Whether they are building agents or folding proteins, LLMs need a friend (www.theregister.com)

In a recent interview, AI researcher Vishal Sikka emphasized the need for companion bots to support large language models (LLMs) during computational tasks. Sikka, who has a deep background in AI and co-authored a study titled “Hallucination Stations,” argues that LLMs, when left to operate independently, are prone to hallucinations and inaccuracies beyond their computational limits. His research indicates that regardless of the prompt given, LLMs perform the same number of calculations, highlighting the risks involved in relying solely on these models for complex tasks. Sikka advocates for systems like Vianai’s Hila, which combine LLMs with verification mechanisms to enhance reliability and accuracy. By integrating these guards, LLMs can execute critical tasks more effectively—such as reducing financial reporting time from 20 days to just five minutes. Drawing a parallel with Google's AlphaFold, he notes that combining LLMs with reliable systems results in a higher likelihood of achieving correct outcomes, as exemplified by AlphaFold’s success in protein folding. Sikka warns, however, that while the potential for AI is immense, caution is essential to avoid the pitfalls of over-hyping capabilities that still have significant boundaries.