MythoTech AI News: 4/18/24
The latest in AI news - Llama 3 is finally here!
LLMSAI NEWS
Very excited to see Meta's Llama 3 finally release! 🦙🦙🦙
Check it out here: https://llama.meta.com/llama3/
Initial reaction.... the Benchmarks 🤯 - but the Context Length 😞
8k context is straight bad for a model released today. This would put it at the very bottom of the rankings in my LLM Factsheet (https://mythotech.ai/llm-api-factsheet), tied with an old version of GPT-4 from a year ago.
We will see how it plays out in time, but my guess is that Llama 3 and its derivatives will be a forgetful genius, incredibly useful for single and straightforward tasks (like classification or summarization) but outclassed in long-form use cases (like complex coding or advanced chatbots).
Still, that kind of power on a small 8b model that can be run fully locally on your average PC is a huge step. Did you know that Mistral-Instruct-7b can already beat out GPT-4 when it is fine-tuned for a specific task?
Llama 3 represents a 15-30% leap in that open-source, locally hosted power.
