SK hynix AI Hackathon 2026 Round 1 Review: Preparation and the Test

I took the first preliminary round of the SK hynix AI Hackathon 2026. This is a review of how I prepared over the five days before it and what I felt while taking it. What was on the test can't be shared publicly, so this post doesn't cover which problems came up. It only covers the published schedule and session structure, my preparation and my impressions.

1. The hackathon and the first round

Here is the schedule from the official site.

You have to pass round 1 to take round 2. Round 1 had two sessions. Session 1 was multiple choice, and in session 2 you solved tasks using the AI assistant built into the test platform. Outside AI tools weren't allowed.

2. Starting without knowing the format

Apart from the one line "basic AI knowledge and ability to use AI", the official materials said nothing about round 1's format. So the first thing I did was keep research notes. I read public prep videos and blogs and the assessment platform's public materials, and wrote down what I had confirmed from the source separately from guesses based on reviews.

What set the direction was how the test is scored.

• The AI-use assessment was known to look at the process through the conversation log with the AI, not just the final answer. In a beta case the assessment platform published, two people with the same result score ended up far apart on the final score because of the process score.
• A news report said the scoring criteria drew on Anthropic's AI Fluency 4D (Delegation, Description, Discernment, Diligence).

So separately from studying the knowledge, I also prepared how to ask AI and how to check its answers.

3. Studying by topic and writing it up

I split the material into topics. For each one I studied with a learning page, did practice questions, went back over what I got wrong and then published a post.

I changed how I studied partway through. While studying evaluation metrics, I got all the terms at once in a table and could barely solve the practice questions. After that I switched to taking one term at a time with a numeric example and checking it with a single question before moving on, and then I could solve them.

4. Check questions and mock exams

I went back over each wrong answer and did one more check question on it.

• Normalization statistics: I picked normalizing train and test each with its own mean. I picked the same option again after the explanation, and only got it right once it was switched to a numeric example. If train has mean 350 and standard deviation 10 but the test period averages 370, a 370-degree wafer is +2 by train's numbers but 0 by test's. The information that it runs hot gets erased. Statistics come from train only.
• Sampling bias vs. benchmark contamination: The situation was test questions leaking into the training data, and I picked sampling bias. I settled on asking "did it see the answers in advance?" first, and only if not, "did it learn only from data from somewhere else?"
• Who carries a legal duty: Under Korea's AI Basic Act the notice and labeling duties fall on the business offering the AI service, but in both the check questions and the final mock I picked the option where the user does the labeling.
• F1: With recall 90% and precision 10%, I picked 10%. F1 is pulled toward the lower value but isn't equal to it: 2 × 0.9 × 0.1 ÷ 1.0 = 18%.

5. Preparing for session 2

A notepad template

Session 2 meant using the platform's AI, so I made a template to paste into questions. Public reviews often said the model could drop conditions given earlier, so I built it on the assumption that the rules get pasted in again every time.

[Rules] 1. Use only what is in the conditions below and the attachment. Otherwise "needs checking".
2. Don't calculate. Formulas only if needed.
3. One table, no explanation, only the rows needed to answer.
4. Cite the evidence by file name and row.

On top of that I added a slot for restating the task's conditions in my own words, and three question templates for pulling information, solving a question and checking evidence.

Rule 3 originally said "a table of at most 10 rows". When I asked why 10, the reason was that long answers get cut off easily and short ones are easier to check. But with more than 10 rows of data the needed rows would get cut, so I changed it from a number to "only the rows needed".

A practice task

I got a short practice task: find the failing chips given chip measurements and shipping criteria. When I first saw it, I asked, "This is overwhelming, what am I supposed to do?" So I worked out the flow first.

1. Read all the questions first and look at their structure: which ones come from the same table, which are arithmetic, which hang on an exception.
2. Restate the criteria in my own words instead of attaching the file. Spell out boundaries as "equal passes".
3. Get just one verdict table from the AI, with a column per criterion.
4. Check only the rows on a boundary or exception against the source, and do the arithmetic myself.

With this flow the AI questions take one or two tries. But I only got as far as seeing the walkthrough for this practice task; I never solved one through on my own.

6. Session 1

There was so much text that it felt closer to a reading comprehension test than basic AI knowledge. It didn't feel hard, but I was short on time.

At first I read everything: the passage, the situation and the options. Once time started running short, I skimmed the passage and worked out the answer by looking at the options. I finished just as time ran out. What I had studied did help.

7. Session 2

At the start, while getting a grip on the tasks, I took some time with an easier problem first to figure out what to ask and how to write it.

The notepad template helped, but the input constraints made it awkward to use as is. Instructions like "output as a table" or "say you don't know if you don't know" did seem good for improving the quality of the answers. On the rule card I had prepared, those are rule 1 (only what is in the material, otherwise "needs checking") and rule 3 (one table).

The AI did get things wrong at times. So I added a verification step and told it to state the evidence for its verification.

I had questions left over, but I ran out of time.

Session 2 was the disappointing part. As I kept going, I came to understand and figure out how to approach it, and that left me with regret. What I figured out was that it is better to pull the information for several problems at once. I should have thought that through and designed my questions around it.

8. What I'd do differently next time

Everything I prepared helped. But it would have been perfect if I'd worked through more expected session-2 problems. I regret not seriously working through session-2 practice problems together with Claude. Even thirty minutes or an hour more would have been good.

If I did it again, when there are different problems inside the same situation, I'd want to work these out first:

• how to ask
• the conditions
• the analysis method
• what to attach as a harness

Working those out first, I'd want to avoid giving the AI the same work twice.

Results haven't been announced yet.

Series: Preparing for the SK hynix AI Hackathon

1. LLM limits and verification
2. Context, RAG, tools and agents
3. Data splits, leakage and imbalance
4. Recall, precision and baselines
5. Prompt injection, jailbreaks and excessive agency
6. Responsible AI principles, copyright in AI output and Korea's AI Basic Act
7. AI Fluency 4D
8. Round 1 review (this post)