The Wall Street Journal reported that AI systems are beginning to help build better AI, making research and development faster and raising concern among people working on the technology. For technical teams, the relevant point is a measurable shift in what models can do in software work.
Longer software tasks
METR’s published work frames the measurement problem as AI ability to complete long software tasks. AIDIRECTORY summarised the related chart as showing top AI models doubling task length about every 7 months.
That figure describes the length of task a model can handle. It does not, by itself, establish that a model can run a software project or an AI lab. It does indicate that the boundary between short assistance and longer delegated work is moving.
For developers and platform teams, task length is a practical measure. Short tasks can be checked locally and quickly. Longer tasks bring in context management, hidden dependencies, testing discipline and review overhead.
Self-improvement needs careful handling
The phrase self-improvement covers a wide range of activity. In the material described here, AI contributes to research and development that can lead to better AI systems. That is different from assuming a fully autonomous improvement loop.
The governance questions follow from that distinction. If models take on longer engineering or research tasks, teams need to decide how they evaluate outputs, record model use and approve changes. Review processes built for code suggestions may not be enough for multi-step work.
The issue also affects how organisations discuss agentic AI. A model that can persist across longer tasks will be treated differently from a chat interface used for isolated prompts. Policies, testing and accountability need to reflect the actual work being delegated.
When AI systems start contributing to AI development, the practical questions are how much autonomy to permit, how to evaluate longer tasks and how to govern model-driven change, and those questions run through ECS's Core AI topic, alongside agentic AI, Azure AI Foundry, the EU AI Act and responsible AI. Sessions and workshops give attendees room to compare evaluation methods, governance choices and practitioner experience from current AI projects. Bring the use cases and control questions that are difficult to settle from model benchmarks alone.
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.