The difference between the conventional model and reasoning is similar to two types of thinking described by the winning economist Nobel Michael Kahneman in his book from 2011 Thinking quickly and slowly: Speedy and instinctive thinking System-1 and slower more prudent thinking of System-2.
The type of model that chatgpt, known as a immense language model or LLM, presents immediate answers to the prompt, asking about a immense neural network. These outputs can be strikingly clever and consistent, but they may not answer questions that require reasoning step by step, including elementary arithmetic.
LLM may be forced to imitate careful reasoning if it consists of developing a plan that must occur. This trick is not always reliable, and models usually fight to solve problems requiring extensive, careful planning. OpenAI, Google, and now Anthropics apply the machine learning method known as reinforcement learning to get the latest models to learn to generate reasoning that indicate correct answers. This requires collecting additional training data from people about solving specific problems.
Penn claims that Claude’s reasoning mode has received additional data on business applications, including writing and setting the code, using computers and answering complicated legal questions. “Things we have introduced improvements are … technical items or entities requiring long reasoning,” says Penn. “What we have from our clients is a great interest in implementing our models in their actual loads.”
Anthropic claims that Claude 3.7 is particularly good in solving coding problems that require step-by-step reasoning, exceeding O1 Openai on some comparative tests, such as Swo-Bench. The company today releases a novel tool called Claude Code, specially designed for this type of coding supported by AI.
“The model is already good in coding,” says Penn. But “additional thinking would be good in cases that may require very complex planning – you will say that you look at the extremely large code base for the company.”