Can an Undergraduate Run Large-Scale LLM Experiments?

Can an Undergraduate Run Large-Scale LLM Experiments?

• Interview with Ji-Hyuk Jung from Sogang University

The reason an undergraduate student was able to keep running experiments? Because the cloud made GPU access this easy.

Ji-Hyuk Jung, an undergraduate student majoring in Electronic Engineering at Sogang University, recently participated as first author in a study titled:
“Is ‘Right’ Right? Enhancing Object Orientation Understanding in Multimodal Large Language Models through Egocentric Instruction Tuning.”
The paper was accepted to CVPR 2025, one of the most prestigious international conferences in the field of computer vision.
(CVPR: Computer Vision and Pattern Recognition Conference, held June 2025 in Seattle, USA.)
View paper →

The study examined how LLMs interpret spatial commands such as “right,” “left,” “forward,” and “backward.”
In high-stakes domains like autonomous driving, incorrect interpretations of directional concepts can lead to critical failures—making it essential for models to understand spatial terms just like humans do.

⸻

When resources run short, cloud computing becomes essential

As his experiments progressed, Ji-Hyuk found that his existing computing resources were no longer sufficient to run the scale of models he needed.
This challenge is common in academia—especially during paper deadlines or rebuttal periods—when a short burst of high-performance computation becomes necessary.

“When you’re working against tight deadlines, the key question becomes:
Can I get exactly the resources I need, for just the time I need them?”

“I knew I needed GPUs, but I had no idea where to start.
Then I came across EliceCloud on social media and decided to give it a try.”

⸻

No setup hassles—just start experimenting

Elice Cloud enabled him to launch a GPU instance and begin running experiments right away.
As a locally operated service in Korea, the support process was quick and seamless.

Most importantly, the on-demand pricing model meant that he only paid for what he actually used—making it manageable even for a student working independently.

“GPU costs used to stress me out, but with Elice Cloud,
I only paid for the hours I actually used.
That gave me the confidence to run more experiments.”

⸻

Even 32B-scale models? Yes, solo testing is possible

Thanks to Elice Cloud, Ji-Hyuk was able to run 32B-scale LLMs that would have otherwise been impossible to test on limited personal hardware.
This ability to quickly move from idea to implementation dramatically improved his research productivity.

“The fact that I could try large-scale experiments on my own—that changed everything.
It gave me a sense of ownership over the research process.”

⸻

A cloud environment that doesn’t get in your way

Compared to previous experiences using Colab—where GPU sessions would disconnect if the browser window closed—EliceCloud offered a far more stable environment.
Now, he could leave experiments running and focus purely on results.

“Back then, I was afraid to step away from the screen—even for the bathroom!
Now, I just spin up the instance and focus on the experiment itself.”

⸻

“Students need reasons to keep experimenting.”

“I know how difficult it is for other undergrads to get reliable compute resources.
EliceCloud offers powerful GPUs like A100 80G for just $2–3 USD per hour, which makes it manageable even for students.
Especially when you need to scale up for a week or so, it’s a practical solution for time-critical workloads.
I’d definitely recommend it.”

⸻

Need a burst of compute power for your next paper deadline or rebuttal?
👉 Try Elice Cloud today