Anthropic, the AI lab created by OpenAI expats, launched its own ChatGPT competitor, Claude, last March. Today, Anthropic is introducing Claude 2, its next-generation AI chatbot, which is said to be faster, smarter at handling longer queries, and better at articulating its responses.
The San Francisco-based AI lab says Claude can now avoid harmful content more effectively. Anthropic’s development process included training the Claude model on rules inspired by the United Nations Universal Declaration of Human Rights. For an AI, this “moral constitution” is nothing more than a set of parameters, but it makes all the difference in making the chatbot more empathetic and more human, and less mechanical.
Anthropic describes this effort in a research paper titled “Constitutional AI: Harmlessness from AI Feedback”. Another notable aspect of this approach is that Claude is supposed to self-improve his behavior and adapt to instances of misbehavior without requiring human intervention.
According to Anthropic’s internal tests, Claude 2 would be twice as effective at avoiding harmful responses. The only downside for now is that Claude 2 is limited to the US and UK, but Anthropic plans to expand it globally in the coming months.
An alternative to ChatGPT that relies on security
Another notable development, Claude 2 is now accessible via its own website. In order to broaden its reach, especially with enterprises, the APIs of the second generation of the Claude model have also been released.
With Claude 2, Anthropic is looking to deliver an update on the same scale that OpenAI got with the GPT-4 update for ChatGPT. Claude 2 is said to offer longer memory, and is also better at solving math, code-aligned, and reasoning-signed queries.
Anthropic claims that Claude 2 scored 76.5% correct on multiple-choice questions on the bar exam, which is a significant jump of over 73% from the Claude 1.3 model.
The company says users can now enter queries worth 100,000 tokens. To put that into perspective, ChatGPT can only process tokens up to a maximum value of 8,000 tokens. In linguistic terms, approximately 2,000 tokens sent to Claude correspond to an essay of 1,500 words. Anthropic updated the tokens in May 2023, noting that Claude could now process an entire novel in less than a minute.
For users, this means that Claude 2 can now easily process long documents like letters and stories, while also being able to condense long documents into concise forms.
Claude 2 seems to seduce companies
Anthropic’s approach seems to resonate with companies. The startup works with “thousands of businesses” that use Claude’s API, including productivity companies like Slack and Notion. She said the 100,000 token (i.e. the amount of information you can enter) popup in Claude 2 enables new use cases such as summarizing long conversations or writing memos and opinion pieces.
In an industry dominated by major players like OpenAI, Anthropic is gaining traction by focusing on developing responsible, transparent, and easy-to-use AI solutions.
This article is originally published on blog-nouvelles-technologies.fr
