Google’s Gemma2-2B, A Compression Marvel

Author(s): Ignacio de Gregorio

Originally published on Towards AI.

How to Build a Small Titan
Source: Generated by author using GPT-4o

Google has reached a new milestone in AI by training a 2 billion parameter model, absolutely minute in today’s terms, that surpasses ChatGPT-3.5, the first version of ChatGPT that came into our lives… despite being almost 90 times smaller.

It’s the first time we've seen such a small model with what might be the best performance-to-size ratio we've ever seen: a ChatGPT-level model that can be run on a consumer laptop.

And the reason behind this amazing achievement resides in a very particular yet elegant and unorthodox way of training LLMs.

Tired of pointless hype?

This piece, among many more weekly content, was first published in my newsletter, the place for AI Executives and Analysts who want to learn the truth behind the hype, spot trends, and take advantage of them.

The newsletter to stay ahead of the curve in AI

thetechoasis.beehiiv.com

When we think of Large Language Models, there’s only one thing we need to know: they are data compressors.

In other words, training them refers to embedding knowledge into their weights, so that the model can replicate it back.

In a nutshell, all LLMs do is, given a text sequence, provide a reasonable continuation that is very similar to the original sequence… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI