition of Recursive Self‑Improvement (RSI)

Strictly, RSI requires a system that can improve not only its output but also the process by which it improves: it must generate ideas, evaluate results, modify methods, and do so without any human input. By this definition most current AI systems fall short; they assist in building better AI but still rely on humans to set goals, define success criteria, and select which changes to retain.

Technical Foundations Leading Toward RSI

Early machine‑learning auto‑tuning, evolutionary algorithms that iteratively generate and select designs, and the past decade’s AutoML frameworks that automate parts of neural‑network architecture search, training, and evaluation constitute the technical groundwork for RSI.

Large Language Models Extending the Trend

Large language models such as OpenAI’s GPT, Google DeepMind’s Gemini, Anthropic’s Claude, and xAI’s Grok have made code generation a primary use case, including code that builds the next generation of models.

OpenAI GPT‑5.3‑Codex

In February 2024 OpenAI reported that GPT‑5.3‑Codex played a significant role in its own development by debugging training runs, managing deployments, and analyzing evaluation results.

Anthropic Claude Code

Anthropic states that most of Claude’s code is now written by Claude Code, though human direction and verification remain required.

AlphaEvolve: An Advanced Coding Agent

DeepMind announced AlphaEvolve (2025), described as an “intelligent coding agent for scientific and algorithmic discovery.” It uses a large language model to guide evolutionary search for neural‑network structures, data‑center scheduling, and chip design. Human researchers still define problems and evaluate outcomes, but each algorithmic breakthrough feeds back to enhance AI‑driven R&D.

Matej Balog, a computer‑science researcher on AlphaEvolve, called the process “highly collaborative” and noted that researchers often gain new insights from AI‑generated solutions.

AI‑Assisted Chip‑Design Roadmap

AI assists humans in chip design.

AI autonomously completes chip development without a specialized team.

AI designs superior AI chips, which are then used to train stronger AI models.

DeepMind alumni founded Ricursive Intelligence to compress the traditional one‑to‑two‑year chip‑design cycle into “days.”

Darwin Gödel Machines (DGMs)

In 2023 the University of British Columbia and Sakana AI released DGMs, which employ evolutionary algorithms to continuously improve LLM‑based code agents. While they cannot yet modify the underlying language model, they increasingly improve themselves and begin to tweak their own improvement mechanisms.

AI Scientist Project

DeepMind’s later AI Scientist system aims to automate the entire research loop—coding, experimentation, evaluation, and knowledge generation—moving beyond code automation toward full scientific discovery.

Remaining Obstacles

Cutting‑edge AI development now costs tens of billions of dollars; no organization is willing to hand over such expensive systems to fully autonomous AI.

Even if AI can design better software, it cannot instantly take over the complex physical infrastructure required for production (data‑center operation, power generation, mining, robotic manufacturing). These capabilities remain deeply dependent on existing human‑built industrial foundations.

Nathan Lambert argues that increasing system complexity introduces friction and coordination costs, resulting in “lossy self‑improvement” that slows the improvement flywheel.

Future Scenarios

Some researchers contend that current AI is only “fairly good” at generating ideas, writing code, and evaluating results, far from fully autonomous. They envision a Cambrian‑like explosion of diverse AI agents, each forming its own ecology, culture, and economy, rather than a single monolithic super‑intelligence.

In such a trajectory, human researchers may first retreat from low‑level debugging to roles akin to professors or team leads, then to project supervisors or CEOs, and eventually to overseers of AI‑driven research.

Illustrations

Code example

来源：ScienceAI
本文
约2200字
，建议阅读
5
分钟
人类正第一次接近“只能闭环”。