Imagine an AI progression that begins with digital co-workers, moves into original scientific discovery and then reaches machines capable of performing tasks in the physical world. Sam Altman has described a rough sequence for the next stage of artificial intelligence. He pointed to working AI agents first, systems producing novel insights next and robots following after that, while acknowledging that predictions about exact dates remain uncertain.
In The Gentle Singularity, Altman wrote that 2025 had brought agents capable of genuine cognitive work, particularly in software development. He expected 2026 to bring systems able to find new insights and suggested that 2027 could see robots completing real-world tasks. These statements are forecasts, not a guaranteed product schedule, but they reveal how he imagines AI crossing from assistance into increasingly independent intellectual and physical activity.
OpenAI’s more recent plan adds another milestone: an automated AI researcher. Altman and OpenAI chief scientist Jakub Pachocki wrote that the company aims to build a system that can accelerate and increasingly automate the research process while remaining steerable and accountable. Their internal expectation is that, by March 2028, a significant share of OpenAI’s research could be performed by AI systems working alongside human researchers.
Each step raises a different verification problem. A coding agent can introduce a bug, a scientific agent can generate a persuasive but false hypothesis and a physical robot can cause immediate damage. Progress in capability therefore cannot be measured only by whether a system completes a task. Reliable evaluation, restricted permissions, monitoring and human approval become more important as the distance between an instruction and its real-world consequences grows.
For businesses, the roadmap suggests a staged approach to adoption. Today’s useful opportunity is often a supervised agent working within a defined digital process. Scientific and physical applications demand stronger evidence and industry-specific controls. Altman’s timeline may prove too fast, too slow or wrong in important details, but the direction is already influencing investment. Organisations should prepare for systems that do more than generate content while resisting the temptation to treat a confident prediction as a reason to deploy unproven autonomy.