Anthropic, a company founded by exiles from OpenAI, has introduced its first hybrid AI model that can produce either conventional output or controlled reasoning to solve complex problems. The new model, called Claude 3.7, aims to make it easier for users and developers to tackle tasks requiring both instinctive output and step-by-step cogitation.
Claude 3.7 features a “scratchpad” that reveals the model’s reasoning process, allowing users to understand how the model is working on a problem and modify or refine prompts accordingly. This feature has proven popular with other AI models, including DeepSeek, which uses it to help users grasp its thought process.
The company’s product leads claim that Claude 3.7 can help increase the capabilities of AI models by enabling them to reason over problems more effectively. OpenAI introduced a similar reasoning model in September 2024, but users have to switch between models to access this feature. Google has also released a similar offering for its Gemini model, called Flash Thinking.
Claude 3.7 is designed to bridge the gap between fast and instinctive System-1 thinking and slower, more deliberative System-2 thinking, as described by Nobel-prize-winning economist Michael Kahneman. The model can produce instantaneous responses but may fail to answer questions that require step-by-step reasoning.
To overcome this limitation, Anthropic is using reinforcement learning to train its models on solving specific problems. This method requires additional training data from humans, which the company has been collecting for business applications such as writing, fixing code, and answering complex legal questions.
Claude 3.7 has shown promise in solving coding problems that require step-by-step reasoning, outscoring OpenAI’s o1 model on some benchmarks like SWE-bench. The company is releasing a new tool called Claude Code to facilitate AI-assisted coding.
Overall, Claude 3.7 aims to provide users with more control over the behavior of their AI models and improve their ability to tackle complex tasks requiring reasoning and instinctive output.
Source: https://www.wired.com/story/anthropic-world-first-hybrid-reasoning-ai-model/