With Claude 2, Anthropic launches a direct competitor to ChatGPT. The chatbot offers similar features as the OpenAI chatbot, Google Bard or Bing Chat.
Anthropic was founded in 2021 by former OpenAI researchers, is funded with around 1.5 billion US dollars, and is seen as a direct competitor to OpenAI. The latter, of course, has a big head start with its hugely successful ChatGPT and Microsoft's billion-dollar backing. However, Anthropic's first model was able to hold its own quite well, at least in benchmarks.
Anthropic is now following up with Claude 2. According to the startup, improvements include better conversational skills, a clearer explanation of its reasoning, more harmless output, a longer memory, and better programming, math, and thinking skills.
The model can be used to write documents, memos, letters, stories, technical documentation, or books. An outstanding feature of Claude 2 is that it can process up to 75,000 words (100,000 tokens) at a time, which is significantly higher than the standard ChatGPT limit of 3,000 words.
This allows the chatbot to consider more context in its responses, improving response quality and task variety. The extra-large context window was already announced by Anthropic in May for the first model, Claude.
On the multiple-choice section of the American Bar Exam, Claude 2 is on par with GPT-4 at 76.5 percent. GPT-3.5, the free ChatGPT, averages only about 50 percent. On the Codex HumanEval Python programming test, Claude 2 achieves 71.2 percent (compared to 56.0 percent for Claude 1.3), and on the GSM8k elementary school math problems, 88.0 percent (compared to 85.2 percent for Claude 1.3).
Additional capabilities are on the horizon and will be rolled out slowly and iteratively over the coming months, according to Anthropic.
Claude 2 launches first in UK and U.S.
The new chatbot took about two months to develop. About 35 people worked directly on the AI model, with another 150 in supporting roles. In developing Claude 2, Anthropic says it paid particular attention to safety.
Unlike OpenAI, Anthropic relies on an AI-based feedback mechanism, rather than involving humans in evaluating the AI-generated content to optimize the model, and sets ground rules through a kind of constitution based on Apple's T&Cs, among other guidelines.
In red team testing, where the chatbot is deliberately provoked to make mistakes, Claude 2 achieved twice the pleasant user experience of its predecessor, according to Anthropic. Claude 2 is not free of hallucinations or misinformation. Anthropic says there are still many hurdles to overcome.
Anthropic says thousands of companies are already using the Claude 2 API, including partners such as Jasper, a generative AI platform for marketing copy, and Sourcegraph, a code AI platform that uses Claude's enhanced reasoning capabilities and larger context windows to help developers write, fix, and maintain code.
Business customers can get the Claude 2 API for the same price as Claude 1.3, and the web chatbot is available as a free beta in the US and UK.