- 4
- June
13 fundamental Claude/LLM terms that everyone in an organization — from executives to developers — should understand together before deciding to buy: the basics (LLM · Pretraining · Fine-tuning · Tokens · Context Window), the answering and tuning that affects UX (Temperature · Latency · TTFT), Anthropic's principles (HHH · RLHF), and connecting Claude into enterprise systems (MCP · MCP Connector · RAG) — this is an original article grounded in concepts from Anthropic's documentation, with added Thai context relevant to Thai organizations (pricing · language · compliance · real use cases).
Group 1 / 4 — LLM Basics: What Claude Is and How It Works
LLM Large Language Model
A large AI model that learns from a massive amount of text. It can write, answer, summarize, translate, and code at a level close to a human · Claude is one such LLM, shaped through pretraining + fine-tuning + RLHF into something that "converses" and acts as an "assistant" — not just a word box.
In the 2026 market the major commercial LLMs are: Claude (Anthropic) · ChatGPT (OpenAI) · Gemini (Google) · Llama (Meta, open-weight) · and from China DeepSeek, Qwen and others — each with different strengths. For a comparison, read ChatGPT vs Claude vs Gemini.
Pretraining
The first step of teaching an LLM — feed it enormous amounts of text and have it repeatedly learn to "predict the next word" until it grasps the patterns of language · The result: a model that "knows language" but is not yet good at following instructions. It must go through fine-tuning + RLHF afterward to become an "assistant".
THAI CONTEXTClaude is pretrained on a large amount of Thai text — meaning it is good at Thai from the "base", not as a bolt-on add-on · It therefore understands Thai context better than an LLM that is pretrained mostly in English and then translates.
Fine-tuning Tuning the model
Training an already-pretrained model further on a specialized dataset so the model "imitates" the patterns of the new data · The Claude we all use is a Claude that has already been fine-tuned into a helpful assistant — not a raw pretrained model.
The Anthropic API currently does not let customers fine-tune Claude (unlike OpenAI, which offers fine-tuning of GPT) — if you are interested, you must talk to Anthropic sales · The more popular alternative: use RAG (read below) + prompt engineering — for most cases this is enough, with no need to fine-tune.
THAI CONTEXTThai customers who want Claude to be "good at something specific" (e.g. Thai law · company SOPs · product catalog) — always start with RAG first · it is safer · it does not change the base model's behavior · and you can update the data instantly (see RAG below).
Tokens
The sub-units an LLM uses to process text · 1 token may be a whole word, a piece of a word, or a character · For English, 1 token ≈ 3.5 characters (≈ 0.75 words).
Why you should care about tokens:
- Billing — the Claude API charges by tokens (input + output)
- Context window — the limit is in tokens, not characters
- Performance — a long prompt is slower and more expensive
THAI CONTEXTThai uses more tokens than English for the same meaning — because Anthropic's tokenizer is not optimized for Thai the way it is for English. Roughly: 1,000 Thai characters ≈ 400–600 tokens (1,000 English characters ≈ 250 tokens) → the cost of Claude for Thai workloads runs about 1.5–2× higher than English for equivalent content · Thai customers planning heavy Claude API usage should budget using the real token ratio.
Context Window
The LLM's "temporary memory space" while answering — not the data used to train it, but the "working memory" of each conversation · A large context window means it can read long documents in one go, use long history, and maintain coherence better.
Claude Opus 4.7 has a context window of 200K+ tokens = it can read a hundred-page contract in a single pass · this is a key difference from older LLMs that were only 8K–32K · For comparison: GPT-4's context window is 128K · Gemini's is 1M (the largest on the market, but with a pricing/quality trade-off).
Note: a large context window does not mean it "remembers the entire conversation" — every call must send the full history each time (the Claude API is stateless) · For how to handle enterprise data safely, read Claude Data Governance & Security for Thai Organizations.
Group 2 / 4 — Answering and Tuning That Affects UX
Temperature
The parameter that controls the "randomness" of answers — a high value → more varied answers, drawing out creativity · a low value → stable, on-point answers · the common range is 0–1 (sometimes 0–2).
Important note: even setting temperature=0 is not 100% deterministic — the same input may yield a different output. If you truly need determinism, you must add extra infrastructure (caching, seed) — Anthropic states this limitation in its documentation.
THAI CONTEXTEnterprise/government customers who need "consistency" for compliance audits — temp=0 alone is not enough; you must add caching/RAG/output snapshots · For creative writing / marketing → temp 0.7–1.0 opens up broader ideas.
Latency
The total time from sending a prompt to receiving the full output · important for chatbots, agents, and real-time apps · Factors: model size, hardware, network, and prompt/response length.
THAI CONTEXTAnthropic's infrastructure is in the US by default — latency from Thailand ≈ 150–250ms (round-trip) · using AWS Bedrock APAC (Singapore) cuts latency by ~50–100ms · for realtime UX (live chat, voice) every 100ms matters.
TTFT Time to First Token
The time from sending a prompt until Claude starts "typing" the first word · different from total latency (TTFT = start, latency = finish) · it matters for UX — users seeing "Claude has started answering" quickly feel good, even if the full response takes another 5 seconds.
Factors affecting TTFT: model size (Opus is slower than Haiku) · prompt complexity · cache hits (see prompt caching in Anthropic's documentation).
Group 3 / 4 — Anthropic's Principles: Why Claude Differs from Other LLMs
HHH Helpful, Honest, Harmless
The 3-H framework Anthropic uses as the goal when training Claude — it is "Anthropic's own framework" that other companies do not use this term for:
- Helpful — Claude tries to do the work / answer as well as it can within its scope
- Honest — it answers truthfully, without hallucinating / confabulating, and admits when it does not know
- Harmless — it does not discriminate, and refuses dangerous/unethical requests while explaining why
Anthropic invests heavily in Constitutional AI research — instead of having people label every output, it uses "principles" (a constitution) for Claude to measure itself against · this is a point of difference between Claude and ChatGPT/Gemini that risk/compliance teams consider when choosing a vendor.
THAI CONTEXTThai executives and legal teams worried about AI giving made-up answers / answering sensitive questions — HHH is the technical reason that clearly answers "why choose Claude" · Claude typically refuses explicit/risky questions (pornography/weapons/medical advice, etc.) more than LLMs from other vendors.
RLHF Reinforcement Learning from Human Feedback
A technique that uses "human opinions" to teach a model to answer the way people prefer · How: people rank 2+ outputs from the model → the model learns to answer closer to what people ranked highly · it is the main technique Claude uses to adjust behavior after pretraining.
Claude (and ChatGPT, Gemini) all use RLHF as a core — the result is AI that "converses" rather than crudely "predicting the next piece of text" · the original research: Anthropic's RLHF paper (arxiv).
Group 4 / 4 — Connecting Claude into Enterprise Systems
MCP Model Context Protocol
An open protocol that Anthropic introduced in 2024 for applications to send context (data, tools, services) to an LLM in a standard format · think of it as the "USB-C of AI" — plug it into anything.
Enterprise use cases: connect Claude to internal databases, file storage, ERP, helpdesk, CRM — without writing your own integration, using existing MCP servers (community + Anthropic + third-party).
Read the dedicated article at Claude MCP — Model Context Protocol for Enterprises.
MCP Connector
A feature in the Claude Messages API that lets customers connect an MCP server without writing their own MCP client · just send requests through the normal API and specify the MCP server endpoint · supports tool calling · currently in beta.
THAI CONTEXTThai dev teams who want to integrate Claude into enterprise systems quickly — the MCP connector cuts integration time from "design your own MCP client" down to "call the normal API" · ideal for a pilot/proof-of-concept before investing in a full integration.
RAG Retrieval Augmented Generation
A technique that lets an LLM "retrieve" data from external sources before answering · How it works: the user asks a question → the system retrieves relevant chunks from a database (vector DB / knowledge base) → those chunks are placed into the context window along with the prompt → Claude reads them and answers with reference to the data.
Benefits:
- Answers with the latest / organization-specific data, without fine-tuning
- Updates data instantly (just edit the database)
- Provides citations — tells you which document the answer came from
- Reduces hallucination because the ground truth is in the context
THAI CONTEXTRAG = the #1 choice for Thai customers who want Claude to be "good at their own organization" such as SOPs, HR policy, product catalogs, or government-specific systems · store data in a vector DB (pgvector, Pinecone, Weaviate, etc.) · safer than fine-tuning because the data does not enter the model · updatable at any time · Grand Linux can help design the RAG architecture + integration.
Summary
These 13 terms are the "foundational words" that let teams in an organization speak the same language about Claude — executives understand Helpful/Honest/Harmless at the governance level · IT teams understand context window / RAG / MCP at the architecture level · procurement understands tokens at the budget level · legal understands pretraining/fine-tuning at the compliance level.
When an organization's teams understand these terms the same way, the decision to buy Claude becomes smoother and more accurate — with no need for a "translator" between dev, procurement, and legal · Next: read Claude AI Pricing — All Plans and Claude Enterprise.
AI terminology is not just a developer's concern — when executives, procurement, and legal understand the same words, decisions about Claude in the organization become far faster and more precise.
- Saeree ERP Team
📖 References
The definitions of LLM, Tokens, Context Window, Temperature, Latency, TTFT, Pretraining, Fine-tuning, RLHF, and RAG are standard concepts in the AI/NLP field, drawn from many textbooks and research papers · HHH (Helpful/Honest/Harmless) and MCP (Model Context Protocol) are frameworks Anthropic introduced — see the source documentation at Anthropic Claude API Glossary · Anthropic's Constitutional AI research: Constitutional AI: Harmlessness from AI Feedback · RLHF research: Training a Helpful and Harmless Assistant with RLHF.
This is an original article by Grand Linux Solution — written from an understanding of these concepts and supplemented with Thai context (pricing · language · compliance · use cases in Thai organizations), not a direct translation of Anthropic's documentation.
Note: Technical details (context window size · token ratio · API features) reflect the state of things in 2026 and may change as Anthropic updates · verify against Anthropic's official documentation before making contract-level decisions.
