My Final Year Project (FYP) has just passed the halfway mark. After wrapping up the midterm report and sorting through the experimental results, I sat in front of my screen, left with an unsettling sense of cognitive dissonance that took a while to fade.

Before diving into hands-on research, my mental model of research was predominantly mathematical. I was accustomed to starting from well-defined formulations, expecting every step to be backed by a coherent chain of reasoning. I envisioned research as an unbroken deductive thread: hypothesis A holds, which naturally leads to method B, and empirical evaluation C confirms the theory.

Doing actual research, however, threw me into an entirely different kind of fog.

The first clash with reality came from the day-to-day execution of the project.

Things did not unfold as smoothly as I had envisioned. For the most part, our project has operated under a noticeable lack of close supervision. Guidance from the top was sparse; determining the research question, designing the algorithmic pipeline, and troubleshooting roadblocks fell almost entirely on my shoulders. I had to read through the literature, formulate actionable ideas, and break them down into concrete tasks to teach and guide my teammates. Carrying both the strategic direction and the implementation mentorship has been mentally demanding.

What added even more pressure was the nature of the experiments themselves. They are immensely time-consuming, requiring long waiting periods between runs. Even more stressful is their volatility. With the exact same architecture, a different random seed or a minor tweak to a hyperparameter can flip the results upside down. Watching volatile metrics repeatedly shatter our working hypotheses while waiting in uncertainty created relentless psychological pressure.

We went through multiple rounds of tearing down and rebuilding. Starting from a broad, chaotic scope, we were repeatedly humbled by failed runs, forced to pivot our assumptions, and gradually narrowed our problem boundary. Only recently have we finally crystallized the exact problem we want to solve and converged onto a clear, well-defined research direction.

Looking back, this process has been genuinely meaningful. It forced me to develop the autonomy to define problems independently and lead a team through ambiguity. Perhaps this is the necessary rite of passage in real research. We are still giving it everything, aiming for solid academic output that can hopefully land at a strong conference.

The Empirical Fog and the “Post-hoc Narrative Tax”

Yet as the direction became clearer, an even deeper intellectual unease surfaced.

In practical machine learning experiments, that clean deductive thread rarely survives. What actually happens is that a mathematically elegant hypothesis falls flat once implemented in code. To keep the project moving forward, you have to pivot constantly: discarding assumptions, tweaking hyperparameters, and applying small engineering hacks.

You run through dozens of heuristic variations, and eventually, one specific configuration happens to click. The benchmark numbers improve.

Does this mean we understand why it works? Not necessarily. It might have stumbled into a favorable local landscape, or perhaps it merely fit the idiosyncrasies of a particular dataset. You might not even know if the performance will survive a change of data distribution.

What feels even more conflicting is the writing that follows. In a research report or paper, you can rarely state that a finding was pure trial and error. To present a coherent narrative, academic conventions subtly nudge us into constructing post-hoc motivations, occasionally dressing up an empirical heuristic with fancy theoretical terminology to make it look principled.

This is not academic misconduct; the empirical results and code are genuine, and the approach does solve the immediate problem at hand. But it is far from my ideal vision of AI research. We often rely on far-fetched justifications to get things done, while lacking deep mathematical conviction. Having to pay this post-hoc narrative tax just to make research move forward feels deeply alien to me.

Am I Being Too Theoretical?

These past few days, I found myself asking: am I simply being too theoretical?

Before stepping into machine learning, it is easy to assume that theoretical rigor and mathematical depth dictate research success. Yet modern AI remains largely empirical and engineering-driven. Breakthroughs frequently stem from massive compute, sharp engineering intuition, and relentless exploration of empirical phenomena, rather than elegant closed-form proofs.

There is a well-known historical parallel: when James Watt refined the steam engine, the laws of thermodynamics had not yet been formulated. Engineering often outpaces theory. You build a machine that works first, and the mathematical framework follows decades later.

If deep learning currently resides in a similar pre-thermodynamics era, heuristic exploration and trial-and-error are the inescapable realities of the field. Insisting on complete theoretical closure before taking each step can paralyze an empirical project.

Unresolved Questions and a Long-Term Stance

To be completely honest, I have not reconciled this tension, nor have I fully accepted this state of affairs.

Deep down, I still want to pursue serious, foundational work. Even if that work does not yield quick, frequent publications, and even if it demands long, grueling investment with uncertain payoffs, I am willing to commit to it. I am fundamentally a long-termist. Rather than chasing fleeting empirical tricks and manufactured narratives of benchmark gains, I crave solid theoretical foundations, transparent boundaries, and unshakable deductive clarity.

I have even wondered: is it possible that one day, I will become completely disillusioned with AI as a field?

If the discipline remains predominantly governed by unprincipled trial-and-error and post-hoc narrative gymnastics, I might genuinely choose to pivot toward pure mathematics, computational mathematics, or applied mathematics. In those disciplines, a step is written down because it is mathematically true, not because it happened to output a higher number on a particular test set.

Passing the midpoint of my FYP, the most valuable outcome is not just finding a viable direction to push toward a top conference, but the stark clarity it gave me regarding my own research taste and identity. This essay does not aim to offer a tidy resolution; it simply records my honest struggle, pressure, and aspirations at this crossroad. Wherever I go next, I still want to pursue work that brings genuine intellectual conviction.