Claude

A family of foundational AI models created by Anthropic, designed to be helpful, harmless, and honest.

How it works

Claude is trained using a technique developed by Anthropic called Constitutional AI. This approach gives the model a set of principles (a 'constitution') to follow during its training phase, allowing it to self-correct and avoid generating harmful, biased, or unhelpful responses. Claude uses a Transformer architecture similar to other large language models, but its fine-tuning process heavily emphasises safety and alignment. Anthropic also invests in mechanistic interpretability research to understand what is happening inside the model.

Why it matters

Claude represents a significant step forward in AI safety and alignment. In enterprise environments and customer-facing applications, businesses need AI models they can trust not to hallucinate wildly or produce toxic content. Claude's focus on being helpful, honest, and harmless makes it a preferred choice for legal, medical, and customer support use cases. Its Constitutional AI methodology has also influenced how other labs approach alignment.