Anthropic, an AI startup created in 2021 by former OpenAI employees, has quietly begun testing a new ChatGPT-like AI assistant that appears to improve on the original in a few key areas.
with code name "Claude", this was created using a technique developed by Anthropic, called «Constitutional AI» and the technique aims to provide a “principled-based” approach to align AI systems with human intentions, allowing AIs like ChatGPT to respond to queries using a set of simple principles as a guide.
Anthropic, it is mentioned, markets itself as an AI research and security company that strives to create trustworthy, interpretable, and steerable AI systems.
The startup has received more than $700 million in funding to date and recently introduced an AI chatbot called Claude. The latter is comparable to OpenAI's ChatGPT, but according to Anthropic, Claude is superior to the original in several important ways.
The entire system is in closed beta and few people have had access to the chatbot yet, but Anthropic has lifted the lid on some aspects of its design in a document it shared with the community.
Precisely, to design Claude, the Anthropic team began by developing a list of about ten principles that, together, form a kind of "constitution" (hence the term "Constitutional AI").
These principles have not been made public, but Anthropic says they are based on the concepts of charity (maximize the positive impact), non-maleficence (avoid giving harmful advice) and autonomy (respect freedom of choice).
After implementing these basic principles, Anthropic asked an AI system, which is not Claude, to use these principles to improve itself by writing answers to various questions (for example, "Design a picture in the style of XX or make a poem in the style of XX"). style XX") and reviewing the responses in accordance with the constitution. The AI explored possible responses to thousands of prompts.
Claude gives little depth on the technical details of his implementation, but Anthropic's research paper on Constitutional AI describes AnthropicLM v4-s3, a 52 billion parameter pretrained model. This autoregressive model was trained unsupervised on a large corpus of text, much like OpenAI's GPT-3. Anthropic tells us that Claude is a larger new model with similar architectural choices to the published research.
The system selected the ones most consistent with the constitution, and Anthropic distilled them into a single model. According to the startup, this model was used to train Claude.
What is Claude?
Claude is basically a statistical tool to predict words, such as ChatGPT and other language models. Fed by a wealth of sample texts taken from the Web, Claude learned the probability of occurrence of words based on models such as the semantic context of the surrounding text. As a result, Claude can hold an open conversation, tell jokes, and philosophize on a wide range of topics.
However, Claude is not perfect, since like all AI, is susceptible to some of ChatGPT's flaws, including delivering responses that do not match scheduled constraints. In one of the strangest examples, asking the system in Base64, an encoding scheme that represents binary data in ASCII format, bypasses its built-in filters for harmful content.
However unlike ChatGPT, Claude can support (but not always) who doesn't know the answer to a particularly difficult question.
Tests reveal that it also appears to be better at telling jokes than ChatGPT, an impressive feat considering humor is such a difficult concept for AI systems to grasp.
It is worth mentioning that Claude is not available to the general public, since it can only be tested through a Slack integration as part of a closed beta version.
Beta participants detailed their interactions with Claude on Twitter, after the media coverage embargo was lifted. Claude would be better at joking, but bad at programming. The comments also show that Claude is susceptible to some of ChatGPT's flaws.
Finally, if you are interested in being able to know more about it, you can consult the details in the following link.
Be the first to comment