AI is already involved in the work of developing AI. It can write code, suggest experiments and help improve algorithms used in computing infrastructure.
That is a substantial development without assuming it has reached the much stronger idea of a system independently building a better successor, then repeating the process.
The distinction makes this story easier to follow—and more interesting than a choice between “nothing new” and “machines are improving themselves without limit”.
There are several levels of “building itself”
| Level | What happens | What that evidence would establish |
|---|---|---|
| Assistance | A person uses AI to write code or analyse a result | Help with part of the development work |
| Automated experimentation | A system proposes and evaluates candidates within a defined setup | Progress against the setup’s measured objective |
| Autonomous successor development | A system designs, builds and validates a more capable successor with little or no human direction | A much broader development capability requiring stronger evidence |
The boundaries are not perfectly sharp. They still prevent a successful coding task from being presented as proof of an entire autonomous research process.
A concrete example: AlphaEvolve
Google DeepMind’s May 2025 AlphaEvolve announcement describes language models proposing programs and automated evaluators scoring the candidates.
The company reports applications to algorithms used in its computing infrastructure and AI training. The important mechanism is a loop: generate a candidate, evaluate it against defined criteria, and use the results to guide further candidates.
This is meaningful evidence of AI contributing to improvements in the systems supporting AI. It is also a bounded process with objectives and evaluation machinery. It does not, by itself, establish an independently self-improving general intelligence.
As with any developer’s report, read the method and conditions alongside the claimed outcomes.
What Anthropic says about its own development
In its essay on recursive self-improvement, Anthropic describes delegating a growing share of development work to AI. It also explicitly says the fully autonomous successor-building process it discusses has not yet been reached.
That makes the essay useful primary evidence of the laboratory’s position. Internal measurements and forecasts remain the company’s reported evidence, rather than independent confirmation of every broader claim.
Keep observed changes, internal estimates and expected future capabilities separate when reading the argument.
Why evaluation is the hinge
If an AI system generates a hundred candidate improvements, something still has to establish which are better.
For a narrow algorithm, a test may offer a relatively clear measure. For a general-purpose model, improvement can involve competing goals: capability, reliability, cost, speed and behaviour across unfamiliar tasks.
A higher score on one test may conceal a regression elsewhere. Faster production of code also does not necessarily mean faster discovery of a valid idea.
The more development is automated, the more important the evaluation process becomes. Who selected the objective? Can the system exploit the test? Who decides whether a candidate should be deployed?
Why an Australian reader should care
This does not need an immediate small-business use case to deserve coverage. It concerns how quickly the technology may change and how much time users, researchers and institutions have to understand those changes.
It may also affect the durability of product plans. A business choosing a tool should expect continuing change without turning a research forecast into a purchasing timetable.
When the next dramatic claim appears, ask what completed the loop: people supervising a task, automated evaluators inside a defined experiment, or a demonstrated successor-development process.
AI helping build AI is already a worthwhile story. The evidence becomes clearer when the headline names which part of the work it actually did.