Researchers Say AI Tools Are Already Helping Build Better AI

Modern AI research is producing clear examples of machines helping to design and improve other machine‑learning systems, reviving an idea first framed by I. J. Good in 1966: that an "ultraintelligent machine could design even better machines." Researchers now describe recursive self‑improvement (RSI) as a spectrum, from narrow automation of specific tasks to fully autonomous self‑modification.

Many building blocks of that spectrum are already in routine use. Over decades, machine learning has automated parameter tuning and design search; evolutionary algorithms iterate on candidate solutions; and AutoML systems automate model structure, training and evaluation. More recently, large language models (LLMs) such as GPT, Gemini, Claude and Grok are being used to write code that contributes to future model versions, including code used for training, debugging and evaluation.

Some companies report concrete rollouts where models materially assist their own development. OpenAI said in February that GPT‑5.3‑Codex played an instrumental role in creating itself by helping debug training, manage deployment and analyze evaluation results. Anthropic has said the majority of its code is now produced by its Claude Code system. In both cases these capabilities operate under human direction and oversight rather than as unconstrained, closed loops.

Research prototypes illustrate how such collaboration can accelerate design. Google DeepMind’s AlphaEvolve uses LLMs to guide evolutionary search for solutions ranging from neural‑network architectures to data‑center scheduling and chip design. DeepMind researcher Matej Balog described the work as "a very collaborative process," noting that teams learn from discoveries made by the system. AlphaEvolve is not a closed loop: people still select problems and set evaluation criteria, but each success can speed subsequent experiments.

Spin‑outs and commercial projects are pursuing similar aims in hardware design. The co‑leads of DeepMind’s AlphaChip launched Ricursive Intelligence to apply AI to chip design; cofounder Azalia Mirhoseini said the team expects to "dramatically reduce the design cycle from one or two years to days." Ricursive’s public roadmap, according to cofounder Anna Goldie, moves from tooling human designers toward automating design for firms without in‑house teams and ultimately to overseen, recursive AI‑assisted chip improvement.

Academic groups are exploring agent‑level self‑improvement as well. Researchers at the University of British Columbia and Sakana AI announced Darwin Gödel Machines, which apply evolutionary algorithms to improve LLM‑based coding agents. These efforts advance important components of RSI while still relying on human goal‑setting, verification and evaluation. For builders the practical lesson is straightforward: tooling that writes, test‑runs and proposes changes to models is maturing and can shorten iteration cycles for architectures, compilers and chips. At the same time, that progress creates concrete needs — more robust evaluation pipelines, stronger human‑in‑the‑loop verification and tighter deployment controls — to manage risks and ensure reliability in production.