Yufei Tao, a fourth-year PhD student in computer science at Maseeh College, spent his teenage years in Wuhu, China learning C++ the way most people learn dead languages—from books, without ever using it. Traditional cultural attitudes in his family kept computers out of children's hands before college, prioritizing academic focus over digital exploration. Tao had programming manuals but no machine to test a single line of code. He studied if-statements and loops the way a linguist might study Sanskrit grammar, memorizing syntax without ever executing it. "I was literally reading it," he remembers. "I was just purely trying to understand, like, how the code is working and how they flow."
Learning a programming language without programming seemed absurd, like studying French conversation from textbooks without ever speaking to another person. But the limitation forced Tao to focus entirely on how code structures logic rather than simply what it does. "I was more learning the logic of like, how the code is working and how they flow," he explains. That early training in pure abstraction—understanding systems before using them—would shape everything that followed: the undergraduate years spent struggling with tools he didn't yet comprehend, the master's degree revelation when suddenly "everything I learned before makes sense," and now doctoral research focused on making AI systems interpretable. Tao's work in natural language processing centers on a question born from those early years of learning without doing: not just how to build language models that work, but how to understand why they work and make that understanding accessible to others.
The Cost of Independence
Coming to the United States at 18 meant navigating American higher education without a guide. Tao grew up the youngest of eight children in a family that ran land restoration contracts in Wuhu, China's smallest city (though still roughly the size of Seattle). The family had resources by then—enough to send Tao to a Florida university and an older sister to Milan—built from an earlier computer school venture that evolved into successful contracting work. But having tuition money and knowing how to navigate a foreign university system proved entirely different challenges. Tao and his sister both became first-generation college students figuring out American academia alone.
The undergraduate years brought isolation compounded by language barriers and a reluctance to ask for help. Tao learned programming tools—object-oriented design, data structures, algorithms—without understanding how they fit together or what they were actually for. "I literally, I didn't really understand, like, even a simple, you know, concept, like object-oriented programs," Tao recalls about those early years. "I couldn't get it." The experience felt like being handed construction equipment without blueprints or context. Tao had learned C++ syntax as pure logic back in China, but American computer science education assumed hands-on familiarity that first-generation students from different educational cultures often lacked.
The master's program at a different university brought revelation. Suddenly the disconnected tools from undergraduate coursework snapped into coherent systems. "Oh my God. Now everything that I learned before certainly makes sense," Tao remembers thinking. The shift came from seeing programming at different scales simultaneously—understanding both the granular mechanics and the larger architectural patterns. "The bachelor's program is more like learning how to build the school. You have the basic concept, and then a master degree is more like using those tools," Tao explains. The metaphor captures something essential about expertise: you can't see the picture when you're still learning to identify the pieces. During the master's program, Tao met Ameeta Agrawal, who would later become his PhD advisor at Portland State, and discovered natural language processing—research that centered on understanding how systems work and fail rather than simply building them.
Why Portland State
When choosing doctoral programs, Tao prioritized fit over prestige, looking for something harder to quantify than rankings or reputation. "It's the diversity and then inclusivity [that matter], because I'm also a part of LGBTQ community," Tao explains about selecting PSU. Finding an advisor who treated students as intellectual peers rather than subordinates, and being in Portland with its progressive culture, sealed the decision.
Agrawal's down-to-earth mentoring style shaped how Tao now works with the two master's students he's guiding toward publication. "I don't consider myself a mentor or more. I just want to consider myself like a peer," Tao says, echoing the collaborative approach he learned from his advisor. The distinction matters in a field where hierarchy can stifle the kind of open questioning that leads to insight. Tao's research happens within the Compassionate Computing Lab (CoCo Lab), whose mission reflects these values: understanding and developing computational technology so that it serves diverse communities.
Tao's current research focuses on making large language models more interpretable—understanding not just that they work, but how and why they generate particular outputs. The work involves analyzing why models hallucinate false information, how they handle context, and whether their responses align with human values. "How do we make this attribute to answer more human or more naturally aligned with what is humanistic?" Tao asks, framing the central challenge. AI systems can now generate fluent text on virtually any topic, but understanding their internal logic remains opaque. Tao's research aims to open that black box.
The Publication Gauntlet
The career trajectory for PhD students in computer science increasingly resembles a high-stakes lottery, with success determined as much by timing and reviewer assignment as by quality of work. Five years ago, top AI conferences received roughly 5,000 paper submissions annually; last year, that number hit 30,000. The explosion in submissions means competing not just against other universities but against researchers globally, all racing to publish incremental advances before someone else does.
The process demands reading hundreds of papers to identify gaps, designing experiments to fill those gaps, hoping for ethical reviewers who judge work fairly rather than protect their own research territory. Tao has managed to publish in several top-tier conferences, the kind of record that leads to industry job offers. But the strain shows in his description of the current landscape—constant pressure to publish, shrinking acceptance rates, outcomes that feel increasingly arbitrary despite rigorous work. What people see—the publications, the job offers—obscures the grind that produces them: the late nights reading papers, the experiments that fail, the revisions responding to reviewer comments that sometimes contradict each other, the psychological toll of having your work's fate determined by anonymous strangers who may or may not have read it carefully.
AI as Tool and Threat
Despite working at AI's cutting edge, Tao holds nuanced views about its impact on education and work, rejecting both utopian enthusiasm and apocalyptic fear. He advises incoming students to use AI tools like ChatGPT without shame, treating them as personal tutors that can explain concepts in multiple ways until something clicks. "I think it's important to use the AI as a tool, like a counselor, instead of doing something for me," Tao explains, drawing a line between AI as learning aid and AI as replacement for thinking.
The distinction matters because over-reliance creates what might be called cognitive debt—the illusion of understanding without the mental structures that support genuine comprehension. Tao compares the current moment to when calculators became ubiquitous in mathematics education: some warned that students would lose arithmetic skills, others argued that calculators freed students to focus on higher-level concepts, and both predictions proved partially true. The question now is whether AI tools will follow a similar pattern or represent something more fundamental, a shift not just in available tools but in what counts as knowledge itself.
Privacy concerns add another layer of complexity. Tao points to Sam Altman's acknowledgment that everything said to ChatGPT can potentially be used in court proceedings—a reminder that convenience often comes with surveillance, that the AI assistant helping with homework is also harvesting data. The legal and ethical frameworks haven't caught up to the technology's capabilities.
Jobs will shift rather than disappear, Tao predicts, but the shift requires expertise that AI can't yet provide. AI can generate code or text that's 80% functional, but fixing that last 20%—making it actually work in complex real-world contexts—demands deep understanding. The future belongs not to people who use AI blindly, but to those who understand its limitations well enough to correct its mistakes.
Freedom in Freefall
Outside the lab, Tao seeks literal escape from constraint. With roughly 800 skydives logged, Tao holds an instructor rating and describes the appeal with characteristic precision before immediately qualifying it. "I just kind of like, like to feel the freedom, like falling," Tao explains, then adds the scientist's caveat: "But actually, you only get 40 seconds of free fall. So it's not that much." The comment captures something essential about Tao's personality—enthusiasm tempered by accuracy, romance grounded in measurement.
Recently, Tao has taken up scuba diving with his husband, exploring a different kind of weightlessness underwater. The couple is also preparing for parenthood; they've matched with a surrogate and expect their baby to arrive in September. The timing means Tao's dissertation defense and graduation will likely coincide with becoming a father, layering personal and professional transitions in ways that make the immediate future both exhilarating and daunting.
What Comes Next
Short-term plans are clear: finish one more publication, defend the dissertation, graduate. Longer-term trajectories remain open, shaped by competing values that don't resolve neatly. Industry offers more computational resources and salary, the ability to work on AI systems at scales academic research can't match. Tao plans to move into industry initially, partly to see what's possible with more powerful infrastructure, partly because becoming a parent shifts priorities around financial stability.
But Tao hasn't closed the door on academia. The collaborative culture at PSU, the satisfaction of mentoring students through their first publications, the intellectual freedom to pursue questions that interest you rather than questions that serve corporate goals—these carry their own appeal. The choice involves familiar trade-offs between autonomy and resources, between research you control and research you can actually execute at scale. For now, Tao is holding both possibilities open, unwilling to foreclose options before understanding what industry work actually entails.
Reflecting on barriers in general, Tao returns to a theme from those early years learning C++ from books without ever running the code. External obstacles matter less than internal drive, Tao suggests—having access to computers wouldn't have made him a better programmer if the desire to understand how code works wasn't already there. Removing barriers helps, but wanting to understand despite barriers makes the difference.
The observation applies to Tao's current research on AI interpretability. Making language models understandable to non-experts requires more than better technical tools—it requires people who care enough about explanation to do the hard work of translation, who remember what it felt like not to understand, who treat understanding itself as the goal rather than simply using what works. Someone who learned programming as pure logic before ever touching a computer, who struggled through undergraduate coursework without comprehension before everything clicked in graduate school, who now works to make AI's black boxes transparent—that person understands something essential about the relationship between barriers and insight. Sometimes you need to learn the language before you can speak it, and sometimes speaking it isn't the point. Understanding how it works, how meaning gets structured, how systems encode logic—that's where the real work begins.