AI researchers suggest open LLMs are not as open as claimed

A trio of AI researchers from Cornell University, Signal Foundation, and Now Institute have published a Perspective piece in the journal Nature, arguing that well-known open LLMs are not nearly as open as their makers claim.

In their paper, David Widder, Meredith Whittaker and Sarah West note that simply making source code available to the public does not make an LLM open. This is because it does not provide access to the underlying training data, and because very few developers have the resources needed to train LLMs independently.

Over the past few years, LLMs like ChatGPT have become very popular, and their popularity has only grown as they have matured. Along with such popularity has come fear—many people, both lay and professional, have started to wonder where AI research is going. Will it lead to loss of privacy? Jobs? Will it become impossible to tell if a picture or video is real, or generated by a neural network system?

Nobody knows the answers to such questions, but in response, LLM creators have tried to make their efforts more transparent to the user community by posting their models as freely open to the public. Anyone who chooses to go to the website of the maker of an LLM can look at or download the code. They can also change the code, and use it however they wish. But, the authors of this new paper ask, do such actions truly make an LLM open?

They strongly suggest the answer is no, because the source code for an LLM is not the same as the source code for a computer program, such as a word processor. When you download the code for a word processor, you have all that is needed to use it as it is, change it, or do whatever you wish.

When you download an LLM, you have the code, and you can modify it if you want, but you cannot modify the underlying knowledge that comes with it, the authors note. That came about due to training done by the maker. Users do not get that when they download the code, and most cannot run their own training regimen—it takes massive amounts of computing power.

Additionally, all the current open LLMs have three main factors impacting openness, the authors suggest—the first is transparency. Some makers make everything transparent, and others do not. The makers of Llama 3, for example, only allow users to use their system via Application Programming Interfaces (APIs). The authors call such practices “openwashing.”

The second factor is reusability—how useable is the open-source code? That depends on how it was written. The third factor is extensibility, which is how users can alter the code to meet their needs.

The authors conclude by suggesting that until users have open access to hardware capable of training LLMs, data easily accessed and/or free access to the underlying data used to initially train an LLM, open LLMs will not truly be open.

© 2024 Science X Network