New chatbots are hitting the market; an ASU AI expert explains their differences
Move over, ChatGPT — you’ve got competition.
New artificial intelligence chatbots are being made available to the public, researchers and practitioners.
With Google, Meta and Microsoft releasing language models (and with a few minor ones out on the market), the AI space is heating up, and people are commoditizing it. AI will eventually make its way into daily use across a variety of goods and services.
Victor Benjamin, an assistant professor of information systems at the W. P. Carey School of Business, is an expert in the AI field.
Here’s what Benjamin had to say about the new models and how we use machine learning.
Editor's note: Answers have been edited for length and clarity.
Victor Benjamin
Question: How do these new models coming onto the market differ from ChatGPT?
Answer: Large language models have two key differences from one another: one, their computational architecture, and two, the data on which they learned.
Regarding computational architecture, we are talking about the math and logic driving the system. All our AI today is rooted in math, and we are finding new ways to push math forward into letting us create machines that can somewhat navigate aspects of the real world and help us do things. That’s all AI is. The math for AI often converges into a subset of techniques that receive most of the attention, and different teams working on AI may try out different tweaks or mathematical tricks to try and further improve their AI. So ChatGPT competitors may use similar math and computational architecture but still possess some differences and twists that produce unique outcomes.
Regarding the data on which AI trained, this is equally important as the math itself. You can have two identical computational architectures for AI but just trained on different data sets, and they will produce two very distinct AI outcomes. Right now, all these large language models rely on text sources from books, magazines, social media, newspapers and whatever they can get their hands on. The exact composition of these sources and how people pre-process them to be fed into AI training produces the second major difference between ChatGPT and its competitors.
To illustrate this, consider a naïve example of a new language model trained exclusively on articles about different sports games. This model would never have received any training regarding other topics, such as musical instruments, and would thus be unable to produce meaningful responses regarding those different topic areas. Likewise, different large language models will have distinct mixtures of input data that can impact their efficacy.
Q: What do those differences mean for the average person and within academia?
A: As more AI services become available, one can expect to receive slightly different outputs from each AI. These different outputs will not be exclusive to large language models; any AI service will have the two aforementioned components of a computational architecture and a large training data set to learn from. So if you are not happy with the results produced by the AI from one vendor, you can always try another.
Q: What is your take on using machine learning or services as a resource for college students?
A: One issue with these AI is that they do not possess ground truth about the real world. Large language models are not conscious of what they are outputting; to the computer, it is regurgitating numbers and statistical patterns. Large language models are just a form of advanced text autocomplete rather than a proper AI. ... (One) dangerous scenario would involve the AI completely misinforming students with inaccurate information, due to lack of ground truth, and the student will believe the misinformation.
I will, however, also speak to the positive potential here. These large language models can eventually be useful educational tools for students. They show promise of being personal tutors and can provide personalized attention for each user. Large language models and related AI could be excellent guides in self-learning environments, such as online courses. The technology will mature eventually, as will the policies regarding AI use.
Q: What’s AI’s impact on the future workforce? Can AI be used to create new industries and career opportunities?
A: AI will just be another tool that lets us do new things. We have a good problem here where AI can help us build new value propositions that have never been considered. We just need time to think about what those new avenues are. It won’t happen overnight, but eventually it will.
These value propositions will include both ways to augment the workforce in traditional roles and help create new career opportunities for the next generation of products and services that AI will enable. Some existing career paths may become obsolete, but AI will create others. It will come just like any other technological change; it just takes time for people to think it through.
Top photo courtesy iStock/Getty Images
More Science and technology
ASU student researchers get early, hands-on experience in engineering research
Using computer science to aid endangered species reintroduction, enhance software engineering education and improve semiconductor material performance are just some of the ways Arizona State…
ASU professor honored with prestigious award for being a cybersecurity trailblazer
At first, he thought it was a drill.On Sept. 11, 2001, Gail-Joon Ahn sat in a conference room in Fort Meade, Maryland. The cybersecurity researcher was part of a group that had been invited…
Training stellar students to secure semiconductors
In the wetlands of King’s Bay, Georgia, the sail of a nuclear-powered Trident II Submarine laden with sophisticated computer equipment juts out of the marshy waters. In a medical center, a cardiac…