To receive industry-leading AI updates and exclusive content, sign up for our daily and weekly newsletters. Learn more
OpenAI rival startup Anthropic released system prompts for its Claude family of AI models yesterday and promised to continue doing so in the future, appearing to set a new standard for transparency in the rapidly evolving AI industry, according to observers.
System prompts act like a playbook for large language models (LLMs), telling the model general rules to follow when interacting with a user, what behaviors or personalities the model should exhibit, and often indicate a cut-off date for the information the LLM learned during training.
Most LLMs have system prompts, but not all AI companies expose them. Uncovering a model's system prompts has become something of a hobby for AI jailbreakers.
But now Anthropic has beaten the jailbreakers to it, publishing operating instructions for the Claude 3.5 Sonnet, Claude 3 Haiku, and Claude 3 Opus models in the release notes section of their website.
Additionally, Alex Albert, head of developer relations at Anthropic, posted on X (formerly Twitter) a promise to keep the public updated on system prompts, writing: “We will be documenting the changes we make to the default system prompts in Claude dot ai and our mobile apps.”
Added a new system prompts release notes section to the documentation to document the changes we made to the default system prompts in the Claude dot ai and mobile apps. (System prompts do not impact the API.) pic.twitter.com/9mBwv2SgB1
— Alex Albert (@alexalbert__) August 26, 2024
What Anthropic's system prompts reveal
The system reveals fascinating details about the three models, Claude 3.5 Sonnet, Claude 3 Haiku, and Claude 3 Opus, including their respective features, knowledge date cutoffs, and various personality quirks.
Claude 3.5 Sonnet is the most advanced version, with an updated knowledge base as of April 2024. It provides detailed answers for complex questions and concise answers for easy tasks, emphasizing both accuracy and brevity. The model handles controversial topics with care, presenting information without explicitly labeling it as sensitive or claiming objectivity. Additionally, Claude 3.5 Sonnet avoids unnecessary filler phrases and apologies, and is particularly careful about how it handles image recognition, never admitting to recognizing a face.
Claude 3 Opus works with a knowledge base updated as of August 2023 and excels at handling complex tasks and writing. It is designed to respond succinctly to simple queries and thoroughly to more complex questions. Claude 3 Opus addresses controversial topics by offering a wide range of perspectives, avoiding stereotypes, and providing a balanced view. It has some similarities to the Sonnet model, but does not incorporate detailed behavioral guidelines such as avoiding apologies and unnecessary assertions.
Claude 3 Haiku is the fastest model in the Claude family, updated in August 2023. It is optimized to provide quick and concise answers to simple questions, and thorough answers when needed for more complex issues. Haiku's prompt structure is easier to understand than Sonnet, focusing primarily on speed and efficiency, without the more advanced behavioral nuances found in the Sonnet model.
Why is Anthropic's system prompt release important?
A common complaint about generative AI systems is the notion of a “black box” – that it's difficult to figure out why or how the model arrived at a decision. The black box issue has led to research into AI explainability, a way to reveal a model's predictive decision-making process. Public access to system prompts is a step toward exposing the black box a little bit, but only to the extent that people can understand the rules that AI companies have set for the models they've created.
AI developers praised Anthropik's decision, noting that Clode's public disclosure of documentation about its system prompts and updates sets it apart from other AI companies.
Now you'll see all three versions of Claude, along with a system prompt for when they were last updated. This is a great change and one I hope will eventually be adopted industry-wide. Great work from Anthropic. Transparency! https://t.co/PmgD0HpnpH pic.twitter.com/2E4zP4LsVz
— Andrew Curran (@AndrewCurran_) August 26, 2024
It's not completely open source
Publishing the system prompts for the Claude model does not mean that Anthropic has published the model family: the actual source code to run the models, the training data set, and the underlying “weights” (or model settings) are solely under the control of Anthropic.
Still, the release of Anthropic's Claude system prompts points the way for other AI companies to bring greater transparency to their AI model development, and it also benefits users by showing them how their AI chatbots are designed to work.
VB Daily
Stay up to date! Get the latest news every day by email
Thanks for subscribing! Check out other VB newsletters here.
An error has occurred.
