Discussion

AI is helping build better AI. The self-improving loop is still a long way off

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#6113

AI firms are using their models to speed up the work of developing other models, from coding to experiments. But the industry’s ambition for AI to design and improve its own successors remains a goal, not an achievement, according to The Logic’s reporting.

Watch Desk analysis

What happened

The Logic reports that Google, Anthropic and other firms are using AI for parts of machine-learning research and development. Google says AI has helped staff write code and iterate on Gemini’s system instructions. The article reports that Google employees wrote 75 per cent of the company’s new code this spring, while Anthropic employees wrote more than 80 per cent.

That is a meaningful shift in the tools researchers use, but it is not yet the full recursive self-improvement described in the industry’s grander ambitions: an AI independently conceiving research, building a successor, assessing the results and deciding what to try next. Mila scientific director Hugo Larochelle told The Logic that current systems are better at work-hours tasks such as coding and testing variations than at deciding which unsolved problems deserve attention.

Why it matters

If AI can make parts of AI development faster, research teams may be able to try more ideas with the same time and resources. That could accelerate progress, but it also makes the boundaries of human oversight a practical question, not just a debate about distant science fiction.

The distinction matters. Automating coding and experiments is already useful; a system choosing the research agenda and reliably improving its own successors is a much bigger claim. The article says no one has achieved that full loop yet.

Our read

This is an important trend, not proof that AI is about to take over its own research department. The useful signal is that AI is increasingly part of the development process itself, while humans still choose the problems and judge the results. The leap from capable assistant to autonomous scientist is where the hard questions begin, and where the evidence needs to catch up with the pitch.

What to watch

  • Whether companies publish concrete measures of AI’s contribution to research and model development.
  • Whether AI systems move beyond coding and experiment support to choosing research questions.
  • How much human review remains in the loop as these tools take on more of the work.

Discussion spark: If AI helps researchers write code and run experiments, should people still make every major choice about what the next generation of AI is built to do?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.