Why Meta is allowing users to see the inner workings of its new AI chatbot

economictimes.indiatimes.com:

Mark Zuckerberg's Meta recently unveiled its Llama 2 chatbot. Microsoft has been appointed as Meta's preferred partner on Llama 2, which will be available through the Windows operating system. Meta's approach with Llama 2 contrasts with that of the company OpenAI, which created the AI chatbot ChatGPT.

That's because Meta has made its product open source — meaning that the original code is freely available, allowing it to be researched and modified. This strategy has sparked a vast wave of discussions. Will it foster greater public scrutiny and regulation of large language models (LLMs) — the technology that underlies AI chatbots such as Llama 2 and ChatGPT? Could it inadvertently empower criminals to use the technology to help them carry out phishing attacks or develop malware? And could the move help Meta gain an advantage over OpenAI and Google in this fast-moving field? Whatever happens, this strategic move looks set to reshape the current landscape of generative AI.

In February 2023, Meta released its first version of the LLM, called Llama, but made it available for academic use only. Its updated version, Llama 2, features improved performance and is more suitable for business use. Like other AI chatbots, Llama 2 had to be trained using online data.

Exposure to this vast resource of information helps it improve what it does — providing users with useful responses to their questions. An initial version of Llama 2 was created through «supervised fine-tuning», a technique that uses high-quality question-and-answer data to calibrate it for use by the public. It was further refined with human feedback reinforcement learning which, as the name suggests, incorporates people's assessments of the AI's performance to align it