Rice Computer Science faculty member Anshumali Shrivastava has been promoted to full professor, effective July 1, 2025. Recognized with numerous awards, Shrivastava is well on his way to revolutionizing how large language and other deep learning models are trained and stored, using new algorithms to make AI scalable and more accessible.
Chris Jermaine, Rice Computer Science Department Chair, has referred to Shrivastava as “a disruptor” and described him as someone who “is trying to fundamentally change how an entire research area operates, challenging the most basic assumptions regarding how we should train modern deep learning models.”
Shrivastava received his Ph.D. in computer science from Cornell, then joined Rice’s faculty, in 2015. In 2017, he received an NSF CAREER Award, and in 2018, was named one of 10 Scientists to Watch by Science News. He was the recipient of an Adobe Data Science Research Award in 2021–the same year he was granted tenure and recognized with Rice Engineering’s Young Faculty Teaching Award. In spring of 2023, he received the Charles W. Duncan Jr. Achievement Award for Outstanding Faculty, which recognizes not just notable research but also success in teaching.
What motivates him, and how he motivates students
Shrivastava said he got interested in mathematics and science when he was growing up because the self-reliant aspect appealed to him. “You are independent. You can validate things, and you can “literally” change the world single handedly if you have enough insights,. That was pretty powerful,” he said.
“And I think in computer science, especially AI, this is very true. If you look at other fields, not so much, you have to rely on experiments. You have to rely on equipment. But in computer science, you can have a single idea developed over a short span of time changing the course of the entire field, and that is still possible, even today.”
Questioning the dominant scientific narrative is another source of Shrivastava’s motivation. “Whatever is being said and whatever the current narrative is, 99% of it is taken for granted, even by top scientists” he said. “And I have seen most of my research and my successes have been when you go beyond what’s taken for granted. You surprisingly see that it was actually easy to make an advance. The only thing was you were blinded by the noise that was going around.” This realization is something he ”would love to teach the students.”
Shrivastava said he doesn’t “like to teach in traditional ways.” Instead, he prefers “to walk with the student, to put myself into the shoes of the student, as in, I don't know anything. And now I'm trying to understand this and walk with them.”
Adopting the student mindset, Shrivastava realized that for students, “one of the biggest problems is motivation that is not obsolete but instead very relevant today. And that is why it's very important to excel in research, because then you can illustrate why they should care about something.”
He gave the example of Huffman coding, an algorithm (or “old school compression technique”) published by David Huffman in 1952. “Why should [students] care about Huffman coding? It's a concept that is a textbook algorithm of the 50s. Everybody will say, ‘We live in the age of LM. What does Huffman coding have to do with what I’m going to do? Why am I learning these algorithms in the age of ChatGPT?’” Using his own research as an illustration, Shrivastava showed “that you can use Huffman coding to actually make large language models 30% more efficient, while mathematically proving that the compression is perfectly lossless. This results in a direct 50% cost savings on LLMs!”
“With this research, we created a new category – lossless compression of LLMs using Huffman coding. When we initially made the paper public on arxiv before its publication, after one day, it was trending No. 1 on Hacker News – all organically!”
Why GPUs?
The inefficiency of LLMs–their unwieldy size and resource-gobbling power needs–is a primary focus of Shrivastava’s research.
“Right now, if you look at AI, it's pretty much synonymous with GPUs,” he said. “And when I looked at the whole history of AI, AI developed in an ecosystem where GPUs were always there.” But GPUs are problematic, “because they're expensive, they are not available, and there is literally a fight to get access to GPUs,” he explained.
“When I started my journey at Rice,” he said, “I started asking this question: Why GPUs? Why should they be synonymous [with AI]? I couldn't find any solid reason for that. I wanted to show people how we can build AI without GPUs.”
It was a bold idea that paid off. “I got tenure based on that research,” Shrivastava said.
The research also led to a startup. ThirdAI (pronounced “third eye,” as in the eye of insight) is the company Shrivastava co-founded with a former student. He said the startup serves as “a classic illustration” for students. “Why should students believe me when I say something? Because look at what I'm doing.”
The company also serves as a counterargument to those who assumed AI could only be done one way. “The fastest way to prove to the current world ” that an alternative is possible is to “create demonstrable value via a startup that is adopted by commercial entities ,” said Shrivastava. “When enterprises see the benefit of it in a real setting, when they see the return on investment, they will be convinced.”
AI and the myth of accessibility
Because of their exorbitant energy consumption and vast storage needs, current deep learning models require very large corporations to keep them going. “Right now, in today's world, there is an illusion that AI is accessible,” Shrivastava said. “Everybody can access ChatGPT, but there are only a handful of companies in the world that can literally control it, and they drive the direction in which it is going.
“If I am a normal individual, the power of AI is not really available to me. What is available to me is what the hyperscalers of the world believe should be available.”
The current state of AI “creates not just a monopoly, but it also creates a privacy issue,” he continued. “If you need AI, you have to give data. I cannot chat without sending my data to a server.”
Shrivastava likens the impact and effect of AI today to that of electricity in the early 20th century. “As a technology, if it's like electricity, it better be available to everyone, and people can switch it on and switch it off and use it,” he said
AI remains inaccessible because of the way it is built, Shrivastava said, but “I don't think that should be necessary. I think there is enough evidence to believe that we can actually make it much, much more accessible by making it cheaper, more available, and private, so it could go into multiple settings.
“While the world may imagine that that will come later, I think it can come sooner, because we don't have to buy into the narrative that you need very expensive infrastructure to build AI,” said Shrivastava, ever the disruptor. “And I think that has a huge societal implication as well in the current world.”