Abstract
Research powered by artificial intelligence made leaps last year, and it is here to stay. AI 'agents' that integrate several large language models (LLMs) to carry out complex, multi-step processes are likely to be used more widely, some with little human oversight. This year might even bring the first consequential scientific advances made by AI. But heavier use could also expose serious failures in some systems. Researchers have already reported errors that AI agents are prone to, such as the deletion of data. This year will also bring techniques that move beyond LLMs, which are expensive to train. Newer approaches focus on designing small-scale AI models that learn from a limited pool of data and can specialize in solving specific reasoning puzzles. These systems do not generate text, but process mathematical representations of information. In 2025, one such tiny AI model beat massive LLMs at a logic test.
NETL
NSTL