Gaurav Aggarwal is Co-Founder of Truva and Forbes U30. Helping SaaS businesses improve customer adoption and reduce churn with AI.
getty
In the ecosystem, the recent announcement of OLMo, which they call an open-source, state-of-the-art large language model, has been sparking discussion. While proprietary models and corporations are under intense scrutiny for having control over the models—hence the direct control on public domain conversations and interface—open-source AI has also been controversial in certain quarters of academia, as well as among regulators. For example, an article published in IEEE Spectrum called open-source AI “uniquely dangerous.” It thus becomes worthwhile to establish a spectrum of openness in AI model development and deployment.
The spectrum ranges from proprietary models at one end to source-available models and, ultimately, to truly open-source models.
The purist, truly open-source distinction advocates for raising the bar on openness in AI development by providing complete access to its development tools, such as pretraining data, training codes, model weights and evaluation tools. Furthermore, this should include granting more democratized access to computing, funding, the ability to scrutinize and modify models and the opportunity to participate in or opt out of the data collection process. Additionally, there should be a focus on AI education to increase participation.
OpenAI’s ChatGPT or Anthropic’s Claude are examples of proprietary systems where the public can’t access the code and model weights that act as an intellectual property for companies. Commercial applications of these models are licensed strictly so that owners can have control over the technology. The pipelines and the data involved in them are not disclosed, thus hiding away the processes through which these models are trained.
This level of secrecy aims to protect intellectual property and commercial interests but hinders collaborative development and transparency in how AI algorithms make decisions that can significantly affect the public domain.
Slightly up the scale are source-available models. An example of this approach could be seen where certain components’ source code can be used under specific conditions—for example, Meta’s LLama 2 model. However, in regard to redistributing or commercializing this technology, there are some serious limitations as mentioned in the tech’s responsible use guidelines. This sort of model demonstrates some openness, which allows some educational and developmental use while safeguarding proprietary advancements and innovations.
The enhanced cost-effectiveness of computing power opens up new avenues for models such as GPT4All or OLMo, which can be examples of the purist distinction. The public availability of their underlying code and datasets invites anyone to scrutinize and understand the methodologies behind their creation. Moreover, the adaptability of these models to less resource-intensive environments means that individuals or entities with limited access to high-end GPU hardware can still harness the advantages of AI innovations. This accessibility democratizes AI, extending its benefits beyond well-resourced labs and companies to a broader audience.
The call for a thorough action on both ends of the spectrum, especially for those who claim to be truly open but instead use the open-source technology for furthering a particular agenda, is a conversation that must be had. It’s worrisome how easily these AI systems can be hijacked for malicious purposes.
For instance, these models can be modified by experienced users to eliminate any internal confines. Those users employ the models to achieve harmful goals. Intensifying this problem are vulnerable channels like social media platforms and messaging apps. These sites have difficulty distinguishing between content generated by humans and those created by machines; hence, they are hot spots for the dissemination of misleading information and fraud. This raises worrying concerns about LLMs’ potential effects on public deliberation, as well as democratic processes, that necessitate concerted governance efforts to improve responsible applications of AI.
Though easier said than done, the best way forward is for all stakeholders to take the following into consideration, no matter their placement on the open-source spectrum.
1. Openness To Knowledge: All of these models should have a framework that allows people to opt out of their data being leveraged to train the models. This implies transparent communication about how data is used, giving people power over the development process. In addition, considering personal data’s value, means for compensation should be designed for those who allow their data to be used.
2. Openness To Compute: The astronomical cost of training state-of-the-art AI models limits access only to a few well-endowed entities and forces significant scholars out of the game. Handling this will involve developments in hardware that are cost-effective and building a more democratic compute and funding landscape.
3. Openness To Decisions: AI should not be a monolith. Instead, it may look like a mosaic characterized by different pieces, each portraying a different part of human society. In other words, we must develop and deploy various models tailored to particular local contexts, while respecting local laws and cultural nuances therein.
Beyond these aspects, the overarching direction of LLMs should prioritize the well-being and benefit of individuals worldwide. Collaborating across borders to establish global standards and ethics for AI development can help us realize a future where AI serves humanity’s broadest spectrum, not just the interests of a privileged few.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
