AI may be incapable of scientific breakthroughs

We might not want to rely on AI for the heavy lifting of innovation.

The Big Picture: While enthusiasts predict that AI will vastly accelerate the pace of scientific progress, some experts are skeptical of that claim. They allege that while AI is highly efficient, it’s not original and can, therefore, only help a human come to their own conclusions, like a “digital yes-man.”

Between the Lines: Thomas Wolf, the co-founder of Hugging Face, explained his reservations at VivaTech in Paris.

• While the structure of Large Language Models (LLMs) makes them good at predicting the next word or stage in a sentence or problem, it doesn’t give them anything resembling the human faculty of originality.
• That’s why, in science, LLMs excel at answering questions but don’t know how to ask good ones.
• Since (Wolf argues) the most important scientific breakthroughs come from reframing old questions, LLMs won’t make breakthroughs — only humans will.

Conclusion: This isn’t to say that AI can’t be helpful to scientists. There are many problems, especially in physics, that require brute-force computation to solve or verify, and AI is well-suited to those purposes. But the spark of genius — that’s still up to us.

Prediction: Don’t expect physics labs to cut most of their human staff. Instead, look out for a more symbiotic relationship between AI and humans than what exists in most industries.