Cheap AI Almost Matches Fancy AI for Science Tasks

Scientists often need to know what a protein does—whether it fights disease or helps digest food. Usually, this takes years of lab work. Now, AI can help by looking at a protein's sequence and predicting its function. But not all AI is created equal, and the most powerful versions can be shockingly expensive.

A new study tested different AI strategies for this protein-matching task. They compared giant language models (like the ones behind ChatGPT) with a smaller, rule-based AI that learned from past examples. The results surprised them: the cheap, rule-based AI nailed the correct answer 88% of the time, while the most expensive model got 92-94%. The cheap one was also lightning fast and never made a mistake due to randomness. The big difference? It can't explain how it reached its answer.

The researchers also tried running these AI systems across multiple institutions—sharing data and models like a joint research team. They expected this "federation" to slow things down or hurt accuracy. Instead, it barely made a dent. That's great news because sharing data across hospitals or research centers is often the only way to get enough information for rare diseases.

So what's the practical takeaway? For routine, well-understood tasks—like checking a protein against known families—a simple, cheap AI is the smart choice. It saves money and is fully consistent. The fancy, reasoning AI is better reserved for brand-new, open-ended questions where you need creative hypotheses. In other words, you don't always need the most expensive brain in the room—sometimes a reliable workhorse gets the job done just fine.