This AI learns to distinguish between good and evil

As artificial intelligence (AI) often produces fictional and offensive content, Anthropic, a company run by former OpenAI researchers, is charting a different path: creating AI capable of knowing what’s right and wrong with minimal human intervention.

El chatbot de anthropic, ClaudeShaped by a unique „constitution,” a set of rules inspired by the Universal Declaration of Human Rights, designed to ensure ethical behavior with robust functionality, along with other „ethical” standards such as Apple’s rules for app developers.

However, the concept of „constitution” may be more metaphorical than literal. said Jared Kaplan, a former OpenAI consultant and one of the founders of Anthropic Wired Claude’s constitution can be interpreted as a set of training parameters that any trainer can model his AI on. This implies different considerations for the model, aligning its behavior more closely with its constitution and discouraging actions deemed problematic.

The anthropological training system is described in a Inquiry Essay Titled „Constitutional AI: Safety from AI Feedback,” it outlines a way to create „harmless” but effective AI that, once trained, can self-improve without human input, capable of recognizing inappropriate behavior and changing its own behavior.

„Thanks to constitutional AI and food safety training, you can trust Cloud to represent your company and your needs,” the company said in its statement. Official website. „Claude is trained to deal with even snobbish or unpleasant conversation partners with grace.”

Notably, Claude can handle over 100,000 data tokens, far more than ChatGPT, Bard, or any other capable large-language model or AI chatbot.

In the field of AI, a “Token„Typically representing a piece of data, such as a word or character, the model behaves like a discrete unit. Claude’s tokenization capabilities allow him to handle long conversations and complex tasks, making him a formidable presence in the AI ​​landscape. In context, you can easily render an entire book, and He knows what to do.

READ  Who should not eat rice

AI and the relativism of good and evil

Concerns about ethics in AI are pressing, but this is a nuanced and subjective area. Norms interpreted by AI practitioners can limit the model if those rules are inconsistent with broader social norms. An overestimation of the trainer’s personal perception of „good” or „bad” can limit the AI’s ability to generate powerful and unbiased responses.

This topic is widespread discussed Among AI enthusiasts, both are praised They criticize (Depending on your own biases) OpenAI’s intervention in your own model is an attempt to make it more politically correct. But as paradoxical as it may seem, an AI must be trained using non-ethical information to distinguish what is ethical and what is not. If AI knows those data points, humans do Inevitably They will find a way to jailbreak the system, circumvent those restrictions, and achieve results that AI practitioners have tried to avoid.

Implementation of Claude’s ethical framework is experimental. OpenAI ChatGPT, which aims to prevent unethical referrals, has yielded mixed results. However, as Anthropic has demonstrated, the effort to directly address the ethical misuse of chatbots is a significant advance in the AI ​​industry.

Claude’s ethics training encourages him to choose responses that are consistent with his constitution, focusing on supporting freedom, equality, fraternity, and respect for individual rights. But can an AI consistently choose ethical responses? Kaplan Greece Technology is much more advanced than many expected. „It works very simply,” he said at the Stanford MLSys symposium last week. „This innocuousness will improve as the process progresses.”

READ  Can Blockchain Technology Help Against Corruption?

Utility rate for innocuousness of the model using constitutional AI (gray) and standard methods (colors). Image: Anthropology

Anthropics clarifies that AI development is not just a technology race; It was a philosophical journey. It’s not just about building a „smarter” AI, but for the researchers on the front lines, it’s about building one that better understands the line between right and wrong.

Interested in learning more about AI? Check out our latest Decrypt U course „Introduction to AI”. It covers everything from history of AI to machine learning, ChatGPT, ChainGPT and more. Learn more here.

Dodaj komentarz

Twój adres e-mail nie zostanie opublikowany. Wymagane pola są oznaczone *