News & Events

ICDS News

ICDS co-hire talks DeepSeek AI

Posted on February 4, 2025

Editor’s Note: A version of this article was published on Penn State News.

UNIVERSITY PARK, Pa. — Dana Calacci, assistant professor of information sciences and technology and Penn State Institute for Computational and Data Sciences (ICDS) co-hire, recently spoke to Penn State News about DeepSeek’s approach to generative artificial intelligence (AI).

DeepSeek, an AI assistant that is designed to accomplish a variety of tasks, is advertised to be strong at formal reasoning tasks such as math and logic problems, according to Shomir Wilson, associate professor of information sciences and technology.

“This is a significant claim because many contemporary AI assistants struggle with formal reasoning tasks,” Wilson told Penn State News.

The system recently made headlines for having a fraction of the cost and computing power as some of its competitors.

“I think the approach the DeepSeek team takes is good for AI development for a number of reasons,” Calacci said. “The first is that right now, many models are evaluated against a “global” idea of what a “good” response is to a given query or prompt. This creates biases and makes models less useful for marginalized groups and specific tasks.”

Calacci suggests that DeepSeek allows for more “fine-grained, task-specific” training that is more affordable and faster than current approaches.

“This means that their strategy could be used to make models that, for some prompts, are more accurate or more useful to specific communities. This is an under-reported, interesting benefit of their approach,” Calacci said.

Calacci also notes that DeepSeek can run on more energy-efficient devices, potentially making their large language models more accessible without relying on cloud computing infrastructure.

“DeepSeek’s approach uses half as much compute as GPT-4 to train, which is major improvement,” Calacci added. “The size of the final DeepSeek model also means probably over a 90% reduction in the energy cost of a query compared to GPT-4, which is huge. Right now, GPT-4 queries are run on big cloud server infrastructure.”

Calacci’s research focuses on crowdsourced AI audits and AI harms, data tools for workers, data rights as labor rights and commercial surveillance.

Calacci and Wilson spoke to Penn State News alongside other University AI experts Akhil Kumar, professor of supply chain and information systems, and Mehmet Canayaz, assistant professor of finance.

Share

Related Posts