Everyone is talking about Generative AI and more and more companies want to incorporate it into their ranks. The immediate accessibility provided by this technology makes it different from all its predecessors. Unlike other technological tools, to use it, interact with it and obtain value, we don’t need knowledge or certifications. Virtually anyone can use it, as the mechanism is simple and available to everyone. A Generative AI platform can give rise to countless applications for audiences of any age or educational level that have access to the Internet.
Having broken this accessibility barrier, Generative AI has become a very attractive technology for everyone, and for the corporate environment is no less so. Its uses are very varied and it can help us to optimize tasks, processes and create original content, among other things. However, the fact that it’s so easy to implement it, may entail risks when it comes to start working with it without any prior preparation. Companies need to be especially careful when introducing Generative AI into their business. They will need to create a solid implementation strategy to avoid hurting the bottom line, losing money, or not utilizing the team in the best possible way. The barrier to entry is lower, yes, but the risk is high.
Risks and pitfalls
One of the main risks to take into account when implementing Generative AI, is what we know as Language Model (LLM) Hallucinations. This happens when the models try to answer a question and, even if they don’t know the answer, they are going to create one. This can be good if we want to use the model in a more creative or subjective way, but if we are looking for a concrete answer that is correct, it can be a problem. If, for example, a company introduces a Q&A chatbot on its website and a customer asks a question whose answer isn’t among the documents with which the AI has been trained, we can’t be sure that the model will not create an answer out of thin air. An answer that should not be controversial may become controversial.
Models aren’t deterministic; you can’t mathematically determine the answers it will give. New guidelines and barriers must be put in place to ensure that the model behaves as expected, and this isn’t straightforward.
Cost is probably another variable to consider before implementing Generative AI in a business. The models aren’t cheap, and neither are the machines needed to run them, as they need to be quite robust, nor is the cost per token.
Best practices in GenAI
To properly adopt Generative AI in our business we need to consider the following:
- We need a solid understanding of what Generative AI is as a whole. At the foundation we find the base models, such as GPT-4 or PaLM 2. These models provide us with the fundamental capabilities for Generative AI applications.
- Next is the data engine that provides the data customization and tuning needed to enable the base model to correctly use the business data.
- Subsequently, we need a development platform to create LLM applications, compare model directions and variants, and deploy applications in production.
As we can see, it is a complex technology that requires a solid knowledge base to understand how it works. In addition, in general, enterprise-level implementations of Generative AI involve some degree of in-house development of their own applications. This allows companies to customize and fine-tune models to optimize performance for their specific use cases and tailor them as closely as possible to their needs. Each business has unique needs that require extensive tuning and rapid engineering of the base foundation models. Commercial and open source models are very general, for enterprise use cases they fall short as they require domain knowledge or business specific data. This is why a solid implementation strategy is necessary.
How can you start creating such a strategy?
- First of all, it is very difficult to integrate something if you don’t talk about it or teach it to your teams. Multiple factors need to be taken into account and a series of steps need to be taken to be responsible in the installation. Teams aren’t always aware of all of that, so you need to train them.
- A sound technology strategy must also be implemented. Choosing the right model cloud provider is a critical step. The three providers (AWS, Google and Azure) have their own offerings, it is a matter of making a study about which option is more beneficial to the business objectives. Therefore, it’s necessary to be very clear about the objectives and what is intended to be achieved with Generative AI.
- Companies must ensure that their models operate in an accountable and transparent manner. Companies that implement governance in their AI initiatives can benefit in several ways. They can better detect and mitigate model risk while strengthening their ability to comply with ethical principles and government regulations.
The past few months have been a very exciting period of innovation and experimentation with Generative AI. Now it is the turn of companies to deal with the usual competition as they worry about responsibly and effectively implementing this technology. As we have seen, it is a process that presents complex challenges for any company, but with a solid strategy and plan, they should feel well positioned to become pioneers in the use of Generative AI.
The integration of Generative AI into companies should happen naturally and through an organic process. This will improve the experience of the company itself, the employees, and even the customers.
Image: Unsplash | Mojahid Mottakin