Artificial intelligence company Anthropic has introduced the first hybrid model, called Claude 3.7, that can produce conventional output or controllable reasoning to solve complex problems. This new model makes it easier for users and developers to tackle issues requiring a mix of instinctive output and step-by-step thinking.
Claude 3.7 features a “scratchpad” that reveals the model’s reasoning process, helping users understand how the model works over a problem. This feature is particularly useful when combined with the ability to adjust the level of reasoning. For example, if the model struggles with breaking down a problem correctly, users can ask it to spend more time working on it.
The development comes as other AI companies, such as OpenAI and Google, focus on getting their models to “reason” over problems to increase capabilities and broaden usefulness. Anthropic’s Claude 3.7 stands out for its ability to solve coding problems with step-by-step reasoning, outscoring OpenAI’s o1 in some benchmarks.
The model’s enhanced reasoning capabilities are the result of reinforcement learning, which involves gathering additional training data from humans on solving specific problems. This approach has also been used by OpenAI and Google to improve their own models.
Anthropic plans to release a new tool called Claude Code, specifically designed for AI-assisted coding. The company believes its model will be particularly useful in technical subjects that require long reasoning, such as writing code or answering complex legal questions.
Source: https://www.wired.com/story/anthropic-world-first-hybrid-reasoning-ai-model