“ChatGPT is an open-source natural language processing (NLP) model based on GPT-2, a transformer-based auto-regressive language model. It enables developers to generate humanlike conversations from limited input data and can be used for applications such as chatbots, conversational AI, question answering systems and more.”
This is how ChatGPT described itself when prompted.
But its capabilities go much further than simple text generation. Navigating through the website shows there are several ways to utilize ChatGPT including image generation, code completion and fine-tuning or training a model for specific cases.
Though ChatGPT rose to fame in recent months, artificial intelligence was a long time in the making, according to previous University of Wisconsin assistant professor of statistics and lead AI educator for Lightning AI Sebastian Raschka.
AI began around 1958 with the Perceptron model, which was the first case where people attempted to teach computers how to learn from data.
“You can have, for example, a very sophisticated set of rules where you say I want to classify emails as spam,” Raschka said. “So what you could do is you could technically code on if the email contains ‘sale’ or ‘discount’ or something like that, then spam, and you can have a chain of these rules to make that more complicated and develop an AI based on these rules.”
People respond positively to humanlike robots, UW researcher finds
From there, AI developed into deep learning around 2012 after outperforming older methods like support vector machines and random forests, Raschka said. These models worked well on tabular data but not on text and images.
Deep learning is a subfield of machine learning that teaches computers to learn by example, similarly to how humans learn, according to MathWorks. This method is what allows driverless cars to perform.
According to UW associate professor in the Department of Biostatistics and Medical Informatics Anthony Gitter, deep learning refers to training neural networks — a group of algorithms that tries to recognize relationships in a data set by mimicking the human brain — or artificial neural networks on example data. Then programmers can apply them to a trained model.
“We can apply some sort of trained model on new data to do all sorts of tasks — so recognize objects and images or generate new text and languages like ChatGPT or translate languages from one to another or transcribe audio to words,” Gitter said.
Faculty associate in the UW Department of Computer Sciences Meena Syamkumar said developers need to carefully train the models on many resources to avoid any biases in programming that may slip into the code.
But Syamkumar said this doesn’t mean the AI will be completely unbiased.
“Of course, they [programmers] need to be really careful about what [training data] they feed into the [large language model] in order to make sure that it’s not biased, which fundamentally is going to be a problem because you have a lot of data out there which can lead LLM models to being more biased than most human beings would be,” Syamkumar said.
Working around biases in AI models just puts more constraints on the work, Syamkumar said. This means humans need to be involved in the process of cleaning up data and not just simply feeding the model any and all data. By doing this, humans can make sure that data gets tagged correctly as to whether it is good enough to use as learning material.
With such a powerful tool accessible to anyone who owns a computer or phone, a big concern circulating UW’s campus is the use of ChatGPT to cheat in class on homework assignments.
Syamkumar said she is currently researching how instructors can leverage ChatGPT as an educational tool instead of worrying about its implications. She said introductory computer science courses, in the format they’re offered now, have been a challenge. They could become obsolete because students can use ChatGPT to solve every problem.
“That is concerning, but rather than focus on the fact that students can use this as a platform to cheat, what I want to do is to tell the students that ChatGPT has a potential to take away the knowledge that they can possibly get out of doing the coursework,” Syamkumar said.
Feb. 24 the UW Teaching Academy hosted an event to discuss the many new questions ChatGPT raises such as “How might we adjust to students using ChatGPT in our classes?,” “How does ChatGPT interface with issues of inclusivity and bias?” and “How should campus address the ethics of using ChatGPT?”
Event organizers Morton Ann Gernsbacher and John Martin approached the event as a way to bring awareness to the fact that ChatGPT doesn’t have to be a threat to learning. Instead, it can be worked into the curriculum, but instructors and curriculum developers need to better understand the tool first.
In fact, the very technologies inside ChatGPT are used all over in science and even all over UW medicine, Gitter said. His research group trains AI on scientific ideas involving proteins.
“There’s this transformer neural network style that is powering ChatGPT that’s a core part of how it trains and predicts words and things like that,” Gitter said. “My research group is using that to study proteins and how to design proteins, and we’re really excited about how these general-purpose AI developments kind of can inspire our new ideas.”
Though there are negative connotations associated with AI, Syamkumar and Raschka agreed the best solution isn’t to fight against it or ban it in classrooms.
Raschka said it’s like Pandora’s Box — it’s open and hard to undo because there will just be more of these models.
“The question is essentially how can we make the best out of it, and how can we make or use that to our [advantage]?” Raschka said. “[I don’t think it’s] the machine replacing humans — it’s more like humans using the machine, replacing humans who don’t use the machine. Because it will, essentially, make you more productive.”
