Transformer
Transformer
What is a Transformer?
- Grasps all word relationships at once
- Attention-based architecture
- The foundation of modern generative AI
The biggest change is the economics of training. In earlier architectures, computation had to finish on one word before moving to the next, so no matter how many GPUs you added, speed stayed tied to 'sequence.' The Transformer broke that sequential constraint, making it possible to train on hundreds of times more sentences in the same amount of time, and experiments confirmed the scaling laws: performance rises as model size and data grow. That is why today's AI race has become a contest over 'who can afford the bigger compute.'
The architecture splits broadly into an encoder and a decoder. The encoder reads a sentence and compresses its meaning; the decoder generates the next token one at a time. BERT-style models used for search and classification use only the encoder, while generative models like GPT and Claude are decoder-only designs stacked layer upon layer. Transformers are also not limited to language: ViT, which splits images into patches and treats them like tokens, and speech recognition models use the same backbone. Knowing that a Transformer is a general-purpose design for handling sequences, not simply a chatbot, makes adoption discussions much easier.
There is a clear trade-off. Because it computes the relationship between every pair of words, compute grows with the square of the token count: doubling the input roughly quadruples attention cost. That is why, even now that context windows have grown to hundreds of thousands of tokens, pushing a long document in whole is not always cheap or accurate. In practice, teams combine it with RAG, which retrieves only the relevant passages, or with document splitting.
The point where things most often go wrong in practice is the belief that 'a big window means it reads everything.' With long inputs, it is well known that the beginning and end survive while the middle gets blurry, so if you paste in a 200-page contract and ask for the unfavorable clauses, you get a summary that omits a penalty clause buried in the middle entirely. If a reviewer signs that off as 'review complete,' the problem surfaces not in the meeting room but after the contract is done. Put this one line at the very top of your internal guide: length is the limit of what it can hold, not a guarantee that it read it.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayIntroduced in 2017 by Google researchers (Vaswani et al.) in the paper “Attention Is All You Need.” The idea of processing sequences with attention alone, without recurrence or convolution, became the foundation of nearly every large AI model since.
Consider a hypothetical case. Company A, a Korean maker of industrial valves, has 320 employees and annual revenue of KRW 89 billion (≈ US$66,586,900). Its overseas sales team had been running the 140 English quote inquiries it receives each day through a translation tool, and mistranslations, such as rendering lead time as 'lead wire,' kept recurring 22 times a month. In the second half of 2024, at a team lead's suggestion, an 8-week in-house AI training program began.
- In the first session, Team Lead Kim of the overseas sales team put 12 mistranslated emails on screen and asked, 'Why does the machine translate the same word differently every time?'
- The instructor fed the same quote to two tools and showed side by side how results differ depending on whether a tool scans the whole sentence at once or pushes through it one word at a time from the start.
- In week 3, when Lee from the quality team pasted in an entire 40-page specification, the result was a summary that dropped the temperature conditions from the middle 15 pages, and the team saw the limits of the context window firsthand.
- The team then set an internal rule of splitting specifications into clauses before asking questions, and added a double check: the person in charge verifies the amount, delivery date, and quantity directly against the original document.
- After 8 weeks, quote mistranslations fell from 22 to 3 per month, and the average handling time per inquiry dropped from 34 minutes to 12.