Chapter 11: The Colloquium
Dojun had presented to audiences of ten thousand people in his previous life—keynotes at CES, TED talks, United Nations panels on AI ethics—and none of them had made his hands shake the way they were shaking now.
The SNU Computer Science Department Colloquium was held in a seminar room on the fourth floor of the engineering building, a space designed for maybe forty people that currently held sixty-three. Every seat was taken. Students sat on the floor along the walls. Two professors had brought their own chairs from their offices. Someone had propped the door open so that the overflow crowd in the hallway could listen.
This was not normal for a colloquium. Dr. Yoon, who co-organized the series, had told Dojun that the typical attendance was twelve to fifteen people, mostly graduate students fulfilling a seminar requirement and a few faculty members who came for the free coffee.
The free coffee was gone. So were the cookies.
“Quite a turnout,” Professor Kim said, standing at the back of the room with his arms crossed and an expression that landed somewhere between pride and concern. “Word travels fast.”
“Too fast,” Dojun muttered, shuffling his note cards. His presentation was loaded on the projector—twelve slides, clean, minimal text, heavy on diagrams. Hana had helped him design them. (“No bullet points,” she had insisted. “Bullet points are where ideas go to die.”)
“You’ll be fine,” Kim said. “Just present the work. The work speaks for itself.”
“The work is a critique of Hennessy’s branch prediction model. Half the faculty in this room use Hennessy as their primary textbook.”
“Then half the faculty in this room will learn something today. That’s what a colloquium is for.” Kim clapped him on the shoulder—an unusually physical gesture for a man who normally communicated through raised eyebrows and strategic silences. “Go.”
Dojun walked to the podium. Sixty-three faces stared back at him. He recognized Dr. Yoon in the front row, notebook open, pen ready. Jinwoo and Soyeon were in the third row—Jinwoo looking supportive, Soyeon looking like she was evaluating whether his presentation would be worth the time she could have spent on her thesis. Several professors he didn’t know. And at the very back, leaning against the doorframe with his arms crossed and his KAIST hoodie conspicuously out of place among the SNU crowd, was Jang Seokho.
Dojun’s eyebrows rose. Seokho gave a small nod. I came to watch. Don’t disappoint me.
“Good afternoon,” Dojun said. His voice came out steady, which surprised him. “My name is Park Dojun. I’m a second-year undergraduate in the CS department, and today I’d like to discuss a potential limitation in Hennessy and Patterson’s branch prediction model as described in their 2005 paper on speculative execution.”
He clicked to the first slide. A diagram of a modern processor pipeline, color-coded to show the branch prediction unit.
“Branch prediction is the foundation of modern processor performance. Without it, pipelined processors would stall every time they encountered a conditional branch—which happens approximately every five to seven instructions. Hennessy’s model assumes that the Branch Target Buffer entries follow a uniform distribution across the address space. My observation is that this assumption fails under workloads with bimodal branch behavior.”
He advanced to the next slide—a graph showing the distribution of BTB entries under a real-world workload, with two distinct peaks instead of the expected flat distribution.
“This is data from a simulation I ran using the SPEC CPU2000 benchmark suite. As you can see, the actual distribution is bimodal, not uniform. The first peak corresponds to loop-heavy code segments, where branches are highly predictable. The second peak corresponds to irregular control flow—error handling, polymorphic dispatch, exception paths—where branches are essentially random.”
He could feel the room’s attention sharpening. This wasn’t a student presentation anymore. This was a research finding.
“The consequence of the bimodal distribution is that Hennessy’s prediction accuracy estimate—which assumes uniform BTB utilization—overestimates performance by approximately 3-7% for mixed workloads. The predictor works well for the regular peak but wastes resources on the irregular peak, where prediction is futile regardless of the algorithm used.”
A hand shot up. Professor Park from the architecture group, a senior faculty member with silver hair and a reputation for asking the questions nobody wanted to answer.
“You’re claiming a 3-7% error in Hennessy’s model? That’s a significant assertion. What’s your confidence interval?”
“The 3-7% range is across ten SPEC benchmarks. Individual benchmarks vary—gcc shows the largest deviation at 7.2%, while mcf shows only 2.8%. The confidence interval for the aggregate is plus or minus 1.1%. I ran each benchmark twenty times with different random seeds to account for simulation noise.”
“And your proposed fix?”
“A dual-mode predictor. Separate the BTB into two partitions—one for regular branches, one for irregular branches—and apply different prediction strategies to each. Regular branches get the standard two-bit saturating counter. Irregular branches get a smaller, simpler predictor that conserves power rather than pursuing accuracy.” He clicked to a slide showing the dual-mode architecture. “The result is a net prediction accuracy improvement of 2.1% with a 15% reduction in predictor power consumption.”
The room was silent. Not the silence of boredom—the silence of people processing something unexpected.
Professor Park spoke again. “This dual-mode approach—has it been published elsewhere?”
“Not to my knowledge. I searched IEEE Xplore, ACM Digital Library, and the major architecture conference proceedings from the last three years.”
“And you derived this independently? As an undergraduate?”
“Yes, sir. Under Professor Kim Taesik’s supervision.”
All eyes turned to Kim, who was still standing at the back. His expression was carefully neutral, but Dojun could see the tension in his jaw—the look of a man who had vouched for something extraordinary and was waiting to see if the bet paid off.
“The simulation methodology is rigorous,” Kim said. “I reviewed it personally. Park’s work is his own.”
The Q&A continued for twenty minutes. Dojun fielded questions on his simulation parameters, his statistical methods, his proposed implementation. Each answer was careful, precise, and calibrated to demonstrate competence without revealing the depth of knowledge that would raise red flags.
The hardest question came from Dr. Yoon.
“Your dual-mode predictor requires a classification mechanism—something that determines which branches are regular and which are irregular. You hand-wave this in your presentation. How would you implement the classifier in hardware?”
Dojun paused. The honest answer was that Prometheus Labs had solved this exact problem in 2022 using a lightweight neural network embedded in the prediction pipeline. But that technology was sixteen years in the future and relied on manufacturing processes that didn’t exist yet.
“A simple profiling pass during program startup,” he said. “Monitor the first thousand branches, compute the variance of their targets, and assign each to the appropriate partition. The profiling cost is amortized over the program’s execution lifetime.”
“Startup profiling has overhead.”
“Less overhead than the prediction errors it prevents. For long-running workloads—servers, scientific computing—the amortization is negligible. For short-lived processes, you’d default to the standard single-mode predictor.”
Dr. Yoon made a note and said nothing further, which Dojun interpreted as grudging acceptance.
The session ended at 3:45 PM. The room erupted in the low buzz of post-seminar discussion. Several graduate students approached Dojun with follow-up questions. Two professors wanted copies of his simulation code. A PhD candidate from the architecture group asked if he was interested in summer research.
Dojun answered everyone, shook hands, accepted business cards. The whole time, he felt like he was floating slightly above his own body, watching a twenty-year-old perform a role that a sixty-three-year-old had rehearsed for decades.
When the crowd thinned, Seokho was still there. Leaning against the doorframe, arms crossed, with the patient stillness of someone who had been studying everything and everyone in the room.
“You took the train from Daejeon for this?” Dojun asked.
“I take the train for interesting things. This qualified.” Seokho fell into step beside him as they walked down the hallway. “Your presentation was clean. No wasted slides, no filler, no hedging. You stated a problem, presented evidence, proposed a solution. Most PhD students can’t do that.”
“I had help with the slides.”
“The designer? Hana?”
“She has strong opinions about bullet points.”
“Smart woman.” Seokho’s tone was approving—a rare thing. “Your Q&A was more interesting than your presentation. Professor Park’s question about confidence intervals—you had the numbers memorized. Not on your slides. In your head. Down to one decimal place.”
“I ran the simulations. I know my own data.”
“Most people know their conclusions. You know your raw numbers. That’s a different kind of familiarity.” He stopped at the stairwell. “The dual-mode predictor. You said you derived it independently.”
“I did.”
“I believe you. But I also think there’s a version of that solution in your head that’s more elegant than what you presented. You simplified it for the audience. You left out the part that would have made people uncomfortable.”
Dojun said nothing. Seokho was, as always, surgically accurate.
“I’m not going to push,” Seokho said. “We all have things we don’t share. But I want you to know that I see the edges of what you’re hiding, and I’m not the only one.” He glanced back toward the seminar room. “Professor Park noticed too. He was watching you during Q&A the way I watch a chess board when I know there’s a move I’m not seeing.”
“What do you suggest I do?”
“What Kim Taesik is already doing. Build a story fast. Publications, presentations, visible achievements—a paper trail that makes ‘gifted undergraduate’ the obvious explanation. Right now, the gap between your visible credentials and your actual ability is wide enough to drive a truck through. Close it.”
“You’re giving me strategic advice.”
“I’m protecting my investment. I’ve decided you’re worth knowing, Park. If you get discredited or marginalized because your story doesn’t hold up, I lose an interesting conversation partner. That’s unacceptable.”
There it was. Seokho’s version of caring—framed as self-interest, delivered with clinical detachment, but genuine underneath. In forty years, this dynamic wouldn’t change. Seokho would always express loyalty as strategic calculation, even when it was something deeper.
“I appreciate the advice,” Dojun said.
“Don’t appreciate it. Act on it.” Seokho headed for the stairs. “I have a 5:20 KTX to catch. Next time, present something that actually challenges me. This one was only moderately terrifying.”
“Only moderately?”
“I’ve seen you solve five contest problems. Critiquing Hennessy is child’s play by comparison.” He disappeared down the stairwell, his footsteps echoing off the concrete.
Dojun found Hana waiting outside the engineering building, sitting on the steps with her sketchbook open, drawing the cherry blossoms that had been replaced by dogwood flowers in the weeks since spring began.
“How did it go?” she asked without looking up.
“Standing room only. Professor Park asked about confidence intervals. Dr. Yoon asked about hardware classification. Seokho came from Daejeon.”
“Seokho was here?” Now she looked up. “He took the train to watch your presentation?”
“Apparently I qualify as ‘interesting.'”
“High praise from someone who considers most human interaction suboptimal.” She closed her sketchbook. “And? Did you survive the Q&A?”
“Survived. Possibly thrived. Two professors want my simulation code. A PhD student asked about summer research.”
“Park Dojun, academic superstar.” She stood and brushed off her jeans. “I helped you with those slides. I expect acknowledgment in the paper.”
“You’ll get co-designer credit. ‘Slide design by Lee Hana, who believes bullet points are where ideas go to die.'”
“Perfect. Put that in the footnotes.” She fell into step beside him. “I talked to Dr. Yoon today. About the design-CS intersection.”
“How was it?”
“Terrifying and exhilarating. She wants me to audit her graduate seminar on Human-Computer Interaction next semester. As a design student. She said, ‘Design without technical understanding is decoration. Technical systems without design understanding are unusable. I need students who can bridge both.'”
“Bridge.” Dojun smiled. “Like the project.”
“Like the project.” Hana’s eyes were bright with the particular energy of someone who had just discovered that their path was wider than they thought. “Dojun, I think something is happening. Not just the grades or the project or the publication. Something bigger. Like all the pieces are falling into place at the same time, and I can’t tell if it’s luck or if we’re doing something right.”
“Maybe it’s both.”
“Maybe.” She was quiet for a moment, then: “The Bridge prototype. I’ve been thinking about the technical architecture. You said you could build the backend. How long would it take?”
“A basic prototype? Two months if I work nights and weekends. A polished demo? Three months.”
“What if we had it ready by August? Before fall semester starts?”
“Why August?”
“Because SNU hosts a startup showcase in September. Student projects, judged by industry professionals. The winning team gets seed funding and mentorship from the SNU Innovation Center.” She pulled a folded flyer from her jacket pocket and handed it to him. “Fifty teams compete. The top three get funded.”
Dojun unfolded the flyer. SNU Innovation Showcase 2006 — Where Ideas Become Companies. The prize for first place was five million won in seed capital, six months of office space, and mentorship from a panel of tech industry veterans.
Five million won. In 2006, that was enough to incorporate a company, buy servers, and sustain two founders for three months of ramen-fueled development.
“You want to enter Bridge in a startup competition,” Dojun said.
“I want to enter Bridge in a startup competition.” She held his gaze, steady and certain. “As a real project. Not a class assignment. A product that could actually exist.”
In his previous life, this moment had happened four years later, in a different city, with a different product. But the energy was the same—the electric certainty that an idea was worth fighting for, the dangerous, beautiful conviction that two people in a basement could take on the world.
“I’m in,” he said.
“You didn’t even think about it.”
“I didn’t need to.”
Hana studied him for a moment. Then she smiled—not the bright, startled laugh, but something quieter. A smile that was less about happiness and more about trust.
“Okay, partner,” she said. “Let’s build something.”
They shook hands. Her grip was firm, her palm warm, exactly as it had been in the study room three weeks ago when they’d first met. But this handshake meant something different. This wasn’t a group project agreement. This was a declaration of intent.
The sun was low, casting long shadows across the campus. Students streamed past them, heading to dinner, to the library, to the thousand small destinations of university life. None of them knew that on these steps, under a dogwood tree that had replaced the cherry blossoms, two sophomores had just decided to build a company.
Neither did Dojun, entirely. Not a company—not yet. That would come later, when the prototype worked, when the showcase went well, when the pieces aligned. For now, it was just a project. Just two people with complementary skills and a shared conviction that technology should be invisible.
But the seed was planted. And Dojun, who had watched this exact seed grow into a hundred-billion-dollar tree in another life, knew exactly how deep the roots would go.
This time, he would make sure the roots didn’t crack the foundation.