The Math on AI Agents Doesn’t Add Up

City: The big AI companies once said that 2025 would be the year of AI agents. But now, it seems that 2025 will just be a year for talking about AI agents. Some experts think we might be waiting a long time for these robots that do tasks for us.

A new paper says that AI models, like large language models, can’t handle very complex tasks. The authors, including a former SAP CTO and his son, used math to show these limits. They think that even advanced AI won’t solve this issue. “They cannot be reliable,” said Vishal Sikka, a leader in AI and head of a startup called Vianai. He believes we shouldn’t expect AI to run important jobs like managing nuclear plants.

However, not everyone agrees. Many in the AI industry claim there have been big steps forward, especially in coding. Recently, Google reported exciting news about reducing mistakes, while a startup named Harmonic showed new ways to make AI more reliable. This company has a tool called Aristotle that uses math to ensure AI can be trusted with coding tasks.

Co-founder Tudor Achim mentioned, “Are we stuck with AI that only creates junk? That would be crazy.” He believes most AI models can handle some practical tasks, like planning a trip.

Both sides have points. Everyone knows that errors in AI, called hallucinations, are a big problem. A report from OpenAI found that even their best models sometimes made up information. They reminded us that AI will never be 100% accurate.