The difference between a conventional model and an reasoning is similar to the two types of thinking described by the Nobel Prize-winning economist Michael Kahneman in his book 2011 Think quickly and slowly: Think fast and instinctive system-1 and think more slowly more deliberative system-2.
The type of model that made Chatgpt possible, known as a large language model or LLM, provides immediate answers to a prompt by asking a large neural network. This output can be remarkably smart and coherent, but may not be able to answer questions that require step -by -step reasoning, including simple arithmetic.
An LLM can be forced to simulate deliberative reasoning if it is instructed to come up with a plan that it must follow. However, this trick is not always reliable and models usually struggle to solve problems that require extensive, careful planning. OpenAI, Google and now anthropic use all use a method for machine learning that is known as reinforcement to get their latest models to learn to reason that points to correct answers. This requires the collection of additional training data from people about solving specific problems.
PNN says that Claude's reasoning mode has received additional information about business applications, including writing and confirming code, the use of computers and answering complex legal questions. “The things we have made improvements are … technical topics or topics that require long reasoning,” says Penn. “What we have of our customers is a lot of interest in implementing our models in their actual workload.”
Anthropic says that Claude 3.7 is especially good at solving coding problems that require step-by-step reasoning, which surpasses O1's O1 about some benchmarks such as SWE-Bench. The company is publishing a new tool today, called Claude Code, specifically designed for this type of AI-assisted coding.
“The model is already good at coding,” says Penn. But “Extra thinking would be good for cases that may require a very complex planning – say that you are looking at an extremely large code base for a company.”
