Claude arrives in his second version, much more accomplished. But by the way, who is Claude 2? It is a chatbot model specialized in text generation. Based on a constitutional AI, it has not learned in the same way as its competitors and is more efficient in certain areas.
Claude was created by Anthropic, a start-up based in San Francisco and created by former executives of OpenAI, the creator of ChatGPT. The company recently announced the arrival of Claude 2, which will be available in beta in the United Kingdom and the United States. Compared to Claude 1.3, its predecessor, it has been greatly improved, in particular on two aspects: performance and ethics.
A performance boost in several key areas
The chatbot Claude, like any of its compatriots, performs its tasks by tapping into billions of different pieces of data. On the other hand, it is much faster than the other chatbots available on the market. The heart of the operation of Claude 2 and the reasons for its speed are not yet clearly explained. We know, however, that it was trained on a huge corpus of texts, much larger than that of ChatGPT-3.
Claude 2 has been evaluated in several areas to compare its results to those of Claude 1.3. According to the website TechCrunch, Claude 2 did better than his ancestor in 4 fields of study: law, medicine, programming and mathematics. If we establish an average on all the tests carried out by the new version of Claude, this one is about 17.8% better than its elder. A very good progress.
Anthropic also doubled down on Claude 2 to improve his reasoning skills and develop self-awareness. This requires training using more recent data from more diverse sources. Now, Claude 2 processes instructions better and faster, while keeping a form of primary awareness of its limitations.
Constitutional AI: an innovative ethical approach
To differentiate themselves from the competition, Anthropic took a different approach to developing their baby. Claude 2 was created using what the company calls “constitutional AI”. This means that Claude is guided by a set of ethical principles to generate his textual answers. According to the society, these principles include self-reliance and beneficence, among others. Even if these data are not completely public, Claude 2 is based on 10 major ethical pillars which allow it to be more honest in its answers. This approach would theoretically act as a filter and limit the misuse that could be made of a chatbot.
However, all is not rosy for Claude 2, Anthropic readily admits it. Limits still persist, and it still sometimes generates inappropriate, even toxic responses. Nevertheless, the company claims that Claude 2 has half the chance of generating harmful content compared to Claude 1.3. Statement that is completely impossible for us to verify factually as users.
This new approach to chatbot learning could lead Claude 2 to make a name for itself quickly in the middle of the AI scene. It is fast and very efficient, but the fact remains that overflows are still possible. Apparently, an ethical approach is not enough to erase this limit by magic.
Sources: TechCrunch, LeBigData
4