Highlights:

  • Claude 2, an update to Claude 1.3, is now available to users in the US and UK as a beta chat experience. The chatbot is also offered to businesses through an application programming interface for the same price as the old form.
  • The new Claude 2 chatbot has many new features that make it better at code, math, and thinking than its predecessors.

Anthropic is an AI research company that aims to build trustworthy AI models to compete with OpenAI LP’s ChatGPT. Recently, it released a new version of its Claude chatbot that is much safer and better at coding.

Claude 2, an update to Claude 1.3, is now available to users in the US and UK as a beta chat experience. The chatbot is also offered to businesses through an API for the same price as the old model.

Claude is an AI chatbot that can understand conversational natural language and respond with comments, research, poetry, analysis, summaries of long papers, and even help with code.

The Anthropic team said in a blog post, “Think of Claude as a friendly, enthusiastic colleague or personal assistant who can be told in natural language to help you with many tasks.”

Claude 2’s training data has been updated with 2022 and early 2023 information. This means that it knows a lot more about what’s going on now than it did before. Users should understand that it may still make mistakes or be confused about things it doesn’t know much about. So, the team warned that Claude might still have “confabulations” or “hallucinations,” which are things that chatbots sometimes generate.

The context window in the model is much longer. Earlier this year, the company increased it from 9,000 to 100,000 tokens, equal to 75,000 words. This means people can send the chatbot huge files to read and examine. After that, it can sum up parts of it or answer questions about it.

The model can also make logical replies of about 3,000 words, much longer than Claude 1.3’s limit of about 400 words. Because of this, Claude can write much longer responses all at once, like notes, emails, and stories.

The new Claude 2 chatbot has a lot of new features that make it better at code, math, and thinking than its predecessors. It is also trained in writing better code in organized data and languages like JSON, XML, YAML, and other markup languages. This makes it easier to change the style of data from one type to another.

Regarding basic coding skills, Claude 2 got 71.2% on the CodexHumanEval Python coding test, up from 56%. This shows that it has improved a lot. In math, it scored 88%, up from a score of 85.2% on the GSM8k, an extensive set of elementary school math questions.

In addition to making Claude 2 better, the Anthropic team has been working to make the model less harmful by improving its safety and making it harder to prompt it to generate hurtful or dangerous outputs.

The Anthropic team said, “We have an internal red-teaming evaluation that scores our models on a large representative set of harmful prompts, using an automated test while we also regularly check the results manually. In this evaluation, Claude 2 was 2x better at giving harmless responses compared to Claude 1.3.”

Anthropic came up with the idea of “Constitutional AI,” a learning model for AI that gives the AI system a set of values it should follow to help get these less adverse outcomes. The goal is to make it less dangerous or toxic by having another AI supervise its reactions and revise its own based on those values.

The company says it is currently working with thousands of businesses using the Claude API to power their AI apps. Salesforce Inc.’s Slack uses Claude to summarize talks and draft documents, and Zoom Video Communications Inc.’s contact center workers can answer customer questions more quickly with the help of Claude.

Sourcegraph, a code search and intelligence tool, uses Claude 2 in its AI-enabled coding assistant Cody, which lets developers prototype code and fix mistakes. It also uses the improvements done to the model and the larger context window. Hence, it can read bigger codebases at once and make larger code snippets. Cody knows about new frameworks because new data has been added to its knowledge base.

Quinn Slack, Co-founder and Chief Executive of Sourcegraph, said, “When it comes to AI coding, devs need fast and reliable access to context about their unique codebase and a powerful LLM with a large context window and strong general reasoning capabilities. The slowest and most frustrating parts of the dev workflow are becoming faster and more enjoyable. “

Anthropic’s free web-based chat is powered by Claude 2, but it’s only available in the U.S. and U.K. The company said it is trying to get it used worldwide in the next few months.