9621 Agnes Crossing, Lake Suzanneview, New Mexico Island 84604-9295.
The new version of Metas Llama-LLM may be released in July. The biggest adjustment will be how the model handles questions it deems “controversial.” Mark Zuckerberg announced in
The new version of Metas Llama-LLM may be released in July. The biggest adjustment will be how the model handles questions it deems “controversial.”
Mark Zuckerberg announced in January that his company Meta had begun training Llama 3, the third generation of the LLM. According to The Information, this will launch to a wider audience in July, although Meta is yet to confirm this. Meta uses its language model in its own social media platforms, among other things, to drive generative AI functions.
According to The Information, the biggest adjustment to Llama 3 will again be made to the handling of “controversial” prompts. To prevent the AI model from saying things that could damage Meta’s reputation, the company built guardrails into the model. The company’s researchers themselves would consider these security mechanisms excessive, since Llama also refuses to answer questions that are not at all controversial.
For example, Lama will not tell you how to win a war just to set an example, but will also give you tips on how to prank a friend, which, in the model’s opinion, goes too far. An illustrative example from The Information is that, for example, the current version of Llama also does not respond to the request turn off an engine. The model doesn’t seem to understand that in this context you’re asking about how to turn off an engine and not literally kill something, so Llama plays it safe.
Lama 3 should be able to understand the context of such “controversial” questions. Of course, that still doesn’t mean you can make the model say everything. Meta also wants to make the model’s responses more nuanced and will appoint someone internally to oversee Llama 3’s training.
Meta’s cautious attitude towards his LLama-LLM is understandable. LLMs have gone off the rails before. Think about the early days of Microsoft’s Copilot, when it was still called Bing Chat and at times gave very strange answers, from arguments with users about the factual accuracy of the answers to declarations of love.
The most recent example of AI derailment comes from Google. The tech giant decided to temporarily stop allowing Gemini to take pictures of people. The model was trained to represent diversity in AI-generated images, but in practice this mainly resulted in historical inaccuracies.
Source: IT Daily
As an experienced journalist and author, Mary has been reporting on the latest news and trends for over 5 years. With a passion for uncovering the stories behind the headlines, Mary has earned a reputation as a trusted voice in the world of journalism. Her writing style is insightful, engaging and thought-provoking, as she takes a deep dive into the most pressing issues of our time.
