ChatGPT usage continues to grow, with more than 1.8 billion monthly visits and 10 million daily queries. It runs on GPT-4, a large language model (LLM), with several competitors, including Google Lamda, Hugging Face’s BLOOM, and others.
There are huge emotions, fears, hype, and investments surrounding ChatGPT, LLM, and other artificial intelligence (AI) generative capabilities. People and businesses are experimenting, and while it’s been less than a year since many of these capabilities became available, it’s worth asking at least two key questions: Where do these services bring business value, and which activities are risky or beyond current capabilities.
The answers are not easy because generative AI skills are evolving rapidly. For example, GPT-4 was first announced in March 2023 and became LLM for all ChatGPT users in May. Also, what works well for one person and company may not generalize well to others, especially now that we must master the new skill of rapid engineering.
But it’s hard for organizations to stand by and ignore the opportunities and risks. “ChatGPT and LLM can change the fundamental equation of business,” says Patrick Dougherty, CTO and co-founder of Trait. “Rather than corporate output being hampered by investment in human time, your only limitation will be the quality of your strategic decision making.” This guides what you can and shouldn’t do and what you should and shouldn’t do.
Do not share proprietary information in public LLMs
Many companies are drafting ChatGPT policies, primarily concerned with the risks of sharing sensitive business information. In a recent case, engineers asked for debugging help by paying for proprietary code on ChatGPT.
“The problem with ChatGPT and many other AI tools is that any data you paste becomes part of your training dataset. If someone enters proprietary information, that information may appear in a competitor’s materials.” So before experimenting and exploring use cases, review the company’s AI and data governance policies and disclose your goals, if necessary, for compliance.
Review LLM capabilities in primary workflow tools
In recent months, many technology providers have announced new AI and LLM capabilities built into their platforms. If you’re looking for business value, review how these capabilities improve productivity, simplify access to information, or provide other new operational benefits. Here is a sampling of several recent announcements:
- Microsoft 365 Copilot is built into Word, Excel, PowerPoint, Outlook, and Teams.
- Adobe Firefly is a generative AI that connects to Photoshop, Illustrator, and Adobe Express.
- Salesforce announced AI Cloud with integrations into its core CRM products, Slack and Tableau.
- GitHub Copilot integrates with IDEs and makes code suggestions.
- Google Duet AI for Google Cloud includes code support, chat support, and AppSheet capabilities.
- Atlassian Intelligence summarizes information and answers questions across Jira Software, Service Management, and Confluence.
- ServiceNow announced integrations with Microsoft Azure OpenAI Service and OpenAI API LLM and enhancements to AI-powered search.
- Crowdstrike introduced Charlotte AI to help stop breaches and reduce the complexities of security operations.
- Coveo Relevance Generative Answering adds LLM capabilities to its intelligent search platform.
Get answers fast, but know the limits of LLMs
A primary use case for ChatGPT and LLM is to get quick answers without doing all the underlying research or learning required to become an expert. For example, marketers may seek help writing customer emails, technologists may want technical terms defined, or human resources can ask for help reforming a policy.
LLMs built on business content also offer many benefits, allowing employees to ask questions to speed up onboarding, understand company benefits, find product information, or identify subject matter experts.
ChatGPT and other LLMs can boost productivity, upskill people, and help create content.
“Generative AI is incredibly useful in helping companies generate rapid analytics and reports by searching the web for open source intelligence like government, economic, and financial data,” says Raul Martynek, CEO of DataBank. “AI is already helping us quickly understand our data center environment, our customers’ intent, and our people’s sentiment to ensure we make fast, informed decisions across all business dimensions.”
But it is essential to understand the limitations of ChatGPT and other LLMs. Alex Vratskides, CEO of Persado, says: “Sam Altman, CEO of OpenAI, was right when he said that ChatGPT creates a ‘misleading impression of grandeur.’ If he’s looking for a productivity boost, ChatGPT is an awesome tool. But ChatGPT alone is still untested, insufficient, and can be misleading.”
Vratskides recommends that greatness comes when AI enables people to make better decisions. “When transformer models are trained on behavioral data from business communications, the language can be personalized to motivate people to engage and take action, thus driving business impact.”
People should also expect AI biases, as models are trained on sources that contain conflicting information, falsehoods, and biased opinions. Marko Anastasov, co-founder of Semaphore CI/CD, says, “Although powerful, language models are ultimately limited by ingrained biases in their training data and the complexity of human communication.”
Lastly, while ChatGPT is an excellent research tool, users should review what data it was last trained on. “ChatGPT is unaware of the latest events or news,” says Anjan Kundavaram, Precisely’s product manager. “It is also trained in text-based human conversations, using potentially inaccurate, false, or misleading data. The integrity of the data that feeds an AI model directly impacts its performance and reliability.”
Simplify understanding of complex information.
In many places in an enterprise’s information and technology stack, it is difficult to identify critical information within complex content and data sources. I hope many companies explore using AI search to improve customer and employee experiences because keyword search boxes are generations behind natural language queries and suggestions.
Finding information is one use case, and solving operational problems quickly is another. For example, performance issues in a multipurpose database can take a team of site reliability engineers, database administrators, and development engineers considerable time to find the root cause. “Generative AI will make it easier to manage and optimize database performance,” says Dave Page, vice president and chief architect of database infrastructure at EDB. “AI-powered tools can automatically monitor databases, detect problems, and suggest optimizations, freeing up valuable time for database administrators to focus on more complex tasks.”
But, Page acknowledges, “Database problems can be complex, and there may be factors that the AI can’t take into account.”
Another use case is simplifying access to large and complex unstructured information sources, such as product manuals and operational training guides. “Our clients generate a ton of documentation that can be hard to follow, not easy to search, or out of reach for the average user,” says Kevin Miller, CTO of IFS North America. “We see LLMs as a great way to help provide context to our users in new ways, including unlocking the power of service manuals and showing how other users have solved similar problems.”
But Phil Tee, CEO and co-founder of Moogsoft, warns of a false equivalence between knowledge and understanding. “ChatGPT and other LLMs provide technical advice and explain complicated processes on a more human level, which is incredibly valuable: no jargon, just information, although we’ve certainly learned to fact-check information,” he says. “But knowing that a set of steps will solve a problem is not the same as understanding whether these steps are right to apply now, and that becomes detrimental if we rely too much on LLMs without questioning their outcome.”
Suppose you’re considering plugging an LLM capability into one of your applications. Phillip Carter, Senior Product Manager at Honeycomb, shares a recommendation in that case. “Challenge yourself to think about what aspects of your product people are having the most difficulty with today, ask yourself first what can be solved without AI, and only look to LLM when you reduce effort or teach new users to help solve those problems.” He adds, “Don’t be fooled into thinking you can put a chat UI in some sidebar of your product’s UI and hope people get excited.”
Get Ready to Build LLMs on Proprietary Data Products
People can use open LLMs like ChatGPT today, use LLM capabilities built into their software platforms, or experiment with generative AI tools from startups. Developing a proprietary LLM is expensive, so it is not an option for most companies. Using an existing LLM to build proprietary capabilities is an option that some companies are beginning to explore.
The most significant opportunities are for companies with expertise in a specific domain that is building the context and layers of knowledge on top of LLMs and using them as translators to deliver personalized interaction with each user.
Domain-specific LLMs include Intuit GenOS, an operating system with custom-trained financial LLMs specializing in solving economic challenges. Another example is BloombergGPT, a 50 billion parameter LLM trained on 700 billion tokens of financial documents and public datasets in English.
“LLMs are already in place and driving business value today, but they just don’t look like ChatGPT,” says Kjell Carlsson, head of data science strategy and evangelism at Domino. “Biotech companies are accelerating the development of proteins for new treatments, while organizations in all industries use LLM to understand customer conversations and optimize customer service operations.”
Integrating LLM capabilities into the existing business model is challenging. The generative capabilities of these models are currently the most brutal ways to drive business value because the business use cases are untested and because of the enormous limitations, including cost, privacy, security, and control of the models. Similar to ChatGPT, that is consumed as a service.
Companies with business models generating revenue from their large, proprietary, and unstructured data sets should consider opportunities to incorporate their data into LLM. “Businesses can run and manage specialized models within their own security boundaries, giving them control over data access and usage,” says Tabnine co-founder and CEO Dror Weiss. “More importantly, companies can customize specialized models using their own data, which is essential for machine learning models to produce accurate results.”
The opportunity to build LLMs in industries with rich data sources, such as financial services, healthcare, education, and government, is significant. So is the potential for disruption is one reason business leaders will explore the opportunities and risks when applying LLMs to their products and operations.