The Expanding Attention Span of AI

Imagine hiring an assistant whose ability to focus independently doubles every few months. Two years ago, they could only work unsupervised for 30 seconds before needing your help. Today, they can manage 40 minutes of complex problem-solving. By next year, they might be running entire projects for days at a time.

This isn't a hypothetical employee—it is the actual trajectory of artificial intelligence. According to Jack Clark, author of the widely read newsletter Import AI, this rapidly expanding "time horizon" of autonomy points toward a startling milestone: by the end of 2028, there is a greater than 60% probability that AI systems will be capable of fully automated research and development. In other words, an AI could autonomously build its own successor.

While the idea of AI building AI sounds like a trope borrowed from science fiction, the evidence driving this forecast is grounded in concrete software engineering metrics. Two major trends suggest that the foundational pieces for automated AI development are already falling into place.

The first trend is raw coding competency. AI models are instantiated through software, and their ability to manipulate code has skyrocketed. On SWE-Bench, a rigorous evaluation that tests how well AI systems can resolve real-world software issues from GitHub, progress has been exponential. In late 2023, top models like Claude 2 hovered around a 2% success rate. Today, advanced preview models are scoring closer to 94%, effectively maxing out the benchmark. AI is no longer just generating boilerplate code; it is debugging, testing, and solving complex architectural problems.

The second, and perhaps more crucial trend, is the ability to chain these tasks together without human intervention. The research organization METR tracks how long an AI can reliably work independently. In 2022, models were reliable for tasks taking about 30 seconds. By 2024, reasoning models like o1 pushed that boundary to roughly 40 minutes. Prominent AI forecasters, such as Ajeya Cotra, suggest that by the end of 2026, AI systems could independently execute workflows that would typically take a skilled human 100 hours to complete.

If you look closely at the day-to-day work of human AI researchers, much of it consists of tasks that fit neatly into these expanding time horizons: cleaning datasets, monitoring training runs, and launching experiments. As AI systems become capable of managing 100-hour workflows, they will naturally begin to absorb the engineering heavy-lifting of AI development itself.

We may see a proof-of-concept within a year or two, likely starting with smaller, non-frontier models training their own iterations. If this scaling continues, the bottleneck in technological advancement will shift away from human labor hours. We are approaching a threshold where AI transitions from a tool we painstakingly refine into an autonomous engine driving its own evolution.

Key Points

• There is an estimated 60%+ chance that AI could achieve fully automated R&D by 2028, building its own successors.
• AI performance on real-world coding benchmarks has surged from 2% to nearly 94% in a very short timeframe.
• The duration an AI can work independently has grown from 30 seconds in 2022 to 40 minutes in 2024, with projections hitting 100 hours by 2026.
• The daily tasks of AI researchers—like data cleaning and running experiments—are increasingly falling within AI's autonomous capabilities.

Why It Matters

As AI systems gain the ability to research and develop their own successors, the pace of technological advancement will decouple from human biological limits, fundamentally accelerating the future of innovation.

Sources:

• Import AI 455: AI systems are about to start building themselves. — Import AI (Jack Clark)