Three months after the introduction of ChatGPT, OpenAI today announced the third major update to the AI platform: GPT-4. The high speed of development is a formidable problem for organizations: they want to be faster than their competitors and use the latest AI tool, but they must do so responsibly, which is especially crucial when adopting potentially groundbreaking technology such as AI.
In most talks, hundreds of business leaders indicate that they are wrestling with a mission-critical question: How can GPT-4 and other analog new technologies be optimally integrated?
GPT-4, according to OpenAI, is the “most advanced system that produces more secure and useful responses,” allowing users to analyze photos and mimic speech, and is intended to be the basic engine that powers chatbots and other systems. The company also stated that Microsoft’s Bing AI chatbot has been using the new software since its February debut.
So, what can companies do to take advantage of GPT-4 and its successors? Regardless of the specific functionality of a particular model, three major efforts will pay off:
Understand the underlying technology
Understanding how generative AI works, its potential and limitations, is the first step to deploying it effectively. The capacity of large language models (LLMs) used in ChatGPT to produce text material similar to what a person might develop is its distinguishing feature. The drawbacks are well reported, and they include a lack of explainability – an LLM cannot specify its sources – as well as a propensity for impreciseness, both of which limit their business applicability. This generation of LLMs has also been trained on unknown, broadly generic data, meaning they typically lack the subject matter knowledge needed for a business application, such as setting a pricing strategy in the healthcare market or improving banking productivity. And there’s always a healthy risk of producing inappropriate content, which we’ve seen several high-profile examples of in the media and online.
Preparation for governance
Crucially, now is the time for companies to develop and implement AI governance: a set of principles and processes that ensure an organization strikes the right balance between rapidly adopting new technology and focusing on business objectives and avoiding risks. Companies can evaluate relevant business applications, such as reducing IT overhead or accelerating data analysis, based on their potential benefit to the business, the resources required to build them, and any associated risks.
The element of risk is critical to this study. While LLM technology has helped software companies like Grammarly build great businesses, applying these technologies to a variety of businesses old and new is a whole new arena. Each company must choose how much risk it is willing to take in exchange for potential benefits and market leadership.
Infrastructure readiness at scale
Most people don’t realize how big the most powerful LLM models are. GPT-3 and the recently released GPT-4 models are larger than standard machine learning models and can continue to expand exponentially. They are too expensive to build and run for anyone but the largest tech companies, and, in the case of the OpenAI models, are closed-source and only available through a paid API as a “model-as-a- service”. Building foundational business capabilities on top of an API, as many who have established their businesses on Facebook or other platforms have found out the hard way, places an organization at the mercy of the API owner and thus poses significant risk to an organization .
Because most organizations don’t have the resources to develop these models themselves, and because access to a closed-source, pay-per-use model involves too much risk, many companies will benefit from working with a smaller , open-source large language model, such as BERT, Flan, GPT-J, or other libraries provided by companies such as Hugging Face. Companies can create tremendous financial value by fine-tuning (ie adapting) these models based on internal, specialized data, even if the platform cannot produce award-winning sonnets in doing so. One of the main benefits could be:
1. Output that is more specific and relevant to the organization. These models are particularly powerful in what is called “single learning,” meaning that the model only needs a few labeled examples to learn a domain.
2. Greater control over moderation to prevent unsavory or inappropriate results while improving the relevance of the comment to the company.
3. All data remains within the corporate firewall, ensuring confidentiality and data retention requirements are met.
4. Controlled costs of running the model, as the organization eliminates exposure to changes in API prices from a for-profit vendor.
While such a model lacks the extensive features of a massive general-purpose language model such as GPT-4, many of those features are useless for targeted enterprise applications. For example, most service desks don’t have to imitate Hemingway’s voice or offer advice on vacations in Mexico; they only require a concise summary of a larger transcript. Although the initial set-up of the model requires specific skills, these models can later be deployed company-wide to serve virtually all industries. Setting up infrastructure to support such reuse is a sensible precondition for the initial expenditure of setting up such a model.
Companies have a variety of infrastructure alternatives, ranging from in-house open-source models to the exclusive use of models-as-a-service and everything in between. Smart techniques enable companies to develop the right strategy for themselves while offering flexibility to react quickly when new technologies emerge and market conditions change.
Several companies are salivating at the power of ChatGPT and looking for a way to use it to propel them to market leadership. Harnessing this intoxicating potential will require building awareness, developing governance and preparing smart infrastructure.
The introduction of GPT-4 marks a watershed moment in the history of AI, but for it to have a real impact in the workplace, enterprises must equip themselves to fully exploit its amazing capabilities.
Source: Forbes Technology Council