Llama3 Is Impressive and Small
I can see why people hate it
Often called the plagiarism machine, LLM’s like ChatGPT become popular because it’s ability to generate text that resembles what a person could output, content is clearly built from referencing others works. OpenAI even has admitted that they scraped the web for content. I’m not sure if there is a way out of this. But I’m hopeful.
At least with local LLM:
- It uses less power The power pull that my machine uses is equivalent to playing a light game. These new models are small but very compressed. Training takes a lot of power still but I’m certain that will become more manageable soon.
- Private LLM Because the model runs locally and doesn’t require access to the web it can be incredibly private. Companies like OpenAI collect your input as more training data.
The Three Models sizes
The initial release of LLaMA-3 includes two primary variants: a smaller model with 8 billion parameters and a larger one with 70 billion parameters. These models are designed to support a wide array of applications, showcasing state-of-the-art performance across various industry benchmarks. As of the latest update, an even more extensive model with over 400 billion parameters is currently in the training phase. I will only be running the 8b model, because I’m not a rich weirdo, just a regular werido.
Models like LLaMA-3 for Tasks
I’m rarely using Llama 3 to simply write content from prompts but actually using the model to get tasks done from my work.
I already use Whisper for generating captions and transcripts, so it’s easy to take those into my workflow. I can drop the transcript of my videos into the document function if GPT4ALL and use it to generate keywords, meta descriptions, and it can be used as a spell and grammar tool.
Llama 3 in all its glory (or not). I get why people hate the idea of AI models like this one being able to generate text so easily, but at least for me, having these tools locally means I can keep my data private and still get stuff done. So for now, I’m mostly impressed.