Less than a year ago, very few people could tell you what a “large language model” was, and “chatbots” were largely confined to the customer support pages of company websites. Thanks to ChatGPT’s extraordinary success, they’re now everywhere, and keeping track of the various runners and riders in the Chatbot Derby is becoming a full time job in itself. ChatGPT and Bing Chat are from the same stable, but have different strengths and weaknesses, while rival bots from Google’s Bard to Inflection’s Pi lead the pack in certain areas, but lag far behind in others. 

‍

One chatbot company many have overlooked is Anthropic, formed in 2021 by former OpenAI staffers who had worked on OpenAI’s GPT-2 and GPT-3 models. Their AI, Claude - intended as a direct competitor to ChatGPT - began life as an unassuming imitator with few distinguishing features, but the recent release of the Claude 2.0 update has seen this dark horse join the front of the pack. The new features are so impressive that Amazon has just laid out $4bn for a minority stake in Anthropic, so it seems likely that Claude’s days out of the spotlight are over; we will definitely be hearing a lot more about this AI in the future, so there’s never been a better time to get acquainted. So, what makes it stand out, and how can you get the most out of it? Read on! 

‍

THE BOOK EATER WITH AN AMAZING MEMORY

There’s one area where Claude completely outclasses the competition: its “context window”. Loosely speaking, this is effectively the chatbot’s memory, the amount of information it can remember and process at any given time. In AI, this is measured in “tokens”, and Claude’s context window is 100,000 tokens, which is roughly equivalent to 75,000 words, or 340,000 characters. For comparison, this is slightly more than three times GPT-4’s capacity (32,768 tokens), and six times that of GPT-3.5 (16,000 tokens).

‍

‍

This has many benefits in practical use. The most obvious is that you can feed Claude an entire medium-sized book - 100k tokens is enough for the first Harry Potter novel, or The Hobbit - and it will digest its contents in seconds. This feature becomes more practical when used for non-fiction purposes, and Claude’s gigantic memory means it’s extremely adept at chewing through large amounts of data, making it a powerful tool for workplace tasks. Even better, this expanded context window gives Claude a second huge edge over its rivals; it will remember the entirety of your conversation in a chat window, and does not need to be constantly reminded of things it said two replies earlier. This enables far more constructive and precise interactions, and allows the user to devote more time to fine-tuning their prompts, rather than merely repeating them. 

‍

‍

THE HARMLESS BOT WHO ADMITS IGNORANCE

While all this capacity is great in theory, the context window is just one part of what makes a chatbot tick. Fortunately, Claude excels in other areas, and has been trained by Anthropic to be a “helpful, honest, and harmless assistant with a conversational tone”. Much has been made about the “harmless” part of that description - Anthropic’s founders left OpenAI because they were concerned about the direction the firm was taking - but in truth the real key word here is “honest”. 

‍

Claude is not immune to hallucination, but it is a lot less prone to it than many of its rivals. While Google Bard now has a (phenomenally useful) “double-check button” to validate its psychedelic responses, Claude prefers to stick to facts whenever possible. This is highly unusual for chatbots, whose desire to please the user tends to outstrip their desire for accuracy. Nonetheless, you should still be careful to avoid taking Claude’s responses at face value, and also take note of Anthropic’s excellent advice on how to keep hallucinations to a minimum. These include ‘ask Claude multiple times’, ‘ask Claude for direct quotes’, and ‘Give Claude a "way out" if it doesn't know the answer’, by explicitly giving it permission to admit it doesn’t know things. We wish this worked with other chatbots!

‍

CLAUDE’S FLAWS

Nobody’s perfect, of course, and Claude has several limitations. At the time of publication, the main area where it is lagging behind is multimodality - interpreting images and audio. Claude can accept attachments of text files and PDFs, but anything that isn’t text-based is beyond its comprehension. It also cannot access the internet directly, but that generous context window helps mitigate this issue; you can simply paste the text of a webpage (or the  source code) directly into Claude. 

‍

‍

Other limitations to bear in mind: Claude is currently only available in the US and UK, and Anthropic says it is for over-18s only. As with ChatGPT, usage is capped, although it’s hard to get a precise idea of exactly what that cap is, not least because Claude will flat out deny any usage cap exists when asked for details. Fortunately, you will get an on-screen warning when you approach the limit (see screenshot), but it’s easy to miss. There is a paid subscription plan for Claude, but this merely increases the limit rather than removing it; at the time of publication we believe free users get 20 prompts every 8 hours, and subscribers get 100 prompts over the same period, but this is subject to change. 

‍

Once again, Claude’s generous memory goes a long way to help reduce the issues caused by the usage cap; even if you only make use of half the context window, those 20 (or 100!) prompts should go a very long way. This means Claude is particularly suited to professional tasks that involve processing large amounts of data, and you will find many use cases within Anthropic’s documentation illustrating the bot’s tremendous power in these areas. Given Claude’s sudden appearance as a frontrunner in the Chatbot Derby, it’s an excellent time to acquaint yourselves with this thoroughbred’s form!

‍