Claude, a new AI chatbot is built by ex

How Anthropic built Claude

Anthropic was founded in 2021 by people who left OpenAI, which included the lead engineer for GPT-3, Tom Brown. So, there are many similarities in the technology that was used to train it. However, the founders claim that they have used a list of rules and principles to train the AI model, which they refer to as Constitutional AI.

Interesting Engineeringhas previously reported how users managed to break into Microsoft's Bing chatbot's initial rules and made it reveal them. Such behavior was something that even made the chatbot a bit defensive and irritated later on.

Claude, as Anthropic claims, has been trained to be helpful and politely refuse to reply to questions that might be harmful. According to a webpage on the company's website, it achieved this by further training the chatbot during its reinforcement learning (RL) phase with another preference model to make the chatbot harmless but also non-evasive.

To test Claude's effectiveness, Anthropic offered companies access to the chatbot. Among them, DuckDuckGo is using the technology to launch a search tool for its users. Google has also found the technology quite useful and invested $300 million in Anthropic late last year.

Now, Anthropic has launched Claude in two variants, Standard and Instant, with the latter offering a faster, lighter, and cheaper way to integrate AI into products. Anthropic suggests that the chatbot could be put to task across search, customer service, sales, as well as back-office work to begin with.

Claude is designed to be self-contained and has no internet access. For improvements, prompts are the best way to teach Claude what is required out of a task. Additionally, it can also be assigned a personality and a tone, something even Microsoft is working into its chatbot.

It will be interesting to see a new chatbot enter the fray and see how Microsoft plans to tackle it.