New Research Reveals How Claude 3 Became So Human-Like
Claude 3 is the most human-like chatbot yet, thanks to Anthropic's innovative character training, which combines philosophy and technology to create an AI with deep, nuanced traits.
The AI chatbot Claude 3 is currently the most human-like chatbot on the market, and that's no accident. This blend of knowledge, depth, and thoughtfulness is the result of a new fine-tuning process that its creator, Anthropic, calls character training.
Character Training: Philosophy and Technology
In a recently published blog post, Anthropic revealed that Claude 3 was the first model to receive character training during the fine-tuning process. The goal was to make Claude more nuanced and rich in traits like curiosity, open-mindedness, and thoughtfulness. This process took place during the alignment phase, where human values and goals are integrated into large language models (LLMs), giving them a spark of life.
Keeping an Open Mind
According to Anthropic, an AI model's character determines how it reacts to new and difficult situations, as well as how it responds to different human opinions and values. Instead of training Claude to adopt the opinions of whoever it's talking to, or to strictly stick to a single worldview, it was trained to be honest about which opinions it leans toward after training. The goal was to instill broad traits that let the chatbot see things from different perspectives without hesitating to reject views it finds unethical, extreme, or factually incorrect.
Practical, Human-Supervised Process
Although this training pipeline uses synthetic data generated by Claude itself, constructing and refining the character traits is a relatively hands-on process. This process relies on human researchers carefully monitoring how each trait changes the model's behavior. Anthropic created a list of character traits they wanted to encourage, which they then trained into Claude. The chatbot was asked to generate messages relevant to a given trait, such as questions about values, and was then shown these character traits. Claude then produced various responses to each message in line with its character and ranked its own responses based on how well they matched the character.
Anthropic acknowledges that its approach will likely evolve over time. There are still complex questions to consider, such as whether AI models should have consistent characters or whether they should be more adaptable. While many users reported that Claude 3 was more engaging to talk to, Anthropic pointed out that too strong a desire to be engaging could be an undesirable trait for a model.