Meta's new AI model, Llama 2, is a breakthrough in natural language processing, but the company is not sharing it entirely with the public. In fact, Meta is joining the likes of OpenAI in claiming its AI is open source when that may not be the whole story, a recent study has shown.
Meta, formerly known as Facebook, recently unveiled its latest AI model, Llama 2, which can generate coherent and diverse texts on any topic, given a few words or sentences as input. The model is based on a massive neural network with 175 billion parameters.
Large language models (LLMs) are a type of artificial intelligence (AI) that are trained on massive datasets of text and code. They can be used for a variety of tasks, such as generating text, translating languages, and writing different kinds of creative content.
Some LLMs are open-source, which means that the code that was used to train them is available to the public. This allows developers to use the code to create their own LLMs, or to improve the performance of existing LLMs.
Major LLMs are Not as Open As they Seem
However, not all LLMs are open-source. Some companies, such as Meta and OpenAI, have released LLMs that are not as open as you may think. This means that the code that was used to train these LLMs is not available to the public.
In a recent paper, a group of AI researchers from Radboud University in Nijmegen, Netherlands, argued that the term “open-source” is being used misleadingly by some companies. They point out that some generative AI LLMs that are labeled as “open-source” are not actually open-source, because the code that was used to train them is not available to the public.
Meta's Llama 2 and the GPT/codex model that underpins OpenAI's ChatGPT are two notable examples. The researchers argue that the lack of open-source LLMs is a problem for the AI community. They call on companies to release more open-source LLMs, so that researchers and developers can have access to the code and improve the performance of these models.
OpenAI's ChatGPT model is the most secretive of all the models that the study team has evaluated so far. It fails to meet any of the openness criteria, except for two: it has a model card that explains what it can and can't do, and a preprint that details its research. But even for these two, it only gets a half mark. Llama 2 is slightly more transparent than ChatGPT, but not by much. It still ranks very low on the openness scale.
Blocking Parts While Claiming Openness
There are a few reasons why companies might choose to release LLMs that are not open-source. One reason is that they may want to protect their intellectual property. Another reason is that they may want to control how the LLMs are used. The lack of open-source LLMs can be a problem for researchers and developers. It can make it difficult to reproduce the results of research papers, and it can make it difficult to improve the performance of existing LLMs.
Meta's position is ironic as the company proudly says that Llama 2 is available for research. Even so, it seems there are parts of the project that the company is locking away from developers and researchers. While OpenAI claims openness – it is right there in the name – the company closed much of the source code to GPT-4 when it was launched in March.
Speaking to Speaking to AI research and podcaster Lex Fridman, Meta CEO Mark Zuckerberg spoke of the company's apparent commitment to AI openness. At the same time, he did admit that the company would keep some technology behind closed doors:
“You know responsibility and getting safety right on these [AI Models] is very important. For the first one, there were a bunch of questions around whether we should be releasing this commercially. So, we kind of punched it in on that for V1 of LLaMA and just released it for research.”