Anthropic's Claude AI Chatbot Is Now Nearly 2X More Capable Than ChatGPT
Token length is the amount of information you can relay to the AI. It describes the maximum length of a prompt that you can give to the network for it to mangle and produce output data. That's what these models do, after all—take input data, interpret it according to the weights from their training, and then output infered information which is converted to sensible output data.
"Tokens" are a fundamental unit of text to be processed by the model, and it can be a word, a part of a word, a punctuation mark, or even a whitespace character. The original GPT-4 had a maximum token length of 8,192 tokens, which was considered impressive at the time of its release. The recently-released GPT-4 Turbo has a maximum token length of 128,000 tokens, but even this huge number is dwarfed by Anthropic's new Claude 2.1 that can accept 200,000 tokens in a single query.
If you need an AI to analyze technical documentation, entire codebases, huge financial statements, or long literary works, you're going to need a gigantic context window, and the bigger, the better. The context window is the amount of data that the AI can process in a single query, and is synonymous with the maximum token length for most purposes. Anthropic says that Claude's 200,000-token context window is an industry first.
Having such a huge context window certainly has its advantages, but it also has downsides. The primary advantages are that it gives Claude immense context when doing things like summarizing long documents, answering questions about huge swathes of text, or generating code that fits in with an existing codebase. However, it also slows down the process immensely.
Anthropic says that "tasks that would typically require hours of human effort to complete may take Claude a few minutes." That's amazing, but "a few minutes" is quite a long time to wait for an AI response relative to its competitors. Of course, its competitors have much smaller context windows, but that's the trade-off.
Besides the gigantic context window, Claude 2.1 has decreased its hallucination rates by half. Anthropic says that the model is much more likely to demur on a topic ("I'm not sure") instead of confidently giving a wrong, hallucinated answer. The company also says that they've reduced the rate of comprehension mistakes by "3-4x", radically improving the model's ability to comprehend and summarize documents.
Furthermore, Claude has a beta feature that allows users to integrate the model with their own existing processes, products, and APIs. This is a pretty exciting feature for Anthropic, as it allows Claude to draw on external tools for advanced reasoning, like using a calculator for complex maths or searching a database for information related to user queries. It also allows users to make use of Claude as a recommendation or search engine.
Anthropic says that tool use is "currently in early development," which is fair enough. If you'd like to get started with Claude, head over to the AI's website.
