What is Anthropic’s Claude and what is it capable of?

Claude is a family of large language models (LLM) built by Anthropic. Like other frontier LLMs, they can be used for many tasks including summarization, search, creative and collaborative writing, Q&A, and coding. They are trained to prioritize helpfulness, honesty, and harmlessness (HHH), though their outputs may still contain bias, factual errors, and hallucinations.

Anthropic considered releasing Claude to the general public in spring 2022, but waited until March 2023, after the release of ChatGPT, with the stated aim of avoiding triggering an arms race. Claude was initially only available in the US and UK, but has since incrementally been rolled out to many other countries.

Notable facts about Claude include:

• Anthropic reports that it has made Claude more resilient to adversarial attacks using Constitutional AI and Moral Self-Correction.

• The models have no access to the internet when answering user prompts.

• It does not remember anything from prior message threads since its memory is cleared between threads.

• As of July 2024, the context window length for the most powerful Claude models is about 200,000 tokens, which works out to about 140,000 words (enough to load the entire novel The Great Gatsby twice).

• Claude 3.0 was trained on data up to August 2023. The model is constantly being updated and the support page offers the latest information. As of July 2024, the latest model is Claude 3.5 Sonnet.