Last month, Chinese artificial intelligence startup DeepSeek sent stocks tumbling after making headlines with its new ChatGPT-like AI model called R1. Even more shocking was the company’s claims that the program was developed at a fraction of the cost of OpenAI’s latest model but is just as effective.
Robin Jia, a USC Viterbi assistant professor of computer science, studies natural language processing and machine learning.
“You have to get a really large number of specialized types of computers [and] with powerful GPUs…and then you have to run those for a very long time on a very, very large amount of data,” he said. “This process is generally considered to be super expensive, which is why there’s only a few big companies in the U.S. who are really doing this, [such as] Google and OpenAI.”
Jia said language models are essentially trained to predict what the next word would be in a document or conversation.
“By repeatedly predicting what word comes next, they can do things like simulate conversation with you,” he said. “Or they can answer questions that you might have because they’re just guessing what is the word that comes [after this that can] answer the question.”
Marko Danial, a USC senior and graduate student studying aerospace and mechanical engineering, said he uses both ChatGPT Premium and DeepSeek AI.
“It’s pretty nice how the free version [of DeepSeek AI] is almost as good as ChatGPT premium,” Danial said. “I think that makes [DeepSeek AI] a lot more popular, and that’s what kind of concerns Open AI, [is] that DeepSeek AI is free but it’s almost as good as their own [model].”
Arjun Balamwar, a USC graduate student studying computer science, said that he thinks DeepSeek R1 is good at solving mathematical problems, but not at addressing creative prompts.
“A big distinction between DeepSeek R1 and other models is that compared to other [large language models] that are out there, it is not able to outperform all the time when it comes to ambiguous queries,” Balamwar said. “So if you tell it to ‘write me a script’ or ‘write me a poem,’ there is no correct answer to this. It is just going to need to be creative.”
Jia said what made ChatGPT-o1 special was that OpenAI was investing a lot of time and money into training the model to have some “kind of reasoning to get the right answer.
“Imagine you’re solving a difficult math problem. The answer itself could be something that’s simple, but you’re not going to be able to get the answer just by reading the question and immediately saying what the number is,” Jia said. “There would be a whole sequence of calculations, or things you might have to prove before you can actually figure out what the right answer is.”
That process is similar to the one R1 employs to get to an answer.
“It starts going into one direction … and then in the middle, it realizes that it was actually wrong, and that helps it in preventing hallucination, and then it pivots and then goes towards the correct, the current, direction,” Balamwar said. “Essentially, it starts to understand its own chain of thought and starts to actually logically answer questions.”
Jia said the issue of “hallucination,” is a well-known issue that causes generative AI models to produce misinformation.
“This is not a solved problem, but it has gotten better over time,” he said. “These R1 models are designed to try and make progress, but those are also where models are likely to have some failure.”
However, the information DeepSeek used to train R1 is still unknown, something Jia said fed into one of his ethical concerns regarding the entirety of the AI industry – a lack of transparency .
“Companies like Open AI reveal very little about how their model was trained. In the case of DeepSeek, they happened [to be] more upfront about describing the methodology, but there is still a lot that’s not described,” Jia said. “All these systems require a lot of data [because] they have to learn by seeing a bunch of examples. So, what [type of] data you’re able to get and show to your model to help it improve has a huge impact on your final results.”
Despite those concerns, Jia said students and educators should be aware of the technology, not only for the ways it can make their lives easier, but the ways in which it can impact the job market.
“There is definitely a lot of potential for disruptions to the labor force, making certain jobs, perhaps, a lot less necessary than they were before,” he said. “So that’s certainly something that students and educators definitely need to stay on top of.”
But, Jia noted, finding ways to prepare students for the future of work while not allowing them to become overly reliant on the technology will continue to be a topic of discussion.
“Maybe for exercises that are supposed to have a lot of instructive value, we should let the students know that it’s important for them to do it [themselves]. Then maybe for things a bit more open-ended, like the final project, it’s much more reasonable to say, ‘you can use any tools at your disposal and just create the best final product’,” he said. “It’s a delicate balance.”
As for students, Danial said they should work to learn more about new AI models and find ways they can be used in a beneficial way.
“I feel like [AI models are] going to continue progressing and [get] better,” Danial said. “And [they’re] something everyone should be ready to learn how to use.”