AI thinking

AI models that think 🤔

There are two kinds of “intelligence”. When we are faced with a problem, we either:

🧠 know the solution instinctively, or

🤔 work out the answer by thinking

Expertise and experience form our instincts in domains that we specialise in. This gives us the ability to, for example, estimate the value of an asset, or the value of a research idea, better than non-experts. But, there are tasks that it would be impossible to solve by instinct alone, such as generating a mathematical proof or debugging an error in a large codebase - for these kinds of problems we need to think.

While AI’s “instinct” is getting better and better with larger models and datasets, we are now transitioning into an era where AI can also "think".

In order to achieve that, AI needs to be able:

🔡 break down a problem into smaller ones

🗓️ develop a plan

✅ validate its reasoning

This is particularly important for workflows requiring advanced levels of introspection, planning and validation. Tasks such as doing research, writing code, or acting as an AI assistant, would greatly benefit from better thinking. As a real world example, a few months ago we helped a retailer build an in store assistant and some of the failure cases where good examples of the model not spending enough time on the problem like working out with prices and dates as well as the need to search multiple times to bring all relevant information to the user.