Improvements in Math, Coding, Reasoning, and Safety
Anthropic, the AI startup co-founded by ex-OpenAI executives, has launched Claude Instant 1.2, an updated version of its text-generating model. The new version incorporates the strengths of the company’s flagship model, Claude 2, and has shown significant gains in areas like math, coding, reasoning, and safety. In tests, Claude Instant 1.2 scored 58.7% on a coding benchmark and 86.7% on a set of math questions, outperforming its predecessor.
Enhanced Performance and Functionality
Claude Instant 1.2 generates longer and more structured responses, while also improving in quote extraction, multilingual capabilities, and question answering. Moreover, it is less likely to produce incorrect or nonsensical text and is more resistant to jailbreaking attempts, a technique used to bypass safety features in large language models.
Context Window Increases Model Performance
Claude Instant 1.2 features a context window of 100,000 tokens, the same size as Claude 2. A larger context window ensures that the model retains the content of recent conversations, reducing the likelihood of forgetting important information. Both Claude Instant 1.2 and Claude 2 can analyze approximately 75,000 words, equivalent to the length of “The Great Gatsby.”
Competition with OpenAI and Other Startups
Anthropic’s Claude Instant is designed to compete with similar entry-level offerings from OpenAI, Cohere, AI21 Labs, and other startups in the text-generating AI space. The company aims to build a next-generation algorithm for AI self-teaching, which could be used to create virtual assistants capable of answering emails, performing research, and generating art and books.
Progress Towards Ambitious Goals
Since its launch in 2021, Anthropic has raised $1.45 billion in funding. However, the company estimates it will need $5 billion over the next two years to develop its envisioned chatbot. Anthropic has already secured thousands of customers and partners, including Quora, DuckDuckGo, and Notion, integrating its AI models into various applications and platforms.