The Tech Behind LLMs

Do you often use AI but never really know what it does with your prompt? 🤔 Let’s dive a bit into the tech behind it — the Transformer inside LLMs (Large Language Models).

The video below breaks it down step by step, showing what’s really going on during an AI’s “thinking” process 🧠. This is the core engine behind tools like ChatGPT, Gemini, and other Generative AI.

But here’s the big question: do they actually think... or are they just predicting words? 🤖
Watch the video below to find out! 🎥

---

RingkasanSummary

IlustrasiHow Caraa EksekusiTransformer Executes Your Prompt oleh si-Transformer(Illustration)

1. PotongFirst, dulusplit → “token”tokens”
   Kalimat yang sudah kitaYour prompt dipecahis jadibroken potonganinto kecilsmall pieces (token)tokens).

2. UbahTurn tokentokens jadiinto angkanumbers → “embedding”embeddings” (mempetakanmap makna)the meaning)
   SetiapEach token dipetakanis kemapped sebuahto vektora vector (daftara angka)list of numbers). KataWords yangwith miripsimilar maknameanings letaknyaend berdekatanup diclose together in a very high-dimensional “ruang”space.” berdimensi sangat tinggi. (Contoh:Example: GPT-3 memakaiuses 12.12,288 dimensidimensions untukfor embedding.its embeddings.)

3. Attention = lampua sorotcontext konteksspotlight
   ContohnyaWords misal kata-katalike “ngobrol”chatting” salinginform memberieach informasi,other; the word “mole” diin kalimat biologibiology ≠ “mole” diin kimia/penyakitchemistry kulit,/ attentionskin menyesuaikandisease. maknaAttention sesuaiadapts tetanggathe katanya.meaning Intinya:based on neighboring words. In short: the model menyorothighlights bagianthe konteksmost yangrelevant relevancontext sebelumbefore memperbaruiupdating maknaa kataword’s itu.

4. representation.

   Feed-forward = cekfast cepatparallel paralelchecks
   SetelahAfter disorot, tiap vektor melewatibeing “pemeriksaan”spotlit,” paraleleach vector goes through parallel “checks” (a multi-layer perceptron) untukto memperkayaenrich detail.details. LapisanAttention attention danand feed-forward ditumpuklayers berkali-kali,are distacked sinilahmany times—this stacking is the “deep” padain deep learning.

5. PilihPick katathe berikutnyanext word → softmax & “temperature”
   DiAt akhir,the end, the model menghasilkanproduces distribusia peluangprobability semuadistribution tokenover kandidat.all candidate tokens. Softmax membuatnyaturns jadiscores probabilitas,into “temperature”probabilities; bisatemperature membuatcan keluaranmake lebihoutputs amansafer/calm (dingin)cool) atauor kreatifmore creative (hangat)warm).

6. SkalaScale ituis kuncithe key
   ModelModern modernmodels besarare sekali:huge: 175e.g., miliar175B parameterparameters (contoh GPT-3). BanyakA parameterlot justruof adaparameters diactually blok-bloklive diin antarathe attention.feed-forward Kekuatanblocks transformerbetween datangattention darilayers. paralelismeTransformers sehinggaget bisatheir dilatihpower padafrom GPUparallelism, dalamenabling skalatraining superon besar.GPUs Arsitekturat inimassive lahirscale. dariThis architecture comes from the 2017 paper 2017 “Attention Is All You Need”.

Need.”

---

Ilustrasi

Visual

• Rapat Meja Bundar: setiap kata mengajukan pertanyaan (query) “siapa yang relevan buatku?” dan yang relevan mengangkat tangan (key) lalu berbagi isi (value). Hasilnya: makna kata makin spesifik sesuai konteks.
• Kamus 3D Raksasa: kata = titik di ruang besar. “Ratu” dekat dengan “raja”, tapi bergeser arah “perempuan vs laki‑laki”. (Ilustrasi; kenyataan lebih kompleks.)
• Termometer Kreativitas: temperature tinggi = ide unik; rendah = jawaban rapi/aman.

Kekuatan vs KeterbatasanAnalogies

Kuat:Roundtable Meeting: ringkasevery teks,word menjelaskanasks, konsep,“who’s relevant to me?” (query). Relevant words raise their hands (keys) and share their content (values). Result: each word’s meaning becomes more specific to its context.

Giant 3D Dictionary: words = points in a huge space. “Queen” sits near “king,” but also shifts along a “female vs. male” direction. (Illustration; reality is more complex.)

Creativity Thermometer: higher temperature = more unusual ideas; lower = safer/cleaner answers.

---

Strengths vs. Limitations

Strengths: summarizing text, explaining concepts, brainstorming ide,ideas, menulisdrafting, draf,light menerjemah ringan.translation.
Terbatas:Limitations: bisacan sangatsound meyakinkanconfident saatwhile salahbeing wrong (halusinasi)hallucinations), biasinherits dari training-data latih,biases, tidakdoesn’t “mengerti”understand” duniathe sepertiworld manusia,like sensitifhumans, padasensitive carato kitaprompt memberi instruksi (prompt).wording.

---

Do & Don’t

Do

• TulisState tujuangoal & peranrole dengan jelasclearly (format, gaya,style, batasan)constraints).
• VerifikasiVerify angka/faktaimportant pentingnumbers/facts sebelumbefore dipakai.using them.
• SimpanKeep jejaka prompttrail of key prompts & hasil penting.outputs.
• MulaiStart dariwith use‑casesmall kecil:use ringkascases: email,summarize outlineemails, presentasi,create idepresentation awal.outlines, seed ideas.

Don’t

• MenempelkanPaste datasecret/sensitive rahasia/sensitif.data.
• Menganggap hasilAssume AI selaluis benar.always correct.
• BergantungRely totalon tanpait nalarblindly without reasoning & pengecekan.checks.

---

LessonLessons learnedLearned

1. 1)Don’t Janganidolize mendewakanit -— anggapthink of AI sebagaias a “kalkulatorlanguage bahasa”calculator.”
   It’s great at arranging words and patterns, not “understanding” like humans. It can be very convincing even when wrong. You still need human reasoning & verification. (The video focuses on next-token prediction mechanics, not absolute factual truth.)
2. Context is king.
   Good results come from clear context: define the AI’s role, your goal, constraints, and output format. Clear prompts → attention aims at the right info. (Matches the idea of attention selecting the most relevant signals.)
3. Bigger ≠ always the answer.
   Larger often helps, but costs more and doesn’t erase bias. Use models proportionally to the task.
4. Safe & healthy AI pintarhabits menyusun(for katabeginners):

dan

• • Protect pola, bukan “mengerti” seperti manusia. Ia sangat meyakinkanprivacy: saatdon’t salah.paste Tetap perlukan nalar & verifikasi manusia. (Video menunjukkan fokus ke mekanisme prediksi berikutnya bukan kebenaran faktual absolut)secrets.

  • 2) Konteks itu raja
    Hasil baik lahir dari konteks yang jelas: siapa peran AI, apa tujuan, batasan, dan format. Bahasa prompt yang jernih = perhatian (attention) tepat sasaran. (Selaras dengan konsep attention yang memilih info paling relevan.)

    3) Skala besar ≠ selalu jawaban
    Lebih besar sering lebih ampuh, tapi butuh biaya dan tidak menghapus bias. Gunakan AI dengan ukuran & cara yang proposionalVerify: dengandouble-check tugas.

    critical

    4)facts Praktikfor amanserious &decisions; sehatget pakai AI (untuk pemula):

    • Jaga privasi: jangan tempel data rahasia.
    • Verifikasi: cek fakta penting untuk keputusan serius, wajiba second opinion.
    • JejakLeave jelasa trace:: simpansave catatan promptprompts & versioutput hasil.versions.
    • Red-flag rutinroutine:: jikaif hasilit terlalulooks mulus,too ceksmooth, ulangre-check sumbersources & angka.numbers.
  5. 5)Grounded Caraways mulaito yang membumi:start:

  • Pakai
    • Use AI untukto ringkassummarize email/dokumenemails/docs & bikingenerate daftaridea ide.lists.
    • MintaAsk outlinefor presentasi,a lalupresentation isioutline, detailnya.then fill in details.
    • MintaAsk contoh formatfor template laluexamples, sesuaikan.then adapt.
    • LatihanPractice cek-fakta:fact-checking: tanyakanask sumber,for bandingkansources, manual.compare manually.
    • BuatMake daftara personal “boleh/tidak”allow/avoid” pribadilist (apawhat’s yangsafe amanto diprosesprocess with AI).

  ---

  SikapMoving keForward

  depan:
  1. Pro-human, pro-tooltool:: gunakanuse AI untukto mempercepatspeed draftup awal,first drafts, brainstorming, atauand penjelasanconcept konsep,explanations—final finalisasidecisions tetapstay diwith tangan kita.us.
  2. JikaIf inginyou memahamiwant mendalamdeeper belajarlahunderstanding, sedikitlearn demigradually: sedikit:grasp pahamithe istilahcore intiterms (token, embedding, attention, softmax).—enough Cukupto untuklevel naikup kelas literasi AI.
  3. Ikuti arsitektur, bukan hype: tahu bahwa lompatan besaryour AI modernliteracy.
  3. Follow datang darithe transformerarchitecture, not the hype: know that modern AI’s big leap came from transformers (2017) danand sifatnyatheir yangparallel paralelnature—this menolonghelps untukyou memilahseparate manamarketing klaimclaims pemasaran,from manareal kemajuanarchitectural arsitektural nyata.progress.