Why is ChatGPT making waves in the AI market?
By Bern Elliot, Distinguished VP Analyst at Gartner
Artificial intelligence (AI) research and deployment company OpenAI recently announced the official launch of ChatGPT, a new model for conversational AI. According to OpenAI, the dialogue provided by this platform makes it possible for ChatGPT to “answer follow-up questions, admit its mistakes, challenge incorrect premises and reject inappropriate requests.”
Since its launch, social media has been abuzz with discussions around the possibilities—and dangers—of this new innovation, ranging from its ability to debug code to its potential to write essays for college students.
ChatGPT is the perfect storm of two current ‘hot’ AI topics: chatbots and GPT3. Together, they offer a wonderfully intriguing method of interacting with and producing content that sounds amazingly human. Each is the result of separate, significant improvements over the last five years in their respective technologies.
Chatbots allow interaction in a seemingly ‘intelligent’ conversational manner, while GPT3 produces output that appears to have ‘understood’ the question, the content and the context. Together this creates an uncanny valley effect of ‘Is it human or is it a computer? Or, is it a human-like computer?’ The interaction is sometimes humorous, sometimes profound and sometimes insightful.
Unfortunately, the content is also sometimes incorrect and the content is never based on a human-like understanding or intelligence. The problem may be with the terms ‘understand’ and ‘intelligent.’ These are terms loaded with implicitly human meaning, so when they are applied to an algorithm, it can result in severe misunderstandings.
The more useful perspective is to view chatbots and large language models (LLM) like GPT as potentially useful tools for accomplishing specific tasks, and not as parlor tricks. Success depends on identifying the applications for these technologies that offer meaningful benefits to the organization.
At a high level, chatbots, or conversational assistants, provide a curated interaction with an information source. Chatbots themselves have many use cases, from customer service to assisting technicians in identifying problems.
At a high level, ChatGPT is a specific chatbot use case where the chatbot is used to interact (chat) or ‘converse’ with a GPT information source. In this case, the GPT information source is trained for a specific domain by OpenAI. The training data used on the model determines the way questions will be answered.
As noted earlier, however, GPT’s capability to unpredictively generate information that is false means that the information can only be used for situations where errors can be tolerated or corrected.
There are numerous use cases for foundation models such as GPT, in domains including computer vision, software engineering and scientific research and development. For example, foundation models have been used to create images from text; generate, review and audit code from natural language, including smart contracts; and even in healthcare to create new drugs and decipher genome sequences for disease classification.
AI foundation models such as GPT represent a huge step change in the field of AI. They offer unique benefits, such as massive reductions in the cost and time needed to create a domain-specific model.
However, they also pose risks and ethical concerns, including those associated with:
Complexity: Large models involve billions, or even trillions, of parameters. These models are impractically large to train for most organizations, because of the necessary compute resources, which can make them expensive and environmentally unfriendly.
Concentration of power: These models have been built mainly by the largest technology companies, with huge R&D investments and significant AI talent. This has resulted in a concentration of power in a few large, deep-pocketed entities, which may create a significant imbalance in the future.
Potential misuse: Foundation models lower the cost of content creation, which means it becomes easier to create deepfakes that closely resemble the original. This includes everything from voice and video impersonation to fake art, as well as targeted attacks. The serious ethical concerns involved could harm reputations or cause political conflicts.
Black-box nature: These models still require careful training and can deliver unacceptable results due to their black-box nature. It is often unclear what factbase models are attributing responses to, which can propagate downstream bias in the datasets. The homogenization of such models can lead to a single point of failure.
Intellectual property: The model is trained on a corpus of created works and it is still unclear what the legal precedent may be for reuse of this content, if it was derived from the intellectual property of others
Start with natural language processing (NLP) use cases such as classification, summarization and text generation for non-customer-facing scenarios and choose task-specific, pretrained models to avoid expensive customization and training. Use cases where output is reviewed by humans are preferred.
Create a strategy document that outlines the benefits, risks, opportunities and deployment roadmap for AI foundation models like GPT. This will help determine whether the benefits outweigh the risks for specific use cases.
Use cloud-based APIs to consume models, and choose the smallest model that will provide the accuracy and performance needed to reduce operational complexity, lower energy consumption and optimize total cost of ownership.
Prioritize vendors that promote responsible deployment of models by publishing usage guidelines, enforcing those guidelines, documenting known vulnerabilities and weaknesses and proactively disclosing harmful behavior and misuse scenarios.