Chatbots for lawyers

ChatGPT-4, a generative large language model, has become notorious for opening a kind of Pandora's box, with BigTech companies racing to outperform each other, and new large language models (LLM, an abbreviation which exclusively meant Master of Law just a couple years ago) popping out at the speed of a bullet train – Phi, Mistral, Gemini, Grok, Claude, Llama to name the few.

As they grow in size exponentially, so do their capabilities.

Llama - the LLM developed by Meta* - is an example. Recently Llama3, the third generation of Llama LLM, was made available to the public.

The biggest Llama3 LLM has 70 billion parameters, was pre-trained and instruction-tuned based on a mix of publicly available data, and has s context length of 8 thousand tokens.

By the way, can you recall last eight thousand words, you’ve just said? Just to compare, "The Little Prince" by Antoine de Saint-Exupéry approximately is of this size.

In an effort to understand its limitations, the Llama3-70B instruction-tuned model, which is intended for assistant-like chat, was given a task to improve a clause from an EPCM contract (Engineering, Procurement and Construction Management).

How the model performed read at my full blog-post at Melling, Voitishkin and Partners.