The question I've committed the next five years to: if you can observe a person's state, and an AI adapts to that person, does their learning and judgment actually get better?
I'm an undergraduate in Bio and Brain Engineering at KAIST, double majoring in Electrical Engineering. I study neural decoding, I build learning systems, and I run my own reasoning system in git. Those look like three interests; they're one question at three levels of observability.
The long version: I think AI should not be an answer machine that replaces judgment, but a challenger — something that surfaces your assumptions, argues back, and expands your thinking. Done right, that enlarges a person's sovereignty over their own thinking. I plan to spend decades on this. So I've broken it into rungs I can actually be wrong about:
One question, many test environments. When I want novelty I change the environment, not the question.
I don't organize this profile as a list of repositories. A repository proves something was built; it doesn't show which problem was worth choosing, or what I changed when I turned out to be wrong. So each item below uses the same grammar:
Problem → Constraints → Decision → Build → Evidence → Counterargument → Update
Problem. In a wireless brain-computer interface, spikes get dropped or falsely detected on the transmission channel. How far can that error rate rise before motor decoding breaks? A 2024 Nature Electronics paper answers this, but the code and methodology were never released.
Constraints. Public data only (CRCNS pmd-1: one macaque, 161 neurons across M1 and PMd, 496 reach trials, 10 ms bins). No lab GPU — a second-hand MacBook Air M1.
Decision. Reproduce the trade-off with a Ridge decoder rather than the paper's spiking network. Two reasons, both stated up front: error propagation stays interpretable, and real-time inference on a low-power implant is actually plausible. This trades peak accuracy for explainability — a deliberate choice, not a shortcut.
Build. 200 ms of history as lagged features (161 neurons × 20 bins = 3220 dims). Spike errors simulated as independent misses and false alarms at equal probability, swept over SER = 10⁻⁴ … 10⁻², five repeats per condition. Scored by Pearson r against measured hand velocity, plus a retention ratio against the error-free baseline.
Evidence. Baseline r = 0.831 — below the paper's spiking decoder (0.932), as expected from a linear model. But retention held above 0.998 even at SER = 10⁻², where the paper's decoder fell to 0.700. The headline isn't "linear wins"; it's that accuracy and error tolerance are separate axes, and a wireless implant may want the second one.
Counterargument. A linear decoder may look robust precisely because it never captured the nonlinear temporal structure that errors would have destroyed — robustness as a side effect of underfitting. Also: one subject, simulated errors rather than a real channel, one dataset.
Update. Next rung: a motor-imagery EEG replication testing whether the same variability–generalization relationship holds across subjects, not just across error rates.
Problem. My AI conversations, projects, and notes kept scattering. What I learned in one place never reached the next decision.
Constraints. One person. Limited and unpredictable time. One laptop, no server. And model churn — whatever I depend on this year gets replaced next year.
Decision. Put the state in git text files, not in a chat history, and treat the model as a
replaceable executor. Every unit of work is a mission file carrying history / decision / next_step;
agents, projects, and repositories live in a registry that CI checks for drift. Automated routines
open branches and pull requests — merges stay human-only, deliberately.
Build. Hudson — the system: agents, registry, routines, decision records. Hunwiki — the knowledge and reasoning corpus it reads from.
Evidence. Daily and weekly routines run unattended and produce reviewable pull requests. Context I used to re-explain every session is now loaded from files. n = 1, running daily.
Counterargument. The structure can outrun the evidence. Document count is not decision quality, and a system that explains everything can absorb its own disconfirming cases. So I log bottlenecks — re-explanation, state hunting, repeated errors — as measurements rather than trusting the architecture.
Update. Memory loading became layered rather than total; reading an entire philosophy to fix a button was the wrong default.
Problem. Students ask three questions no curriculum answers: Isn't this too hard? Why am I learning this? Where would I ever use it?
Decision. Four rules, in use since 2022: explain why before what; design the curriculum backwards from the goal; ask questions instead of giving answers; and when a student is stuck, review the design before blaming the student.
Evidence. A student taught this way ranked first in mathematics after entering a science high school, and kept studying with me online after I moved cities. Rating 5.0 across six reviews.
Counterargument. A handful of self-selected students, outcomes reported by me. Suggestive, not validated.
Update. Dukky exists to test the same loop at a scale where I can't flatter myself:
news → concept check → hypothesis → Socratic counter-questions → reflection. In one week of real
use it produced six friction reports and two bug reports, all of which became shipped changes.
Earlier, FinStudy (five-person team, ROK Air Force startup competition) tested the question
engine itself: multiple choice → why do you think so → misconception analysis.
Problem. I asked my agent system to draft an introduction, pointing it at a page I had pasted as a formatting example — someone else's real academic bio.
What happened. The session read the example as my own history and wrote three channel-specific bios full of another person's degree, lab, and publications. I caught it while reviewing. Nothing was ever published.
Update. The failure wasn't hallucination — every fact was real, just not mine. It was a command interpretation failure: the word "example" was read literally instead of checked. So the rule changed: when a source is marked example / for reference / format only, its content cannot be adopted as fact without an explicit confirmation. The invalid files were kept with a correction banner rather than deleted, so the failure stays legible.
This is what I mean by AI as a challenger rather than an oracle. The evidence that you're using AI well isn't output volume — it's a record of judgments you actually changed.
I'm ambitious about where this goes and precise about where it currently is. I am not an expert in cognitive science or philosophy. I have not built a validated human-learning system, and I have not solved BCI personalization. Today's accurate verbs are explore, build, and update on counter-evidence — and the table above is what I intend to be held to.
Neural decoding, learning systems that adapt to the user, or designing AI as a challenger — I'd like to hear from you, especially if you disagree with something above.
- Essays (Korean): brunch.co.kr/@dbw1gns
- Email: youjhun0225@kaist.ac.kr