August 2023 Summaries

Meta Platforms, Inc. has introduced Code Llama, an advanced model derived from Llama 2, aimed at enhancing programming tasks with features such as infilling, large input context support, and zero-shot instruction following. Code Llama is accessible through Ollama, offering models with 7 billion, 13 billion, and 34 billion parameters, each with specific memory requirements. In addition to general code generation, Code Llama includes foundation models and specializations for Python, allowing users to efficiently perform code generation and completion tasks.

In May 2023, machine learning engineer Eric Hartford discussed the merits of uncensored AI models in his blog post "Uncensored Models," providing examples of how these models differ from their censored counterparts. Hartford highlights various uncensored models, such as the Llama 2 and Nous Research's Nous Hermes Llama 2, which are noted for their long responses and reduced hallucination rates. By comparing the output of censored and uncensored models, the text illustrates how the latter provides straightforward answers to questions, such as identifying who made Rose promise to never let go, offering a recipe for spicy mayo, and locating religious text references, while the censored models often refrain from providing direct responses under the guise of ethical considerations. The uncensored models, however, come with risks, and users are advised to exercise caution when using them, as they may not adhere to the same ethical guidelines as their censored counterparts.