702: Llama 2 — It's Time to Upgrade your Open-Source LLM — with Jon Krohn (@JonKrohnLearns)

Time Awareness

Llama 2 introduces a groundbreaking feature called time awareness, allowing it to adapt its responses based on temporal context. Jon Krohn explains that this feature enables the model to provide historically accurate answers, such as acknowledging the earth as flat in the year 800, while recognizing it as round in 2023 1. This advancement, coupled with a doubled context window of 4000 tokens, significantly enhances the model's ability to handle complex queries and diverse documents.

   

Reinforcement Learning

Llama 2's chat capabilities are powered by a two-stage reinforcement learning from human feedback (RLHF) process. This process includes rejection sampling and proximal policy optimization, enhancing its generative capacity beyond other open-source models 2. Additionally, the model employs ghost attention, which improves its ability to maintain context in multi-turn conversations, allowing for creative interactions like responding solely with emojis.

   

Safety & Investment

The development of Llama 2 involved a substantial investment of $25 million, emphasizing its commitment to safety and alignment testing. Jon Krohn highlights that these efforts have resulted in AI safety violation percentages lower than any other open-source LLM, even surpassing ChatGPT in some metrics 2. This rigorous testing ensures that Llama 2 not only excels in performance but also adheres to high safety standards.