What Enterprises Can Do To Maximize The Impact

Table of Contents

Clément Stenac, Chief Technology Officer, Dataiku.

Three months after the release of ChatGPT, OpenAI just revealed the third upgrade to the blockbuster AI platform: GPT-4. The rapid pace of change presents organizations with a daunting challenge: They want to move faster than their competition and incorporate the latest AI tool, but they need to do so responsibly—which is extra important when deploying a potentially world-changing technology like AI.

I’ve heard in conversations with dozens of enterprise leaders that they’re grappling with a business-critical question: How can you best integrate GPT-4 and other, comparable emerging technologies?

OpenAI claims that GPT-4 is its “most advanced system, producing safer and more useful responses,” allowing users to analyze images and mimic speech, designed to serve as the underlying engine that powers chatbots and other systems. The company also announced that Microsoft’s Bing AI chatbot had already been using the new software since its launch in February.

So, what can companies do to take advantage of GPT-4 and its successors? Three main efforts will pay dividends, regardless of the specific functionality of any one model:

Understanding The Underlying Technology

Deploying generative AI effectively starts by first understanding how it works, its capabilities and its limitations. In the case of large language models (LLMs) used in ChatGPT, their defining characteristic is their ability to produce text content on par with what could be created by a human. Their limitations have been widely documented and include a lack of explainability—an LLM cannot cite its sources—and a propensity for inaccuracy, both of which will limit its enterprise applications. This generation of LLMs has also been trained on unknown, broadly generic data, meaning that they usually lack domain expertise that’s necessary for an enterprise application—like setting pricing strategy in the healthcare market or improving productivity for a bank. And there’s always a healthy risk of producing inappropriate content, which we’ve seen several high-profile examples of in media and online.

Gearing Up On Governance

Critically, now is the time for enterprises to create and implement AI governance—a collection of practices and processes to ensure that an organization strikes the right balance between rapidly taking advantage of these new technologies while focusing on business needs and minimizing risk. Companies can assess useful business applications, such as eliminating IT overhead or speeding data analysis, according to their potential benefit to the business, the resources required to develop them and any associated risks.

Risk is the essential element of this analysis. Though tech companies like Grammarly have built strong businesses on LLM technology, deploying these technologies in a range of enterprises old and new is a whole new frontier. Each organization will need to decide for itself how much risk they are willing to take for potential benefits and market leadership.

Readying Infrastructure At Scale

Most don’t appreciate that the most powerful LLM models are truly massive. The models garnering attention recently—GPT-3 and the newly-released GPT-4—are larger than traditional machine learning models and may continue to grow exponentially. They are prohibitively expensive for all but the largest tech companies to develop and execute and, in the case of the OpenAI models, are closed-source and available only via a paid API as a “model-as-a-service.” As many who built their business on Facebook or other platforms have learned the hard way, building core business capabilities on top of an API puts an organization at the mercy of the API owner and is thus a notable risk for an organization.

Since most organizations don’t have the resources to develop these models themselves, and if accessing a closed-source, pay-per-use model poses too much risk, for many businesses, it will make sense to work with a smaller, open-source large language model, such as BERT, Flan, GPT-J or other libraries provided by companies like Hugging Face. By “fine-tuning” (i.e., adapting) these models on internal, specific data, companies could achieve impressive business value, even if the platform can’t churn out award-winning sonnets on the side. A few key advantages could include:

1. Output that’s more specific and relevant to the organization. These models are particularly powerful in what’s called “few-shot learning,” meaning that the model only needs a few labeled examples to learn a domain.

2. More control over moderation to prevent unsavory or inappropriate outputs, while also improving the relevance of the response to the business.

3. All data stays within the organization’s firewall, helping meet confidentiality and data residency requirements.

4. Controlled costs for running the model, as the organization eliminates exposure to changes in API pricing from a for-profit supplier.

Such a model would not have the broad capabilities of a general-purpose, large language model like GPT-4, but many of those capabilities are irrelevant for targeted enterprise applications. For instance, most service desks don’t need to emulate the voice of Hemingway or offer tips on Mexican vacations; they just need a quick summary of a longer transcript. And while initial model setup requires specialized competencies, these models can then be deployed throughout the organization, serving nearly all lines of business. Putting infrastructure in place to enable such reuse is a smart precondition for the initial investment of setting up such a model.

Enterprises have many infrastructure options, from open-source models running in-house to the exclusive use of models-as-a-service and everything in between. Smart approaches will enable organizations to tailor the right strategy for itself while leaving room to adjust nimbly as new technologies emerge and changing market conditions.

Many companies are salivating at the power of ChatGPT and searching for a way for it to catapult them to market leadership. Building understanding, establishing governance and readying a smart infrastructure will be essential to capitalize on this heady promise.

The release of GPT-4 is a milestone in the history of AI, but for it to have a real impact on the enterprise, organizations need to prepare themselves to be able to take full advantage of its impressive capabilities.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Leave a Reply