I’ve recently been looking back at the news from 2023, when everyone was touting prompt engineers as the next high-paying opportunity in Silicon Valley, and even the best path for liberal arts majors to enter the tech industry, like this article: link
It was content like this that prompted me to try transitioning from the financial industry to AI, little did I know I’d get so deeply involved.
Looking back at the summer of 2023, it was the year when ‘Prompt Engineers’ were elevated to mythical status. At that time, observing how Midjourney in its early days required extremely precise, almost mystical keywords to generate usable images, I too once thought: perhaps this is the future—an art where humans use ‘intuition’ and ’experience’ to explore the boundaries of models. We would accumulate human experience to find the ‘correct way to open’ this black box, figuring out how to awaken this grand magic box to get what we needed, and with significant variance at that.
But standing here today in 2026, this myth has been completely shattered.
- Good Prompts Are Not Talent, But Model ‘Deficiencies’
Back then, we thought that being able to write good prompts was a rare talent; now, it appears to have been merely compensatory behavior resulting from imperfect model training distributions. I’ve recently been re-reading CLIP (Contrastive Language-Image Pre-training), which is a gold standard for combining images and natural language, still useful five years later. Yet, some tricks were considered model deficiencies, and these deficiencies, as mentioned in the papers at the time, should have been addressed through the evolution of model architecture, not prompt engineering. Prompt engineering back then seemed more like a reluctant measure to optimize performance.
In the early stages of architecture, models’ understanding of semantics was full of noise. The reason we wrote ‘a photo of…’, ‘4k resolution’, ‘hyper-realistic’ was essentially to manually perform ‘feature calibration’ for the model. We weren’t creating; we were helping the model find the few correctly labeled coordinate points within that distorted high-dimensional manifold.
With the brutal breakthroughs in computational power and advancements in data cleaning techniques, models now possess powerful semantic alignment capabilities. When models can accurately understand ’natural language,’ those deliberately piled-up ‘keyword spells’ become worthless. As models get smarter, human ’tuning experience’ depreciates.
- Back to Basics: Why Underlying Architecture Is the Truth?
When ‘spells’ are no longer mysterious, the real moats are revealed. Instead of meticulously studying ‘how to talk to the model,’ it’s better to understand ‘why the model talks that way.’ For example:
- Next-token Generation: Understand the essence of autoregressive models, and you’ll understand why models hallucinate and why ‘CoT’ enhances logic.
- Bi-encoder and Cross-encoder: Understand the alignment and interaction of vector spaces, and you’ll know why some retrieval tasks (RAG) are never precise and why CLIP can achieve zero-shot learning.
This understanding of underlying architectures is the hard knowledge that transcends model iteration cycles. When you grasp the underlying logic, you won’t feel anxious when Midjourney releases V7 or Gemini releases a new version, because you know that no matter how the surface changes, the mathematical essence remains the same.
Additionally, Professor Hung-yi Lee’s courses also strive to identify and teach what content will remain crucial for years to come in this era of rapid development. I believe this embodies the spirit that a true professional should have. I am also very grateful for his videos over the past few years, which have given us the correct intuition about LLMs.
- The Collective Anxiety of Startups: Your Feature Is Just Someone Else’s ‘Afterthought’
This is also a significant warning for job seekers and entrepreneurs. In early 2024, countless startups invested heavily in developing prompt-based applications (e.g., helping you write resumes, edit images), or even connecting to APIs to analyze financial information. The result? When Claude or Gemini updated a small feature, these startups were often ‘dimensionally crushed’ within a day.
Because if your moat is built upon others’ ‘imperfections,’ you will disappear when they become perfect.
- The Future’s ‘Golden Combination’: Underlying Understanding + Domain Expertise
So, in 2026, what kind of talent is truly valuable? The answer is: people who possess a deep understanding of underlying AI architectures and can combine it with profound domain knowledge.
If you understand law and comprehend how AI retrieves case precedents through vector spaces, you can develop tools that professional lawyers would genuinely trust and use.
If you (truly) understand finance and recognize the limitations of Transformers in processing time series and structured data, you won’t blindly trust the predictions given by models. A technological moat is never about cutting-edge ’trendy knowledge’ that can be easily superseded, but rather a deep coupling of underlying logic with real-world professional domains. This combination is the moat that brute-force computational power cannot easily breach.