Anthropic co-founder shifts timeline for self-improving AI models to 2027

Jack Clark, a co-founder at AI research company Anthropic, has moved up his expected timeline for self improving artificial intelligence from 2028 to 2027. According to his newsletter Import AI, this revision comes after observing recent leaps in automated scientific research skills and mathematical problem solving. Researchers are seeing software make faster gains in areas that once required human intuition, such as deciding which technical experiments are worth running next.

If software becomes capable of upgrading its own capabilities without human help, the pace of technological development could accelerate exponentially. AI models are already being tested on tasks like data curation and scientific research taste, while separate studies highlight rapid progress in visual reasoning and autonomous systems. Should these self improving systems arrive by 2027, the shift could radically change how software is built, leaving tech companies and regulators with much less time to adapt to autonomous systems.

Despite these revised predictions, major bottlenecks remain before AI can safely manage its own code. Current models still struggle with basic visual intuition and real world physical reasoning, failing frequently on tasks that humans solve easily. It remains unclear whether today's software architectures can truly achieve safe recursive self improvement, or if these shorter timelines overemphasize narrow benchmark gains while ignoring deep reasoning gaps.