AI Is Starting to Build Better AI
This IEEE Spectrum feature examines how far AI has come toward recursive self-improvement (RSI), the long-imagined idea of machines that design better machines. The article traces decades of stepping-stones, from machine learning and AutoML to today's large language models writing the code for their own successors. It argues that while many systems now help build better AI, they still rely on humans to set goals, define success and decide which changes to keep, so the question is how much of the self-improvement loop has actually been closed.
Key Takeaways
- Recursive self-improvement (RSI) is the idea, dating to 1966, that an intelligent machine could design even better machines.
- RSI is best understood as a spectrum, from fully autonomous loops to nearly any use of tech to build tech.
- Today's systems can help build better AI but still rely on humans to set goals, define success and keep changes.
- OpenAI reported that GPT-5.3-Codex was instrumental in creating itself by helping debug training and analyze evaluations.
- Anthropic claims the majority of its code is now written by Claude Code.
- Google DeepMind's AlphaEvolve guides the evolution of solutions but still needs humans to set problems and evaluations.
The long-held idea of self-improving AI
The concept predates modern AI by decades.
- ›In 1966, mathematician I. J. Good wrote that an ultraintelligent machine could design even better machines.
- ›Good predicted an intelligence explosion that would leave human intelligence far behind.
- ›AI researchers have long seen RSI as something both to desire and to fear.
The article notes RSI means different things to different people. Some use it as a bogeyman to scare up regulation, others brandish it in marketing, and definitions range from a fully autonomous loop to nearly any use of tech to build tech. The author concludes it is safest to call it a spectrum.
How much of the loop is closed
The strictest definition sets a high bar that today's systems do not meet.
- ›At its strictest, RSI describes systems that improve not just outputs but the process by which they improve.
- ›Such systems would generate ideas, evaluate results and modify their own methods with zero human direction.
- ›By that standard, many of today's systems fall short.
The article reframes the question: not whether self-improvement exists in some form today, but how much of the loop has actually been closed. Current systems can help build better AI but still depend on humans to set goals, define success and decide which changes to keep.
Decades of stepping-stones
Researchers have built the elements of RSI over many years.
- ›Machine learning algorithms automatically tune parameters of programs that play games or create new programs.
- ›Evolutionary algorithms diversify and iterate on design solutions, including other algorithms.
- ›AutoML has automated parts of the pipeline for structuring, training and evaluating ML models.
The article places today's large language models such as GPT, Gemini, Claude and Grok as an extension of this trend, noting that one of their biggest use cases is writing code, including code to produce future versions of themselves.
LLMs helping build their successors
Recent examples show AI assisting in its own development.
- ›In February, OpenAI reported GPT-5.3-Codex was instrumental in creating itself.
- ›It helped debug training, manage deployment and analyze evaluation results.
- ›Anthropic claims the majority of its code is now written by Claude Code.
The article stresses these systems still rely on humans to direct and verify the work, so they assist rather than autonomously improve themselves.
AlphaEvolve, chip design and supervision
Google DeepMind projects show collaboration between humans and AI.
- ›AlphaEvolve is a coding agent for scientific and algorithmic discovery that uses LLMs to guide the evolution of solutions.
- ›It has optimized neural-network architectures, data-center scheduling and chip design.
- ›It is not a fully recursive loop, since people still set the problems and define how to evaluate performance.
DeepMind's Matej Balog called it a collaborative process where researchers learn from what the system discovers, and said there are first demonstrations that using AI to find algorithms beyond human intuition is not a wild dream. Separately, the co-leads of DeepMind's AlphaChip launched a startup called Ricursive Intelligence to use AI to design AI chips, aiming to cut the design cycle from one or two years to days, with cofounder Anna Goldie noting the recursive phase would still happen under human supervision.
Frequently Asked Questions
What is recursive self-improvement?
It is the idea, dating to I. J. Good in 1966, that an intelligent machine could design even better machines, potentially leading to an intelligence explosion. The article treats it as a spectrum rather than a single definition.
Are today's AI systems truly self-improving?
Not by the strictest definition. They can help build better AI but still rely on humans to set goals, define success and decide which changes to keep, so much of the loop remains open.
How have AI models helped build their own successors?
OpenAI reported that GPT-5.3-Codex was instrumental in creating itself by helping debug training and analyze evaluations, and Anthropic claims most of its code is now written by Claude Code.
What is AlphaEvolve?
It is a Google DeepMind coding agent for scientific and algorithmic discovery that uses LLMs to guide the evolution of solutions, but it still needs humans to set problems and evaluation criteria.
What is Ricursive Intelligence trying to do?
Founded by the co-leads of DeepMind's AlphaChip, it aims to use AI to design AI chips and cut the design cycle from one or two years to days, with the recursive phase still under human supervision.
The article concludes that AI is genuinely starting to help build better AI, but the self-improvement loop remains open because humans still set the goals, define success and approve the changes.
Continue Learning
Comments
Sign in to join the conversation