Anthropic launches the world’s first AI model “Hybrid Reflection”

The difference between a conventional model and reasoning is similar to the two types of thinking described by Nobel Prize winner Michael Kahneman in his 2011 book. Thinking quickly and slowly: Fast and instinctive thinking of the system-1 and slower thinking of the system-2.

The type of model, which made Chatgpt possible, known as a large language model or LLM, gives instant answers to prompted by requesting a large neural network. These exits can be strikingly smart and agreed, but they may not answer questions that require step -by -step reasoning, including simple arithmetic.

LLM may be forced to imitate conscious reasoning if it is instructed to come up with a plan that should then follow. However, this trick is not always reliable and models usually fight to solve problems that require wide, careful planning. Openai, Google and now Anthopic use all Machine learning method known as strengthening training To make their most recent models learn to generate reasoning that point to the right answers. This requires the collection of additional training data from people to solve specific problems.

Pen says Claude’s reflection regime has received additional business applications information, including writing and fixing code using computers and answering complex legal questions. “The things we made improvements are … technical topics or objects that require long reasoning,” Pen. says. “What we have from our clients is a great interest in implementing our models in their actual loads.”

Anthropic says Claude 3.7 is particularly good at solving encoding problems that require step-by-step reasoning, exceeding O1 on Openai on some indicators such as SWE-Tala. The company today launches a new tool called Claude Code, specially designed for this type of AI-assisted coding.

“The model is already good at encoding,” Pen. says. But “Additional thinking would be good for cases that may require very complicated planning – let’s say you are looking at an extremely large code base for company.”