Anthropic, an AI firm founded by former OpenAI researchers, has revealed its updated chatbot Claude 2, which aims to take on rivals ChatGPT and Google Bard. The new model boasts impressive features such as a higher token limit, improved coding skills, and a public-facing beta website.
The creators ambitiously describe Claude as a “friendly, enthusiastic colleague or personal assistant who can be instructed in natural language to help you with many tasks.” They may be right. Claude 2 can handle 100,000 tokens per prompt, meaning it can engage with content as long as a 75,000-word book, far surpassing Claude 1.0’s 9,000 token limit and Bard and ChatGPT4’s 8,192 token limits.
Claude 2’s expanded capabilities are reflected in its test results. The chatbot scored 76.5 percent on the multiple-choice section of the Bar exam and ranked in the 90th percentile in the GRE reading and writing exams. The chatbot’s coding skills are also advanced, with a 71.2 percent score on a Python coding test.
But Claude 2 isn’t just about smart engagement and complex tasks; it also emphasizes safety. Anthropic has strived to make Claude more harmless and resistant to generating offensive or potentially dangerous outputs. This update comes just five months after the debut of Claude 1.0.
The developers have designed Claude with a unique “constitution” inspired by the UN Declaration of Human Rights and DeepMind’s Sparrow principles. This set of rules allows the chatbot to self-improve, identify inappropriate behavior, and adapt its conduct without human feedback.
Beyond its immediate functional use, Claude 2 has broader implications in the field of digital art and creation. For instance, the use of AI like ChatGPT in projects such as the Turbo meme coin and lore building for Forgotten Runes demonstrates AI’s creative potential. The advent of Claude 2 signifies another step in this journey of text, visual, and video-based AI artwork, benefiting the entire ecosystem.
Editor’s note: This article was written by an nft now staff member in collaboration with OpenAI’s GPT-4.