What is OSS?
When we speak about Open Source Software, we are typically speaking about code which has been released under or covered by an Open Source License. These licenses are licenses granted under the copyrights of the licensor. They differ from traditional proprietary copyright licenses in several fundamental ways, most importantly:
- transparency over the source code allowing for research, testing and reproduction
- unrestricted distribution, use and modification as long as license terms are respected
- no discrimination between types of users, industries, technological applications
Some OSS licenses are not so business friendly
It is commonly known that there are OSS license types that are inimical to the corporate tech companies who might be interested in cloning or forking the OSS for their own commercial gain. This is because of the terms in those licenses that require all re-distributions or instances based on the licensed code (derivative works) to be released under the same underlying OSS license, including release of all relevant source code. In other words, certain licenses such as the GPL family of licenses by perpetuating the above conditions for all downstream releases, tend to warn off would-be commercial enterprises wanting to use the Open Source code as the basis of a custom-packaged proprietary application.
The OSS trade off
Outside of some well established OSS projects such as the Linux Core operating system, many OSS projects would struggle to gain the needed traction from the corporate user community to bring their code to the wider market if the choice of license causes these very users to self select out of these projects i.e. not want to contribute improvements to making the code a de facto standard that they can build on and incorporate as the base layer of their own customized solution. Enter the ‘permissive’ form of OSS license that gives these companies the freedom they need to do just that while remaining respectful to the underlying license.
The many varieties of permissive
Permissive licenses come in different shapes and sizes. Many of them, notably the MIT license, are rather spare. Apart from requiring re-distributors of the software to give credit to the authors and maintain warranty exclusions, there are few if any obligations imposed on the licensee. The Apache license (latest Apache 2.0) is the outlier. This license provides protection to users and licensees in the case that the licensor, other contributors of code to the project or any licensee should try to assert patents.
Why is this important? Well, in the US at least software algorithms can in some circumstances be patented. If patented algorithms show up in the source or binary code of software applications a bare copyright license will give the licensee only partial protection.
Sometimes relevant patents might be in the hands of third parties and this is always a risk for the re-distributor of OSS code. However, the Apache license at least ensures that those companies who are contributing to and re-distributing the OSS code don’t get to hold other licensees and their customers to ransom.
The ‘Open LLM’ debate
Now consider the Large Language Models (LLM) that are intrinsic to GenAI. There is much discussion about what the term ‘Open Source’ or even the word ‘Open’ actually mean in the context of these models. Many commentators have observed that the likes of Meta’s Llama and Apple’s DCLM are not truly Open Source because they are not made available under one of the OSS license formats recognized and approved by the Open Source Initiative (OSI).
This may be true but it also misses a far bigger point. Software code and algorithms form a relatively small component of these models. Take a look at this description of the components of the Llama model that appeared recently in a professional media post:
– PromptStore: For storing and managing prompts for Llama models.
– Batch Inference: A tool for making requests on batches of data.
– Continual Pretraining: A method for continuously pretraining Llama models on new data.
– Realtime Inference: A tool for making predictions in real-time.
– Quantized Inference: A method for optimizing inference performance through quantization.
– Evals: A component for evaluating the performance of Llama models.
– Fine Tuning: A method for fine-tuning Llama models on specific tasks.
– Reward Scoring: A tool for scoring rewards for Llama models.
– Synthetic Data Generation: A method for generating synthetic data for Llama models.
– Data: A component for managing and processing data for Llama models.
– Models: A component for managing and deploying Llama models.
– Hardware: A component for managing and optimizing hardware resources for Llama models.
– Accelerators: A type of hardware accelerator for Llama models.
– Storage: A component for managing and storing data for Llama models.
– Safety: A component for ensuring the safety and reliability of Llama models.
So what about the patents?
This list is no doubt representative of most if not all LLMs. So what about the patents? The general public might be lulled into a false sense of security, thinking that since Meta has opened access to much of this information there are no patents covering any of these components.
Ask any patent attorney and they will tell you this is not necessarily the case. Apple is very transparent that their Apple Sample Code License grants rights only under their copyrights. No patent rights are granted expressly or by implication and no protection is offered to licensees whose derivative applications may be found to infringe the patents of others (whether DCLM licensees or otherwise). The message is unfashionably clear: when it comes to patents you are on your own.
The GenAI patent landscape
A quick review of the Generative AI Patent Landscape report published earlier this year by WIPO should be enough to convince you that patents are an important and increasing phenomenon in this space. It might be enough to look at the headline figures but as the report states, the lengthy gap between patent filing and patent grant means that the numbers shown in this report may be significantly short of the emerging reality. Chinese entities both in the commercial and research sectors feature most strongly in the ranking but as you would expect the likes of IBM, Microsoft and Alphabet are not too far behind. While Meta itself does not figure strongly in the report other than for its number of scientific publications, it might be naive to think that it has not amassed enough at least to provide the company with a solid defensive position, especially given the patent portfolios of its key competitors.
Patent wars have loomed large over every industry in recent times where massive profits have been at stake. Customers of the protagonists, both large and small, have often found themselves as targets or have otherwise been swept up in lengthy and time consuming litigation.
Conclusion
However the LLM in question is being marketed (Open Source, Open Access, Source Available, Fair Source etc,etc.), the time-honored warning ‘caveat emptor’ was never more appropriate. Let the buyer beware. There be dragons for sure.
Your mOSS team
Leave a Reply