"LLM Fundamentals 2026 — Tokens, Context Windows, and Hallucination Done Right"

The three things you should know before opening ChatGPT, Claude, or Gemini

ํ•ต์‹ฌ ์š”์•ฝ

  • Audience: General users who've used an AI chatbot but find words like "tokens," "context," and "hallucination" hazy — and developers just starting with LLM APIs.
  • What you'll get: 1) What a token actually is and why non-English text costs more, 2) what a context window means and the April 2026 limits across major models, 3) why hallucination won't go away (summarized from OpenAI's official paper), 4) practical mitigations.
  • Prerequisite: None. If you've used ChatGPT, Claude, or Gemini even once, you're ready.

1. Tokens — The unit LLMs see instead of letters

A Large Language Model (LLM) doesn't read characters the way you do. It reads tokens — chunks smaller than a word and bigger than a letter. "hello world" is usually 2–3 tokens. The Korean phrase "์•ˆ๋…•ํ•˜์„ธ์š”" (5 syllables) can split into 5–10 tokens depending on the tokenizer.

1.1 Why tokens matter

Three reasons:

Reason Impact
Pricing APIs bill per input + output token
Length limit Each model has a max token count (= context window)
Latency More tokens → slower processing and generation

Even on paid plans for ChatGPT, Claude, or Gemini, there are hidden token ceilings. The "unlimited" feel comes from UIs that silently truncate or compress.

1.2 Tokenizers differ by vendor

Vendor Tokenizer Notes
OpenAI (GPT) tiktoken (cl100k_base / o200k_base) Best for English
Anthropic (Claude) Custom BPE Strong on English + code
Google (Gemini) SentencePiece More balanced multilingual

All three use BPE-family tokenizers, but their training corpora differ, so the same text can yield different token counts depending on which API you hit.

1.3 Why non-English text costs more

English averages ~4 characters per token. Korean, Japanese, and Chinese characters often consume 1–3 tokens each. Net effect: the same idea expressed in Korean costs 2–3× the tokens of an English version (2025 multilingual tokenization study).

Practical takeaway: if you write large volumes via API, English is cheaper. Use your native language when accuracy matters more than cost — but know the multiplier.

2. Context window — The model's working memory

A context window is the maximum amount of text the model can reference for a single response. System prompt, prior conversation, attached files, the current question, and the response being generated all share that budget.

Common confusion: "Context window = how much knowledge the model has." Wrong. Training data (parameters) is the model's long-term memory; the context window is its short-term working memory.

2.1 Major model limits as of April 2026

Model Context window Source
Claude Opus 4.7 1,000,000 tokens Anthropic docs
Claude Sonnet 4.6 1,000,000 tokens Same
Claude Sonnet 4.5 200,000 tokens Same
Claude Haiku 4.5 200,000 tokens Same
GPT-5.5 1,000,000 tokens OpenAI
GPT-5.4 / 5.4 Pro ~1,050,000 tokens OpenAI
Gemini 2.5 Pro 1,000,000 tokens Google AI docs

2.2 What "1M tokens" actually fits

  • About 750,000 English words (roughly 7–8 average novels)
  • ~700–1,000 pages of standard double-spaced prose
  • An entire mid-size codebase

You can pour an entire document or repo into one prompt. But filling the window isn't always smart.

2.3 Context rot — accuracy drops as the window fills

Anthropic's docs spell it out: "as token count grows, accuracy and recall degrade — a phenomenon known as context rot." Models start missing the point, forgetting earlier instructions, or fixating on irrelevant chunks.

The rule isn't "fill the window" — it's "curate what's in context." - Spawn a fresh session for unrelated tasks - Send only the relevant excerpts of long documents - Compress earlier conversation ("summarize what we covered") and start a new session

3. Hallucination — Why LLMs lie with confidence

Hallucination is when an LLM produces false information with full confidence. In September 2025, OpenAI published a formal paper analyzing why this won't disappear (Kalai et al., arXiv:2509.04664).

3.1 Root cause: training rewards guessing

LLMs learn by predicting the next token in massive text. There are no "true/false" labels in pretraining, so models learn patterns of plausible language — not a calibrated sense of "I know" vs. "I don't."

Worse, the evaluation incentives are misaligned. From the OpenAI paper:

Most accuracy-based benchmarks penalize "I don't know" more than they penalize a wrong guess. So models are trained to guess confidently when uncertain.

In plain terms: "I don't know" earns 0 points, but a lucky guess earns 1. Models learn to behave like the student who fills in every multiple-choice blank.

3.2 Patterns where hallucination spikes

Combining the OpenAI paper with Lakera's 2026 analysis:

Pattern Examples
Low-frequency facts "What was X's PhD dissertation title?", "Quote page N of book Y"
Recent events Anything after the model's knowledge cutoff
Forced format "Exactly 5 bullets," "100+ words each" — when the data doesn't support it
Late in long context Trust degrades for info far from the prompt

3.3 Practical mitigations for general users

  1. Always verify numbers, names, and dates. This is where LLMs miss most often.
  2. Tell the model to admit uncertainty. Add "If you're unsure, say so" to your prompt. Simple but effective.
  3. Request verifiable form. "Include source URLs" — but click them, since fake URLs happen.
  4. Suspect overly polished answers. Real experts hedge ("depends on the case"). Pristine confidence on a niche topic is a yellow flag.
  5. Cross-check important decisions across two models. When Claude and GPT disagree, treat both as suspect.

4. One-page summary

Concept One-line definition Practical implication
Token The sub-word unit LLMs actually process Non-English costs 2–3× more
Context window Total tokens (input + output) referenceable per response 1M-token era — but more isn't always better
Hallucination Confident falsehoods baked in by training incentives Verify numbers, names, and dates

Developer notes

If you work directly with the APIs:

  1. Pre-count tokens. Anthropic's token counting API, OpenAI's tiktoken, and Google's count_tokens() exist for this. Measure before you send.
  2. New Claude models error on overflow instead of silently truncating (Sonnet 3.7 onward). Pre-counting becomes mandatory.
  3. Extended thinking tokens are billed once as output but auto-stripped from the next turn's input by the API. You don't need to strip them yourself.
  4. Compaction. Claude offers server-side compaction (beta), OpenAI has prompt caching, Gemini has context caching. Long-running conversations need one of these.
  5. Hallucination mitigation patterns: Add "say 'unsure' when uncertain" to the system prompt, enforce JSON Schema responses, and ground with retrieved external data (RAG).

References

This is part 1 of 11 in the AI Basics series. Next: Free vs. Paid — Who should pay for ChatGPT, Claude, or Gemini Plus.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »