What ChatGPT said about Claude and Anthropic AI

I was having a chat with our good friend ChatGPT. We were writing product descriptions for my Etsy store as I turned out work and I decided to pause and ask some questions about Large Language Models, NLP and chat bot training.

I got curious, so I had to start asking it about Anthropic AI’s claim of creating language models that follow a “constitution”.

Constitutional AI is what Claude runs on, Anthropic’s answer to ChatGPT and pre-trained generative chatbot models. They tout it on their home page thusly;

With Constitutional AI built in, Claude is designed to reduce brand risk. Best in class data retention, and no training on your data.

Anthropic AI

Looking at their model card and published paper reveals what this means;

Our core research focus has been training Claude models to be helpful, honest, and harmless. Currently,
we do this by giving models a Constitution – a set of ethical and behavioral principles that the model uses
to guide its outputs. You can read about Claude 2’s principles in a blog post we published in May 2023. Using this Constitution, models are trained to avoid sexist, racist, and toxic outputs, as well as to avoid helping a human engage in illegal or unethical activities.

However, Claude 2 certainly isn’t perfect and can still make mistakes. Like all models, Claude can be
jailbroken, and our work to make Claude more helpful, harmless, and honest is ongoing.
Ethical considerations also shape our Acceptable Use Policy (AUP)

Anthropic AI’s Claude 2 Model Card

We can see from the write up and the published materials that Claude 2 is trained like ChatGPT, to decline requests, to monitor chats for acceptable use and respond appropriately.

Here’s what ChatGPT had to say about all this:

The term “constitution” used by a company in the context of AI training is likely a form of marketing speak and may refer to the set of guidelines or rules that are applied during the training process. These guidelines are designed to shape the behavior of the AI model and ensure it adheres to specific ethical or safety standards.

In the context of AI training, reinforcement learning (RL) is one of the techniques that can be used to fine-tune the model’s behavior. RL involves a feedback loop, where the model receives rewards or penalties based on its generated responses. By using RL, developers can train the model to improve its behavior over time and align it with desired outcomes.

The “constitution” may contain rules and guidelines for the AI model, specifying what responses are encouraged or discouraged based on various inputs. For instance, it might include guidelines for refusing inappropriate or harmful requests, promoting helpful and respectful responses, and adhering to ethical considerations.

It’s important to note that the term “constitution” is not a standard term in the field of AI and machine learning. Different companies may use unique terminology to describe their AI systems and the training processes. When evaluating claims made by AI companies, it’s essential to look into the details of how the model is trained, what guidelines are applied, and how ethical considerations are addressed.

In any case, the focus should always be on transparency, responsible use, and ensuring AI systems are designed to benefit users while adhering to ethical principles and avoiding harmful behaviors.

ChatGPT, August 1st 2023