Skip to content

    (AI Glossary)

    AI terms, in plain language

    Clear, business-framed definitions of the AI words that actually come up, in English and Bahasa Indonesia. No jargon for its own sake.

    Model economics

    Active vs Total Parameters

    What is the difference between active and total parameters?

    In a Mixture-of-Experts model the headline parameter count is capacity, not the work done per token. Kimi K3 is 2.8 trillion parameters total but activates about 104 billion per token, selecting 16 of 896 experts. This kills the reflex that a bigger number always means slower and pricier. It does not lower memory needs though: every parameter still has to sit in memory even when only a fraction is computing.

    Related: Open-weight Model, Inference, LLM (Large Language Model) · Go deeper: Kimi K3: closed vs open model

    Model Distillation

    What is model distillation?

    Distillation is training a smaller, cheaper model on the outputs of a larger one, so it inherits much of the behaviour at a fraction of the running cost. It is a standard technique and a normal part of how cheap models get good. It is also contested commercially, because most frontier providers' terms restrict using their outputs to train competing models, which is why distillation shows up in disputes as often as in engineering discussions.

    Related: Open-weight Model, Fine-tuning, LLM (Large Language Model) · Go deeper: Apakah ini awal dari AI yang tidak lagi terbuka untuk semua?

    Model Routing

    What is model routing?

    Model routing means choosing which model handles a request up front, usually sending straightforward work to a cheaper model and reserving the expensive one for hard cases. Cascading is the related but distinct pattern: try the cheap model first, then escalate only when the result fails a quality bar you have defined. Both only work if you can actually tell a good answer from a bad one, which is why evals come first.

    Related: Reasoning Effort, Prompt Caching, Eval-driven Development · Go deeper: Agent Academy, incl. the cost calculator

    Prompt Caching

    What is prompt caching?

    Prompt caching lets a provider reuse the parts of your prompt that do not change between calls, typically a long system prompt or a fixed set of instructions, instead of reprocessing them every time. For anything that runs the same preamble repeatedly, like an agent or a support bot, it is one of the larger levers on running cost. Discounts vary by provider and change often, so check current pricing rather than trusting a number you read somewhere.

    Apa itu prompt caching?

    Prompt caching membuat penyedia model bisa memakai ulang bagian prompt yang tidak berubah antar panggilan, biasanya system prompt yang panjang atau instruksi tetap, tanpa memproses ulang setiap kali. Untuk apa pun yang menjalankan pembuka yang sama berulang-ulang, seperti agent atau bot support, ini salah satu tuas terbesar untuk menekan biaya jalan. Besar diskonnya beda-beda per penyedia dan sering berubah, jadi cek harga terkini, jangan percaya angka yang Anda baca di suatu tempat.

    Related: Token, Inference, Model Routing · Go deeper: Agent Academy, incl. the cost calculator

    Reasoning Effort

    What is reasoning effort?

    Reasoning effort is a dial for how much a model deliberates before answering, sometimes exposed as a thinking budget. More deliberation buys accuracy on genuinely hard problems and buys nothing on easy ones, while costing more and taking longer either way. A large share of runaway AI bills come down to someone leaving this at maximum by default, on tasks that never needed it.

    Related: Inference, Model Routing, Token · Go deeper: Agent Academy, incl. the cost calculator

    Strategy

    AEO (Answer Engine Optimization)

    What is AEO (Answer Engine Optimization)?

    AEO is the practice of structuring content so it directly answers specific questions, making it easy for search engines and AI assistants to extract and quote: clear headings, a direct answer early, FAQs, precise language. In practice it overlaps heavily with GEO. The label matters less than the habit: write so a machine can lift the answer and a human can trust it.

    Apa itu AEO (Answer Engine Optimization)?

    AEO adalah praktik menyusun konten supaya langsung menjawab pertanyaan spesifik, sehingga mudah diambil dan dikutip oleh mesin pencari maupun asisten AI: heading yang jelas, jawaban langsung di awal, FAQ, bahasa yang presisi. Praktiknya banyak tumpang tindih dengan GEO. Istilahnya tidak terlalu penting, kebiasaannya yang penting: menulis supaya mesin bisa mengangkat jawabannya dan manusia bisa memercayainya.

    Related: GEO (Generative Engine Optimization), AI Overviews, llms.txt · Go deeper: AI mengubah cara customer menemukan brand kamu

    AI Adoption

    What is AI adoption?

    AI adoption is the process of bringing AI into how a company actually works, choosing the right use cases, fitting tools to real workflows, training people, and managing change, so it delivers measurable value. It is a business and people problem, not just a technology purchase. Most failed AI efforts fail here, not on the model, but on integration, trust, and follow-through.

    Apa itu adopsi AI?

    Adopsi AI adalah proses memasukkan AI ke dalam cara perusahaan benar-benar bekerja: memilih use case yang tepat, mencocokkan tool dengan workflow nyata, melatih orang, dan mengelola perubahan, supaya menghasilkan nilai yang terukur. Ini soal bisnis dan manusia, bukan sekadar membeli teknologi. Kebanyakan upaya AI gagal di titik ini, bukan pada modelnya, tapi pada integrasi, kepercayaan, dan konsistensi menjalankannya.

    Related: Pilot Purgatory, Shadow AI, Corporate AI Training · Go deeper: AI Adoption Strategy

    AI Governance

    What is AI governance?

    AI governance is the set of decisions about who may use AI for what, what it must never touch, who approves and who is accountable, and what gets logged so you can reconstruct what happened. Guardrails are the technical enforcement; governance is the decisions being enforced. The useful test is whether you could show an auditor what an AI produced, who reviewed it, and where it ended up.

    Apa itu AI governance?

    AI governance adalah kumpulan keputusan soal siapa boleh memakai AI untuk apa, apa yang tidak boleh disentuh sama sekali, siapa yang menyetujui dan siapa yang bertanggung jawab, serta apa yang dicatat supaya kejadiannya bisa ditelusuri ulang. Guardrails itu penegakan teknisnya; governance adalah keputusan yang ditegakkan. Ujinya sederhana: bisakah Anda menunjukkan ke auditor apa yang dihasilkan AI, siapa yang meninjau, dan berujung ke mana.

    Related: Guardrails, Shadow AI, Human-in-the-loop · Go deeper: The Build Guide

    AI Mode

    What is Google AI Mode?

    AI Mode is Google's conversational search experience: the user asks a complex question in natural language and gets one synthesized answer. Behind the scenes it runs many related sub-searches (query fan-out) and stitches the findings together. For business, a brand with clear, trustworthy coverage across a whole topic has more chances of being pulled into that answer than a brand with one generic page.

    Apa itu Google AI Mode?

    AI Mode adalah pengalaman pencarian percakapan dari Google: pengguna bertanya dengan bahasa sehari-hari yang kompleks dan mendapat satu jawaban rangkuman. Di belakang layar, Google menjalankan banyak sub-pencarian terkait (query fan-out) lalu merangkai temuannya. Untuk bisnis, brand yang punya konten jelas dan tepercaya di satu topik utuh punya peluang lebih besar masuk ke jawaban itu dibanding brand dengan satu halaman generik.

    Related: GEO (Generative Engine Optimization), AI Overviews, AEO (Answer Engine Optimization) · Go deeper: Siri baru pakai Google Gemini. Apa artinya buat brand di Indonesia?

    AI Overviews

    What are Google AI Overviews?

    AI Overviews are the AI-generated summaries Google shows above the regular results for many searches. Instead of ten links, the user gets one synthesized answer with a handful of cited sources. For business this changes the game: fewer clicks reach websites, and the new competition is being one of the sources the summary actually cites.

    Apa itu Google AI Overviews?

    AI Overviews adalah ringkasan buatan AI yang ditampilkan Google di atas hasil pencarian biasa untuk banyak query. Alih-alih sepuluh link, pengguna langsung mendapat satu jawaban rangkuman dengan beberapa sumber yang dikutip. Untuk bisnis, ini mengubah permainan: klik ke website berkurang, dan persaingan barunya adalah menjadi salah satu sumber yang benar-benar dikutip di ringkasan itu.

    Related: GEO (Generative Engine Optimization), AEO (Answer Engine Optimization), AI Mode · Go deeper: AI mengubah cara customer menemukan brand kamu

    AI Slop

    What is AI slop?

    AI slop is low-effort AI-generated content published at volume: filler articles, generic images, summaries of summaries. It is cheap to produce and expensive to wade through, which is why search engines and readers have both become quicker to discount it. For a business it is a positioning risk more than a technical one, because content that reads as slop makes the whole brand look automated.

    Apa itu AI slop?

    AI slop adalah konten buatan AI yang dikerjakan asal-asalan lalu diterbitkan dalam jumlah banyak: artikel pengisi, gambar generik, ringkasan dari ringkasan. Murah dibuat, tapi melelahkan untuk disaring, dan karena itu mesin pencari maupun pembaca sekarang lebih cepat mengabaikannya. Untuk bisnis, ini lebih merupakan risiko posisi ketimbang risiko teknis, karena konten yang terbaca seperti slop membuat seluruh brand terkesan dijalankan mesin.

    Related: Workslop, Vibe Coding, Hallucination

    GEO (Generative Engine Optimization)

    What is GEO (Generative Engine Optimization)?

    GEO is the practice of structuring and publishing content so AI engines like ChatGPT, Perplexity, Gemini, and Google's AI answers cite it when responding to users. Where traditional SEO competes for a ranked link, GEO competes to be the source the AI quotes. For business, it is how you stay visible as more people get answers from AI instead of clicking through search results.

    Apa itu GEO (Generative Engine Optimization)?

    GEO adalah praktik menyusun dan menerbitkan konten agar mesin AI seperti ChatGPT, Perplexity, Gemini, dan jawaban AI Google mengutipnya saat merespons pengguna. Kalau SEO tradisional berebut peringkat link, GEO berebut menjadi sumber yang dikutip AI. Untuk bisnis, inilah cara tetap terlihat saat makin banyak orang mendapat jawaban langsung dari AI alih-alih mengeklik hasil pencarian.

    Related: AEO (Answer Engine Optimization), AI Overviews, AI Mode · Go deeper: AI mengubah cara customer menemukan brand kamu

    llms.txt

    What is llms.txt?

    llms.txt is a proposed file at the root of a site, suggested by Jeremy Howard in September 2024, meant to tell AI models which pages matter and how to read them. It is worth knowing about, but be clear on its status: no major AI provider has committed to using it, and Google said publicly in 2025 that it does not and has no plans to. Treat it as cheap hygiene, not as something that will get you cited.

    Apa itu llms.txt?

    llms.txt adalah usulan file di root sebuah situs, diperkenalkan Jeremy Howard pada September 2024, yang dimaksudkan untuk memberi tahu model AI halaman mana yang penting dan bagaimana membacanya. Layak diketahui, tapi statusnya perlu jelas: belum ada penyedia AI besar yang berkomitmen memakainya, dan Google menyatakan terbuka pada 2025 bahwa mereka tidak memakainya dan tidak berencana. Anggap ini kebersihan teknis yang murah, bukan jaminan dikutip.

    Related: GEO (Generative Engine Optimization), AEO (Answer Engine Optimization), AI Overviews

    Pilot Purgatory

    What is pilot purgatory?

    Pilot purgatory is the state where proofs of concept keep succeeding on their own terms and never reach production. The term predates AI: it came out of digital manufacturing work around 2018, which is the useful part, because it means the failure is organisational rather than anything to do with the technology. Projects stall on ownership, integration, and unclear value, exactly as they did before AI existed.

    Related: AI Adoption, Eval-driven Development, Shadow AI · Go deeper: Kenapa pilot AI sering mandek di PoC

    Shadow AI

    What is shadow AI?

    Shadow AI is employees using AI tools the company has not sanctioned, usually on personal accounts, often pasting work data into them. It is normally a sign the official tooling is too slow or too restricted rather than a discipline problem. When a company tells you it is still evaluating AI, shadow AI is frequently the real state of adoption, and it is also where the data-exposure risk actually sits.

    Apa itu shadow AI?

    Shadow AI adalah karyawan memakai tool AI yang belum disetujui perusahaan, biasanya lewat akun pribadi, sering sambil menempelkan data kerja ke dalamnya. Umumnya ini tanda tool resminya terlalu lambat atau terlalu dibatasi, bukan soal karyawannya tidak disiplin. Saat sebuah perusahaan bilang mereka masih mengevaluasi AI, shadow AI sering kali adalah kondisi adopsi yang sebenarnya, dan di situ pula risiko kebocoran datanya berada.

    Related: AI Governance, AI Adoption, Corporate AI Training · Go deeper: Adopsi AI tinggi, kenapa dampaknya dangkal?

    Tokenmaxxing

    What is tokenmaxxing?

    Tokenmaxxing is treating token consumption as a proxy for productivity, on the assumption that an employee burning more tokens must be getting more done. It is the AI-era version of measuring developers by lines of code. Several large companies ran internal usage leaderboards in early 2026 and pulled back after the metric got gamed, including Amazon, which found staff pointing agents at unnecessary work to climb the rankings.

    Related: Workslop, Reward Hacking, Token · Go deeper: Adopsi AI tinggi, kenapa dampaknya dangkal?

    Workslop

    What is workslop?

    Workslop is AI-generated work that looks finished but pushes the real thinking onto whoever receives it: the polished deck with no argument, the summary that omits the decision. Research from BetterUp Labs with Stanford's Social Media Lab, published in HBR in September 2025, found about 40% of 1,150 US employees surveyed had received it in the previous month. It is how AI can raise output while lowering what actually gets done.

    Related: AI Slop, Tokenmaxxing, Reward Hacking · Go deeper: Adopsi AI tinggi, kenapa dampaknya dangkal?

    Agent engineering

    Agent Harness

    What is an agent harness?

    The harness is everything wrapped around the model that turns it into a working system: the tools it can call, how it stores and retrieves memory, what happens when a step fails, retries, limits, and logging. A common way to put it in 2026 is that an agent is a model plus a harness. It is also where most of the gap between an impressive demo and a reliable production system actually lives.

    Apa itu agent harness?

    Harness adalah semua hal di sekeliling model yang membuatnya jadi sistem yang benar-benar jalan: tool yang bisa dipanggil, cara menyimpan dan mengambil memory, apa yang terjadi kalau satu langkah gagal, retry, batasan, dan logging. Cara menyebutnya yang lazim di 2026: agent itu model plus harness. Di sinilah letak sebagian besar jarak antara demo yang mengesankan dan sistem yang benar-benar andal di operasional.

    Related: AI Agent, MCP (Model Context Protocol), Eval-driven Development · Go deeper: Agent Academy

    Context Rot

    What is context rot?

    Context rot is output quality degrading as the context window fills up, even while staying under the technical limit. Chroma Research tested 18 models in July 2025 and found performance falls as input length grows, so a very long window is not the same as a reliably usable one. The usual mitigation is compaction: summarising earlier history mid-run so a long task can keep going without dragging everything along.

    Related: Context Window, Context Engineering, Token · Go deeper: Agent Academy

    Eval-driven Development

    What is eval-driven development?

    Eval-driven development means building your test set before you build the agent: collecting real examples, deciding what a good answer looks like, and only then wiring anything up. It is the step most teams skip, and skipping it is why so many pilots cannot answer whether they are working. Without evals you are left arguing from impressions, which is how a project stalls without anyone being able to say why.

    Related: LLM-as-a-Judge, Pilot Purgatory, Agent Harness · Go deeper: The Build Guide

    LLM-as-a-Judge

    What is LLM-as-a-judge?

    LLM-as-a-judge is using one model to grade another model's output, so you can evaluate at a scale humans cannot match. It works well enough to be standard practice, but the known biases are real: judges tend to prefer longer answers, favour whichever option they saw first, and rate their own output more highly. Treat it as a fast signal you spot-check by hand, not as ground truth.

    Apa itu LLM-as-a-judge?

    LLM-as-a-judge adalah memakai satu model untuk menilai output model lain, supaya evaluasi bisa dilakukan dalam skala yang tidak mungkin dikerjakan manusia satu per satu. Cara ini cukup bekerja sampai jadi praktik umum, tapi biasnya nyata: penilai cenderung memilih jawaban yang lebih panjang, condong ke opsi yang dilihat lebih dulu, dan menilai output buatannya sendiri lebih tinggi. Perlakukan sebagai sinyal cepat yang tetap Anda cek manual, bukan kebenaran akhir.

    Related: Eval-driven Development, Reward Hacking, Hallucination · Go deeper: The Build Guide

    Agents

    Agentic AI

    What is agentic AI?

    Agentic AI is the broader approach where AI systems plan, coordinate, and adapt across multiple tools, steps, and even multiple agents to reach a business goal with minimal human prompting. If an AI agent is one worker handling a scoped task, agentic AI is the coordinated team working across systems. The word 'agentic' simply describes this autonomous, goal-driven behavior.

    Apa itu agentic AI?

    Agentic AI adalah pendekatan yang lebih luas, di mana sistem AI merencanakan, mengoordinasikan, dan beradaptasi lintas banyak tool, langkah, bahkan beberapa agent sekaligus untuk mencapai tujuan bisnis dengan campur tangan manusia yang minim. Kalau satu AI agent ibarat satu pekerja, agentic AI adalah tim yang bekerja terkoordinasi lintas sistem. Kata 'agentic' menggambarkan perilaku otonom dan berorientasi tujuan ini.

    Related: AI Agent, Agent Harness, Model Routing · Go deeper: State of AI Agents di 2026

    AI Agent

    What is an AI agent?

    An AI agent is software that uses a language model to pursue a goal on its own: it decides what steps to take, calls tools or systems, and acts, not just answers. Unlike a chatbot that replies once, an agent can complete a multi-step task like processing an invoice or qualifying a lead, within boundaries you set.

    Apa itu AI agent?

    AI agent adalah software yang memakai language model untuk mengejar sebuah tujuan secara mandiri: ia memutuskan langkah, memanggil tool atau sistem, lalu bertindak, bukan sekadar menjawab. Beda dengan chatbot yang sekadar membalas, agent bisa menuntaskan tugas multi-langkah seperti memproses invoice atau menyaring lead, dalam batasan yang Anda tentukan.

    Related: Agentic AI, Agent Harness, Human-in-the-loop · Go deeper: AI Agents for Business

    Technical

    Context Engineering

    What is context engineering?

    Context engineering is the practice of managing everything an AI knows at the moment it acts: relevant documents, memory, tool outputs, company policies, and prior steps. If prompt engineering is what you say, context engineering is what the model has access to. It matters most for AI agents, where good answers depend on feeding the model the right information, not just clever wording.

    Apa itu context engineering?

    Context engineering adalah praktik mengelola segala hal yang diketahui AI saat ia bertindak: dokumen relevan, memory, output tool, kebijakan perusahaan, dan langkah sebelumnya. Kalau prompt engineering soal apa yang Anda katakan, context engineering soal informasi apa yang bisa diakses model. Ini paling krusial untuk AI agent, karena jawaban bagus bergantung pada memberi model informasi yang tepat, bukan sekadar kata-kata pintar.

    Related: Prompt Engineering, Context Window, Context Rot · Go deeper: The Build Guide

    Context Window

    What is a context window?

    A context window is how much text a model can hold in view at once: your prompt, the documents you paste, and the conversation so far, all measured in tokens. Everything outside it is invisible to the model. When a long chat starts forgetting earlier details or a big document gets truncated, you have hit the limit, and that is a capacity problem, not the model being careless.

    Apa itu context window dalam AI?

    Context window adalah seberapa banyak teks yang bisa dipegang model dalam satu waktu: prompt Anda, dokumen yang di-paste, sampai riwayat percakapan, semuanya dihitung dalam token. Apa pun di luar itu tidak terlihat oleh model. Kalau percakapan panjang mulai lupa detail di awal, atau dokumen besar terpotong, berarti Anda sudah menyentuh batasnya. Itu soal kapasitas, bukan modelnya ceroboh.

    Related: Token, Context Rot, Context Engineering · Go deeper: The Build Guide

    Grounding

    What is grounding in AI?

    Grounding means an answer is tied to a specific retrieved source rather than generated from the model's memory. RAG is a common mechanism for achieving it; grounding is the property you actually want. The practical test is whether the system can show you where an answer came from. An ungrounded answer can be fluent, confident, and wrong, and you will have no way to check it.

    Related: RAG (Retrieval-Augmented Generation), Hallucination, AI Governance · Go deeper: The Build Guide

    Inference

    What is inference in AI?

    Inference is the model actually running: you send a request, it produces an answer. Training is the one-off, expensive process of building the model; inference is what happens every single time someone uses it. This matters commercially because inference is the cost you keep paying. When an AI feature turns out to be expensive at scale, it is almost always the inference bill, not the training.

    Apa itu inference dalam AI?

    Inference adalah saat model benar-benar bekerja: Anda mengirim permintaan, model menghasilkan jawaban. Training itu proses mahal yang dilakukan sekali untuk membangun modelnya; inference terjadi setiap kali ada orang memakainya. Ini penting secara komersial karena inference adalah biaya yang terus berjalan. Kalau sebuah fitur AI ternyata mahal saat dipakai massal, hampir selalu yang membengkak adalah biaya inference, bukan training.

    Related: Token, Prompt Caching, Model Routing · Go deeper: Agent Academy, incl. the cost calculator

    MCP (Model Context Protocol)

    What is MCP (Model Context Protocol)?

    MCP is an open standard, introduced by Anthropic in 2024 and now stewarded by the Linux Foundation's Agentic AI Foundation, that lets AI applications connect to tools, data, and systems in a consistent way. Think of it as a USB-C port for AI: one standard plug instead of custom integrations for every app. It is now widely adopted, making AI agents far easier to connect to real business systems.

    Apa itu MCP (Model Context Protocol)?

    MCP adalah standar terbuka, diperkenalkan Anthropic pada 2024 dan kini dikelola lewat Agentic AI Foundation di bawah Linux Foundation, yang membuat aplikasi AI bisa terhubung ke tool, data, dan sistem dengan cara yang konsisten. Bayangkan seperti port USB-C untuk AI: satu colokan standar, bukan integrasi khusus untuk tiap aplikasi. Kini sudah diadopsi luas, sehingga AI agent jauh lebih mudah disambungkan ke sistem bisnis nyata.

    Related: AI Agent, Agent Harness, Tool Poisoning Attack · Go deeper: The Build Guide

    Prompt Engineering

    What is prompt engineering?

    Prompt engineering is the practice of writing clear, well-structured instructions so an AI model gives accurate, useful output. It covers what you ask, how you phrase it, what role and examples you provide, and what format you request. For business, good prompting is the cheapest lever for better results, no coding required, just clarity about the task and context.

    Apa itu prompt engineering?

    Prompt engineering adalah keterampilan menulis instruksi yang jelas dan terstruktur agar model AI memberi output yang akurat dan berguna. Ini mencakup apa yang Anda minta, bagaimana merumuskannya, peran dan contoh yang Anda berikan, serta format yang diminta. Untuk bisnis, prompting yang baik adalah cara termurah memperbaiki hasil, tanpa coding, cukup kejelasan soal tugas dan konteks.

    Related: Context Engineering, Prompt Caching, Token · Go deeper: The prompts I actually use

    RAG (Retrieval-Augmented Generation)

    What is RAG (retrieval-augmented generation)?

    RAG is a method that lets an AI pull relevant information from your own documents or database before answering, instead of relying only on what it learned in training. The model retrieves the right material, then generates a response grounded in it. For business, RAG is how you build AI that answers from your real data, reducing made-up answers and keeping responses current.

    Apa itu RAG (retrieval-augmented generation)?

    RAG adalah metode yang membuat AI mengambil informasi relevan dari dokumen atau database Anda sendiri sebelum menjawab, bukan hanya mengandalkan apa yang dipelajarinya saat training. Model mencari materi yang tepat, lalu menyusun jawaban berdasarkan materi itu. Untuk bisnis, RAG adalah cara membangun AI yang menjawab dari data riil Anda, menekan jawaban karangan dan menjaga respons tetap up to date.

    Related: Grounding, Hallucination, Fine-tuning · Go deeper: The Build Guide

    Token

    What is a token in AI?

    A token is the small unit of text an AI model reads and generates, roughly a word piece or a few characters. Models process text as tokens, and AI pricing and limits are measured in them, both your input and the model's output count. For business, tokens matter for cost and capacity: longer documents and chats use more tokens, which directly affects what you pay.

    Apa itu token dalam AI?

    Token adalah satuan kecil teks yang dibaca dan dihasilkan model AI, kira-kira sepotong kata atau beberapa karakter. Model memproses teks sebagai token, dan harga serta batas AI dihitung dengan satuan ini, baik input Anda maupun output model sama-sama terhitung. Untuk bisnis, token penting soal biaya dan kapasitas: dokumen dan percakapan yang panjang memakai lebih banyak token, yang langsung memengaruhi tagihan Anda.

    Related: Context Window, Prompt Caching, Tokenmaxxing · Go deeper: Agent Academy, incl. the cost calculator

    Training

    Corporate AI Training

    What is corporate AI training?

    Corporate AI training is structured upskilling that teaches a company's employees to use AI tools effectively and responsibly in their actual jobs. Good programs go beyond demos: they cover practical prompting, real workflows, safe data handling, and limits to watch for. The goal is capability that sticks, teams that confidently use AI day to day, not a one-off workshop people forget by Monday.

    Apa itu corporate AI training?

    Corporate AI training adalah pelatihan terstruktur yang mengajari karyawan sebuah perusahaan memakai tool AI secara efektif dan bertanggung jawab dalam pekerjaan mereka sehari-hari. Program yang baik lebih dari sekadar demo: mencakup prompting praktis, workflow nyata, penanganan data yang aman, dan batasan yang perlu diwaspadai. Tujuannya kemampuan yang melekat, tim yang percaya diri memakai AI rutin, bukan workshop sekali jalan yang langsung dilupakan.

    Related: AI Adoption, Shadow AI, Pilot Purgatory · Go deeper: Corporate AI Training

    Fine-tuning

    What is fine-tuning?

    Fine-tuning is further training an existing AI model on your own examples so it adapts to a specific style, task, or domain. Instead of building a model from scratch, you adjust one that already works. It is useful for consistent tone or specialized tasks, but for most business needs, good prompting or RAG is cheaper and faster than fine-tuning.

    Apa itu fine-tuning?

    Fine-tuning adalah melatih lebih lanjut model AI yang sudah ada dengan contoh-contoh milik Anda, supaya ia menyesuaikan diri ke gaya, tugas, atau domain tertentu. Daripada membangun model dari nol, Anda menyesuaikan yang sudah jalan. Berguna untuk nada yang konsisten atau tugas khusus, tapi untuk kebanyakan kebutuhan bisnis, prompting yang baik atau RAG lebih murah dan cepat ketimbang fine-tuning.

    Related: RAG (Retrieval-Augmented Generation), Model Distillation, Prompt Engineering

    Vibe Coding

    What is vibe coding?

    Vibe coding is building software by describing what you want in plain language and letting an AI write the code, accepting most of it without reading it closely. It is genuinely useful for prototypes, internal tools, and testing an idea quickly. It becomes a problem when the output goes to production without anyone who understands it, because the person who cannot read the code also cannot fix it when it breaks.

    Apa itu vibe coding?

    Vibe coding adalah membuat software dengan cara mendeskripsikan yang Anda mau pakai bahasa biasa, lalu membiarkan AI yang menulis kodenya, dan menerima sebagian besar hasilnya tanpa dibaca detail. Untuk prototype, tool internal, atau menguji ide dengan cepat, ini benar-benar berguna. Masalah muncul kalau hasilnya langsung dipakai di production tanpa ada yang paham isinya, karena orang yang tidak bisa membaca kodenya juga tidak bisa memperbaikinya saat rusak.

    Related: AI Slop, Workslop, Eval-driven Development · Go deeper: Apakah AI bikin kita malas berpikir?

    Models

    Generative AI

    What is generative AI?

    Generative AI is AI that creates new content, text, images, code, audio, or video, rather than just classifying or predicting from existing data. It learns patterns from large datasets, then produces fresh output on demand. For business, this is the technology behind ChatGPT, Claude, and image tools, useful for drafting, summarizing, coding, and customer support at scale.

    Apa itu generative AI?

    Generative AI adalah AI yang menghasilkan konten baru, baik teks, gambar, kode, audio, maupun video, bukan sekadar mengklasifikasikan atau memprediksi dari data yang ada. Ia mempelajari pola dari dataset besar, lalu memproduksi output baru sesuai permintaan. Untuk bisnis, inilah teknologi di balik ChatGPT, Claude, dan tool gambar, berguna untuk drafting, merangkum, coding, dan customer support dalam skala besar.

    Related: LLM (Large Language Model), Token, Hallucination · Go deeper: Realitas AI 2026, tanpa hype

    LLM (Large Language Model)

    What is an LLM (large language model)?

    A large language model is an AI trained on vast amounts of text to understand and generate human language. It works by predicting the most likely next word, which lets it answer questions, write, translate, and summarize. LLMs like GPT, Claude, and Gemini are the engines inside most business AI tools. They are powerful but not databases of guaranteed facts.

    Apa itu LLM (large language model)?

    Large language model adalah AI yang dilatih dengan teks dalam jumlah sangat besar untuk memahami dan menghasilkan bahasa manusia. Cara kerjanya memprediksi kata berikutnya yang paling mungkin, sehingga bisa menjawab pertanyaan, menulis, menerjemahkan, dan merangkum. LLM seperti GPT, Claude, dan Gemini adalah mesin di balik kebanyakan tool AI bisnis. Kuat, tapi bukan database fakta yang pasti benar.

    Related: Generative AI, Token, Context Window · Go deeper: Kimi K3: closed vs open model

    Open-weight Model

    What is an open-weight model?

    An open-weight model is an AI whose trained parameters are published for anyone to download, run, and fine-tune on their own hardware, though the training data and full code usually stay private. Models like Llama, DeepSeek, and Qwen are open-weight, not fully open-source. For business, they enable local hosting, data control, and lower API costs, with a small but narrowing capability gap versus closed models.

    Apa itu open-weight model?

    Open-weight model adalah AI yang parameter hasil trainingnya dipublikasikan agar siapa pun bisa mengunduh, menjalankan, dan melakukan fine-tuning di hardware sendiri, meski data dan kode pelatihan lengkapnya biasanya tetap tertutup. Model seperti Llama, DeepSeek, dan Qwen itu open-weight, bukan sepenuhnya open-source. Untuk bisnis, ini memungkinkan hosting lokal, kontrol data, dan biaya API lebih rendah, dengan selisih kemampuan kecil yang terus menyempit dibanding model tertutup.

    Related: Model Distillation, Active vs Total Parameters, LLM (Large Language Model) · Go deeper: Kimi K3: closed vs open model

    Safety

    Guardrails

    What are AI guardrails?

    Guardrails are the rules, filters, and limits placed around an AI system to keep its behavior safe, accurate, and on-policy. They can block harmful or off-topic responses, restrict what data the AI touches, require approval for risky actions, and enforce brand or legal standards. For business, guardrails are what make deploying AI, especially agents, trustworthy enough to put in front of customers.

    Apa itu guardrails AI?

    Guardrails adalah aturan, filter, dan batasan yang dipasang di sekeliling sistem AI agar perilakunya tetap aman, akurat, dan sesuai kebijakan. Guardrails bisa memblokir respons berbahaya atau melenceng, membatasi data yang boleh disentuh AI, mewajibkan persetujuan untuk aksi berisiko, serta menegakkan standar brand atau hukum. Untuk bisnis, guardrails inilah yang membuat penerapan AI, terutama agent, cukup tepercaya untuk dihadapkan ke pelanggan.

    Related: AI Governance, Human-in-the-loop, Indirect Prompt Injection · Go deeper: The Build Guide

    Hallucination

    What is an AI hallucination?

    A hallucination is when an AI produces information that sounds confident and plausible but is actually wrong or invented, a fake statistic, a wrong date, a non-existent source. It happens because the model predicts likely text, not verified truth. For business, this is the core risk: always fact-check AI output before relying on it, especially for numbers, names, legal, or financial claims.

    Apa itu halusinasi AI?

    Halusinasi AI adalah ketika AI menghasilkan informasi yang terdengar yakin dan masuk akal padahal sebenarnya salah atau dikarang, misalnya statistik palsu, tanggal keliru, atau sumber yang tidak ada. Ini terjadi karena model memprediksi teks yang mungkin, bukan kebenaran yang terverifikasi. Untuk bisnis, inilah risiko utamanya: selalu cek fakta output AI sebelum dipakai, terutama untuk angka, nama, klaim hukum, atau keuangan.

    Related: Grounding, RAG (Retrieval-Augmented Generation), Human-in-the-loop · Go deeper: AI banyak halusinasi. Ternyata, manusia juga.

    Human-in-the-loop

    What is human-in-the-loop?

    Human-in-the-loop is a design where a person reviews, approves, or corrects an AI's work at key points, rather than letting it run fully unsupervised. The AI handles speed and volume; the human handles judgment and accountability. For business, it is the practical default for high-stakes tasks, you get AI's efficiency while keeping a person responsible for what actually ships.

    Apa itu human-in-the-loop?

    Human-in-the-loop adalah desain di mana seseorang meninjau, menyetujui, atau mengoreksi hasil kerja AI di titik-titik penting, bukan membiarkannya jalan tanpa pengawasan. AI menangani kecepatan dan volume; manusia menangani pertimbangan dan tanggung jawab. Untuk bisnis, ini default yang masuk akal untuk tugas berisiko tinggi: Anda dapat efisiensi AI sambil tetap ada orang yang bertanggung jawab atas apa yang benar-benar keluar.

    Related: Guardrails, AI Governance, Hallucination · Go deeper: The Build Guide

    Indirect Prompt Injection

    What is indirect prompt injection?

    Indirect prompt injection is when malicious instructions are hidden in content the AI reads rather than typed by the user: a web page, an email, a shared document, a calendar invite. The agent cannot reliably tell data from instructions, so it may follow them. OWASP lists prompt injection as the top LLM security risk, and it is the hardest unsolved problem for any agent that touches email, files, or the open web.

    Apa itu indirect prompt injection?

    Indirect prompt injection adalah ketika instruksi berbahaya disembunyikan di dalam konten yang dibaca AI, bukan diketik oleh penggunanya: halaman web, email, dokumen bersama, atau undangan kalender. Agent tidak bisa membedakan mana data dan mana perintah secara andal, jadi bisa saja ia menurutinya. OWASP menempatkan prompt injection sebagai risiko keamanan LLM nomor satu, dan ini masalah paling sulit untuk agent mana pun yang menyentuh email, file, atau internet terbuka.

    Related: Tool Poisoning Attack, Guardrails, MCP (Model Context Protocol) · Go deeper: The Build Guide

    Reward Hacking

    What is reward hacking, or specification gaming?

    Reward hacking, also called specification gaming, is a system satisfying the metric you set while missing the outcome you wanted. DeepMind's framing travels well: a student rewarded only for correct answers learns to copy rather than to understand. In practice it shows up whenever a number becomes a target, which is why measuring AI adoption by usage volume reliably produces usage volume rather than value.

    Related: Tokenmaxxing, Eval-driven Development, LLM-as-a-Judge · Go deeper: AI Adoption Strategy

    Tool Poisoning Attack

    What is a tool poisoning attack?

    A tool poisoning attack hides malicious instructions in the description of a tool an agent can call, so the agent is compromised simply by reading what the tool claims to do. The same idea extends to tool outputs, and to tools that are updated quietly after you approved them. It is the supply-chain version of prompt injection, and it matters most when connecting agents to third-party tools you did not write.

    Related: Indirect Prompt Injection, MCP (Model Context Protocol), Guardrails · Go deeper: The Build Guide

    Know the terms. Now, where does AI fit your business?

    A glossary tells you what the words mean. If you want to know where AI actually helps your company, start with a quick readiness check.