1. Motivation
Large language models (LLMs) have rapidly matured into sophisticated reasoning machines. Their ability to approximate, or even outperform, human responses across a broad spectrum of intellectual tasks has garnered the world’s attention. The success of these models raises an important question: Should the general public be able to access a model’s code, given its potential impact on society?

Currently, developers are divided on whether to release (i.e., open-source) or restrict access to the inner workings of their most capable models. Some developers favour open-sourcing, publishing every detail of their model so that anyone can utilise, scrutinise, and customise the code. Proponents of this approach argue that it fosters innovation and distributes the economic and epistemic value that their models generate. In contrast, some developers prefer to keep their code private, limiting access to a select group of trusted collaborators. They argue that withholding their code from the public domain is the only way to guarantee that it won’t be fine-tuned for malicious applications, from writing fake news to designing malware.

In this article, we examine the benefits and drawbacks of various strategies for AI dissemination. We start by considering the three main approaches to model release, and end by outlining some lesser-known tactics. The goal is to provide an introduction to the field of model dissemination, rather than a comprehensive survey. That being said, most models can be broadly categorised into one of the following three descriptions:

Closed models, like Gopher, are entirely inaccessible beyond the lab or company that produced them.
Queryable models, like Chat-GPT, can be prompted, with responses appearing on a user interface, but the code and weights are kept from the public domain.
Open-Source models, like GPT-J, make both the code and model weights publicly available.

2. Three Common Strategies for Model Dissemination

2.1 Closed Models
Closed models minimise the likelihood of misuse by restricting access to a small group of trusted individuals. This approach has been applied to control other dangerous technologies where legislator’s main concern is preventing misuse. For instance, countries that restrict firearm ownership tend to have lower rates of firearm-related homicides. On a more existential scale, nuclear non-prolif...

Posts

Wikitag Contributions

Comments