Public announcements or reports from major tech companies.

Meta's Llama 3 AI Model Shows Advanced Capabilities, Struggles with Quantization

May 12, 2024, 04:32 PM

Meta's recent launch of Llama 3, an advanced language model, has sparked significant interest in the AI community. Llama 3, notable for being trained on a record 15 trillion tokens, demonstrates a nuanced understanding of data relationships, which becomes evident in its performance degradation when quantized compared to its predecessor, Llama 2. The model's sensitivity is attributed to its use of BF16 precision, which captures even the minutest decimals. Specifically, the LLAMA3-8B model under LoRA-FT quantization shows that low-rank finetuning on the Alpaca dataset cannot compensate for the errors introduced by quantization. Additionally, Llama 3, which offers models like the 8B and 70B, is being integrated into platforms such as Promptitude for enhanced language processing and task management. The model's development was highlighted during a 24-hour hackathon involving over 500 AI engineers, showcasing its capabilities and potential applications in real-world scenarios.