Teams That 'Use' vs. Teams That 'Understand' AI
The quality of product decisions differs significantly between teams that just adopt AI tools and those that understand their underlying principles. Rinda shares our journey of diving into transformer architecture, what we discovered, and how it fundamentally changed our product design.
Teams That 'Use' vs. Teams That 'Understand' AI
🔑 Key Takeaway (TL;DR) The real difference in AI adoption strategy lies in 'using' vs. 'understanding' the tools. Teams that understand the practical mechanics of LLMs can diagnose output errors and make high-quality product decisions. You don't need to write code to benefit—mastering three core concepts is enough to shift your AI strategic approach.
If your team is struggling with an AI adoption strategy, you've likely experienced this: Have you seen those "Intro to LLM" videos hitting over 300,000 views? At first, we just scrolled past them. But looking closer, we realized that number wasn't just about general content popularity. Are you one of the many who use AI tools daily but constantly pause, wondering, "Why on earth is it giving this response?" We believe that specific anxiety is what drove those 300,000 people to those videos.
What Does It Mean to 'Understand' AI Adoption?
To be honest, our team started by just plugging in the ChatGPT API. We thought, "If we put in a prompt, something comes out, and we use that—done." Adopting the tool itself wasn't hard. The problem arose later. When outputs became erratic or inconsistent with the same inputs, we had no idea where the issue lay or how to explain it.
This is where the line between "Using Teams" and "Understanding Teams" is drawn:
- Users: When the result is weird, they dismiss it as "the model being glitchy."
- Understanders: They immediately question, "Is the tokenization method unsuited for this domain?" or "Is the context window hitting its limit?"
This single difference completely changes the quality of product design decisions.
Borrowing from Gartner's AI maturity model, most teams are stuck somewhere between 'Awareness' and 'Adoption.' To reach true 'Integration,' your team must be able to explain why the tool behaves the way it does.
Practicing LLM Literacy: Why We Read the Transformer Paper
Our motivation was simple. While building automated export AI features, we faced recurring instability in LLM outputs. We blamed the prompt, then the model version—yet nothing fixed it.
Then someone asked: "Are we failing to distinguish between a bug and the nature of a probabilistic model?" That question launched our study. When we first opened Vaswani et al.'s 'Attention Is All You Need' (2017), it felt impossible. We questioned why we were reading something filled with complex math.
However, as we unpacked tokenization, embeddings, and positional encoding, practical intuition emerged. We finally grasped, "Why does the output degrade when we provide long documents?" From a practical LLM perspective, introductory courses provide intuition, but that is distinct from actual implementation capability. Acknowledging that gap is crucial.
Three Ways LLM Literacy Changed Our Export AI Architecture
Here is how our theoretical study connected to real-world product decisions:
Adopting Chunking Strategies. Before understanding context window limitations, we were feeding long export documents in their entirety. Post-study, we implemented logical chunking based on semantic units, drastically improving output stability.
Shift in Model Selection Criteria. Benchmarks aren't everything. We started evaluating models based on their alignment with export domain terminologies and document structures, and their potential for fine-tuning.
Changing Error Handling. Knowing LLMs behave probabilistically, we redesigned our retry logic. It moved beyond simple "re-triggering" toward designing conditional branching.
However, a word of caution: moving toward Agentic AI (multi-step reasoning, tool-calling, memory management) is a much higher hurdle than simply understanding LLMs. Simply using a framework like LangChain does not mean you can design an agentic system. We are still learning this ourselves.
Beyond Prompt Tips: What You Actually Need
Every day, feeds are flooded with "Master LLMs in one book" or "DM me for the link." I find the mass interest more intriguing than the content itself. That anxiety—using AI daily without knowing how it works—is real.
However, be realistic about the "mastery" an introductory book provides. Given the complexity of the Transformer architecture, an intro book offers intuition, not expert-level engineering. If you feel the limits of ChatGPT prompts but cannot explain them, your expectations will consistently be misplaced.
For export managers and sales leads, you don't need to write code. You just need to master these three points:
- Why outputs can be unstable.
- Under what conditions you should not trust the AI.
- When fine-tuning adds value.
If you have that practical intuition, your decision-making changes. If you want a deeper dive, Stanford's CS224N lectures or the Hugging Face official documentation are far more effective than "DM for link" schemes.
Why Technologically Literate Teams Build More Trust
We continue to study for a simple reason: teams that understand the technology can be more honest with their customers. We can be transparent about what AI excels at and where it falls short. A team that says, "You shouldn't blindly trust this output in this specific condition," earns more long-term credibility.
We are currently digging into RAG (Retrieval-Augmented Generation) pipeline optimization and agentic workflow design. We are exploring how to make LLMs handle real-time buyer information in the export domain—it's a complex, evolving puzzle, and we'll keep sharing our progress.
Author · RINDA Export Sales Research Team (Researchers in Buyer Prospecting & Export Automation)
Based on data from over 200+ Korean export companies and internal Rinda platform insights, we edit strategies and checklists for practical use in export operations.
At Grinda AI, we embody the integration of technical literacy and export automation. RINDA handles international buyer prospecting and cold email automation, while Grinda AI explores the wider landscape of AI export tools. If you're curious about how we solve these problems, feel free to reach out.
Q. How deep does my understanding of LLMs need to be?
A. You don't need to be able to code. Grasping "what a context window is," "that outputs are probabilistic," and "when fine-tuning is necessary" is sufficient to improve your strategic decisions. Anything beyond that is depth reserved for those building the models themselves.
Q. Where should I start studying Transformers?
A. We recommend the Stanford CS224N lectures and the Hugging Face NLP course. Read the original Vaswani et al. paper after you have a grasp of the fundamentals; reading it cold can be quite daunting.
Q. How is Agentic AI different from just understanding LLMs?
A. LLM literacy is your base, while Agentic AI design is a specialized capability built on top. Multi-step reasoning, tool-calling, and memory management are design challenges that go far beyond just "knowing how to use" an LLM. We are still learning this area ourselves, and confusing the two often leads to misaligned product expectations.