A naukri.com initiative
Analyticsindiamag
2M
114
Image Credit: Analyticsindiamag
Meta Likely to Release Llama 4 Early Next Year, Pushing Towards Advanced Machine Intelligence (AMI)
- Meta is exploring ways to make its Llama models more advanced, with designs focusing on evaluating decisions and planning details in real-time to manage real-world scenarios.
- Meta’s Manohar Paluri said in an interview that Llama would include self-supervised learning (SSL), a combination of versatile and task-specific models involving both planning and evaluating models.
- Meta will employ Chain of Thought to improve Llama's ability to dynamically adapt by skillfully breaking down complex tasks.
- The release of Llama 4 by Meta is imminent, with the company already pre-training the model, CEO Mark Zuckerberg confirmed in interview.
- The model will include advanced machine intelligence that can apply perception, reasoning and planning in an integrated way, and the team has hinted at a release date of Q1 2022.
- Meta claims that Llama's major strength comes from its self-supervised learning, a process Paluri described as a model's ability to pick up general knowledge from vast data autonomously.
- Meta AI's chief, Yann LeCun, told AIM that advanced machine intelligence is essential if AI is to help people in daily life, making systems that can understand cause and effect, and model the physical world more valuable.
- Meta aims to blend SSL with RLHF (reinforcement learning with human feedback) models to create Llama models that can provide broader and more flexible general knowledge, as well as more intensive and defined task-specific CPR support.
- To enhance Llama's performance on a device, Meta recently integrates the Quantisation of LLMs, enhancing on-device AI performance with up to four times faster inference speeds, a 56% reduction in the model size.
- The frequent iteration of Llama's model upgrades is aimed at ensuring predictable, substantial improvements with every release.
Read Full Article
6 Likes
For uninterrupted reading, download the app