AI FOUNDATIONS
AI is a family of methods, not one kind of tool
Rules and expert systems
Follow explicit if–then rules written by people. They are predictable for narrow policies and calculators, but do not learn new patterns unless someone updates the rules.
Predictive machine learning
Learns statistical relationships from examples to classify, rank, forecast, or recommend. The answer is usually a score or category, not a written explanation; performance depends on representative, well-maintained data.
Language models (LLMs)
Learn patterns in text and generate a response token by token from the prompt and conversation context. They can draft, explain, translate, summarize, and answer questions, but can state plausible errors confidently.
Reasoning models
Spend additional computation on multi-step problems, checking constraints, or planning. Extra reasoning can improve hard tasks, but does not add missing facts or guarantee a correct result.
Multimodal and vision models
Process combinations of text, images, audio, video, or documents. They map inputs into internal representations and answer across them; small text, charts, motion, and ambiguous images can still be misread.
Image and video generators
Create or edit pictures and clips from instructions or references. Many image systems iteratively refine a latent representation; outputs can contain visual mistakes, odd hands, or inaccurate text.
Speech models
Transcribe speech, identify speakers, translate speech, or synthesize a voice. Accuracy depends on the language, accent, microphone quality, and background noise.
Embeddings and retrieval (RAG)
Convert text or other content into vectors so a search system can retrieve relevant passages from a trusted library. A language model then uses those passages to answer. Better grounding helps, but the source or answer may still be incomplete.
Agents and tool-using systems
Combine a model with instructions and tools in a loop: choose a step, request a tool, inspect its result, then continue. The application executes the tool. People should approve consequential actions and check the final result.
FROM PROMPT TO ANSWER
How an AI answer is produced
A simple model of the process: an assistant does not simply look up an answer in a perfect database. It generates from learned patterns and the context or tools available in that session.
1 · Training
A model is optimized on examples so its internal parameters capture useful statistical patterns. Training gives it broad capabilities, but not guaranteed truth or live awareness.
2 · Instructions and context
Your prompt, earlier messages, attached files, system rules, and any retrieved passages form the context. Clear goals, audience, constraints, and examples make the task easier to follow.
3 · Input encoding
Text is divided into tokens. Images, audio, and video are encoded into model-readable representations. Context limits and supported input types depend on the model.
4 · Generation and tools
The model predicts output tokens. A reasoning setting may allocate more work. If tools are enabled, it can request a search, calculation, or function; the software runs it and returns the result for another model step.
5 · Review and use
The assistant returns text, media, structured data, or a tool result. Check sources, numbers, language, and missing assumptions. A human remains accountable for decisions and external actions.
FUNCTIONS / PLAIN LANGUAGE
What the assistant functions actually do
Chat and writing
Draft, rewrite, explain, translate, summarize, and answer questions from the prompt. The model predicts a likely continuation; it does not automatically verify every claim.
Reasoning or thinking mode
Uses more inference effort on complex logic, planning, or code. It can take longer and still make mistakes; use a small test set to see whether the extra effort helps.
Web search
Looks up current pages and may attach source links. Search can be useful for recent facts; check that each citation supports the sentence beside it.
Deep research
Plans a multi-step search across sources and synthesizes a report. It is for broad questions that need research, not a substitute for checking critical sources.
Files, images, and vision
Reads supported documents or pictures and answers questions about them. It can extract, summarize, compare, or describe; verify small print, totals, and chart values.
Data analysis and code
Runs code or calculations over supplied data to create summaries, charts, and transformations. Check formulas, assumptions, units, and sample sizes before using the output.
Image and video creation
Creates or edits visual media from text and reference inputs. Give clear subject, composition, style, and constraints; inspect text, details, rights, and factual accuracy.
Voice and transcription
Turns audio into text, speaks a response, or supports live conversation. Review names, numbers, Khmer/English switches, speaker labels, and noisy sections.
Projects, memory, and connectors
Projects keep instructions and files together; memory can personalize future chats; connectors can retrieve information from connected services. Check what is retained and which sources or accounts are connected.
Function calling and agents
The model returns a structured request to a named tool. Your application executes that function and sends its result back. Tool calls can read data or change it, so restrict permissions and require approval for important actions.
CHOOSE AN AI
Which AI model or assistant should I choose?
Open an AI provider below to see its popular models in simple language. Each model has a button with a real example. Model menus differ by app, plan, and account.
Popular AI assistants and their models
Choose a provider to open its model list. Some services show model names; others choose a model for you or offer separate modes for search and research.
ChatGPT · OpenAI
These are the GPT choices shown in your screenshot and related model families. Default lets ChatGPT choose. Other choices suit different types of work.
Default
Let ChatGPT choose a suitable model for you. Good when you are not sure, or for everyday work.
Read a simple example
Prompt: Turn these meeting notes into a short action list. Show the task, owner, and due date. Mark missing details as “Not provided.”
GPT-6.1 Sol
A strong all-round choice for complex work when you want high capability without always choosing the most powerful option.
Read a simple example
Prompt: Compare three supplier proposals against our delivery, quality, and price requirements. Recommend one, cite the proposal evidence, and list what we still need to confirm.
GPT-6.1 Terra
Fast and efficient for understanding images or screens, using a computer, and producing clear writing.
Read a simple example
Prompt: Review this screenshot of our registration page. Point out the three most confusing steps and suggest clearer wording. Do not change the page.
GPT-6 Astra
The most capable choice in this family for especially demanding, multi-step work. Save it for a problem where careful reasoning matters.
Read a simple example
Prompt: Compare these monthly operations reports, check the totals, identify any conflicting figures, and draft a decision brief. Show calculations and mark assumptions.
GPT-6 Sol
Designed for complex coding and workflows where the model may use tools or complete several connected steps.
Read a simple example
Prompt: Help plan a stock-reordering workflow from this sample spreadsheet. Explain each step, flag missing information, and wait for approval before any change is made.
GPT-6 Luna
A quick, efficient choice for focused, repeatable, or high-volume tasks.
Read a simple example
Prompt: Turn these 12 support notes into a table with one short issue category per note. Do not add details that are not in the notes.
GPT-5.6 Sol
A previous model family option for demanding professional work such as research, coding, or long reports, when it is available in your picker.
Read a simple example
Prompt: Turn this procedure draft into a clear operations checklist. Keep every safety requirement and mark any unclear step as a question.
GPT-5.6 Terra
A balanced option in the GPT-5.6 family for useful results with a focus on efficiency.
Read a simple example
Prompt: Sort these customer comments into a few clear themes and give one short example from the comments for each theme.
GPT-5.6 Luna
An efficient GPT-5.6 option for straightforward tasks where a fast first draft is enough.
Read a simple example
Prompt: Extract the dates, tasks, and responsible people from these notes into a simple table. Use “Not stated” when a detail is missing.
OpenAI model catalog ↗ · ChatGPT model and thinking-level guide ↗
Gemini · Google
A family of models: choose a reasoning model for hard work, a Flash model for fast multi-step tasks, or an image model for visual creation.
Gemini 4 Argon
Google’s flagship reasoning model for difficult questions and long, detailed work.
Read a simple example
Example: Review these three project reports, find the biggest delivery risk, and support your recommendation with evidence from the files.
Gemini 3.8 Flash
A fast model for complex tasks, software work, and workflows that use tools.
Read a simple example
Example: Turn these meeting notes into an action plan with owners, deadlines, and open questions.
Gemini 3.5 Flash-Lite
A fast, lower-cost choice for many simple, repeatable requests.
Read a simple example
Example: Label these 100 feedback comments as positive, neutral, or negative. Return only a table.
Nano Banana 2.1
A specialist model for creating or editing images; it is not the same as a text-chat model.
Read a simple example
Example: Create a clean illustration for a staff training poster. Leave clear space at the top for a title; do not add words into the image.
Claude · Anthropic
Claude offers several text-and-image models. Pick by the balance you need between deep analysis, speed, and volume.
Claude Fable 5.1
For difficult reasoning and tasks that need an AI agent to work through many steps.
Read a simple example
Example: Read this operations plan, find missing approvals and risks, and prepare a review checklist. Separate facts from assumptions.
Claude Opus 5.5
For long-running coding and demanding knowledge work.
Read a simple example
Example: Compare our current and proposed standard operating procedures. List every important change and who needs to act.
Claude Sonnet 5.5
A balanced choice for capable, responsive everyday work.
Read a simple example
Example: Rewrite this customer update so it is clear and polite, while keeping all dates and commitments unchanged.
Claude Haiku 5.5
A fast choice for sorting, extracting, and other high-volume tasks.
Read a simple example
Example: Extract the names, dates, and action items from these short notes into a table. Use “Not stated” for missing details.
Grok · xAI
Grok 4.7 is the main model for chat and coding. Separate Imagine models handle image and video creation. Web or X search must be enabled for current information.
Grok 4.7
The main Grok model for chat, coding, and tool-using work.
Read a simple example
Example: Draft a response to this customer complaint. First list what the message says, then write a calm reply without promising a refund.
Grok Imagine Image 2.0
A specialist for making or editing images from a written description or reference image.
Read a simple example
Example: Create a square social graphic showing a team preparing for a workshop. Keep it professional and leave room for our logo.
Grok Imagine Video 1.5
A specialist for turning a prompt or image into a short video clip.
Read a simple example
Example: Make a 6-second video concept showing a training room changing from empty to ready for learners. Keep the camera steady.
DeepSeek
DeepSeek’s current API family has a capable Pro option and a faster Flash option. Flash accepts images; check the exact product before choosing.
DeepSeek V4-Pro
For demanding reasoning, coding, and work that needs strong tool use.
Read a simple example
Example: Review this process and find three likely causes of late deliveries. For each, show the evidence and a low-risk next step.
DeepSeek V4.1-Flash
A faster model for everyday work, structured output, and image questions.
Read a simple example
Example: Read this screenshot of a stock sheet and list the items that appear below the reorder point. Flag any number you cannot read clearly.
Meta AI · Muse
Meta’s assistant uses Muse models. The model names and features can differ between Meta AI, Muse, and developer tools.
Muse Wonder / Wonder Full
Meta’s frontier models for personal and professional tasks that involve planning and taking several steps.
Read a simple example
Example: Help plan a small staff training day from these notes. Make a timetable, list what is missing, and ask before assuming a venue or budget.
Muse Spark 1.2
A Meta model with a coding agent called Muse Code for working across larger software projects.
Read a simple example
Example: Review this small website project and suggest the three most useful fixes. Explain the effect of each before changing any files.
Microsoft Copilot
Copilot is a product that can offer Auto plus different model choices. The available choices depend on the Copilot app, subscription, and organization settings.
Auto
Let Copilot choose an available model for the task. A good starting point for everyday work.
Read a simple example
Example: Summarize this meeting transcript and list decisions, owners, and follow-up dates.
GPT or Claude options
Some Copilot experiences let you choose an OpenAI or Anthropic model. Choose based on the task and options your account shows.
Read a simple example
Example: Ask two available models to summarize the same project brief, then compare what each missed before sharing the summary.
Perplexity
Perplexity combines answer modes, web search, and model choices. “Best” can select a model automatically; deeper modes spend more effort researching.
Best mode
A quick search that chooses a suitable model for the question.
Read a simple example
Example: Find the official opening hours for this government office and show the source links.
Research mode
A multi-step search that gathers sources and prepares a fuller report.
Read a simple example
Example: Compare three official sources about a new business requirement. Summarize what agrees, what differs, and link each source.
What does the reasoning or effort level mean?
A setting such as High asks a model to spend more effort before answering. Use the standard setting for routine work and a higher one for complex plans, conflicting information, or careful analysis. It may take longer and still cannot guarantee a correct answer. Other AI apps may use different names or settings.
OpenAI · model and thinking guide ↗ · Microsoft · model and effort guide ↗
01 / MODEL WATCH
A practical evaluation plan for Cambodian teams
1. Pick one real task and create a small set of representative examples. 2. Compare the same prompt and reference material across two or three models. 3. Score factual accuracy, Khmer/English quality, editing time, privacy fit, and total cost. 4. Keep a person responsible for approval, and stop if the workflow creates more risk or work than it removes.
Model names and features change quickly. This briefing reflects official announcements checked on October 29, 2026; provider claims are identified as provider claims, not independent benchmark results.
CHATGPT · GEMINI · DEEPSEEK · CLAUDE
Four AI assistants: current models and what their functions do
Think of these as product families, not a fixed feature checklist. The same company may offer a chat app, API, and separate specialist models; access, limits, and data settings can differ.
MODEL FAMILY / 01
ChatGPT · OpenAI
The current API family includes GPT-6 Astra for demanding reasoning and coding; GPT-6.1 Sol for a strong capability/cost balance; GPT-6.1 Terra for visual reasoning and computer use; and GPT-6 Luna for efficient, focused work. Separate specialist families handle image generation, speech, and transcription.
Functions explained
- Chat: write, revise, explain, translate, and brainstorm from your instructions.
- Search: look up recent web information and return links to sources.
- Deep Research: investigate a broad question across sources and produce a cited report.
- Files and vision: summarize uploads, extract details, and interpret images or charts.
- Data analysis: run calculations and code on supplied spreadsheets or files; inspect the formulas and outputs.
- Image generation/editing: create or revise visual assets from text and reference images.
- Voice: speak and listen in real time; separate transcription models turn audio into text.
- Projects, memory, and connectors: group materials, retain permitted preferences, or retrieve connected app data; availability varies.
- Tools and computer use: search, retrieve files, run functions, or interact with a computer when the selected model and product permit it.
When it fits
Try when you need an all-purpose work assistant, a cited research brief, document analysis, or visual and voice workflows. Ask for sources and calculations; verify important claims. ChatGPT app features and API model tools are different surfaces.
MODEL FAMILY / 02
Gemini · Google
Gemini 4 Argon is listed as Google’s flagship reasoning model. The Gemini 3 family includes 3.8 Flash for complex and agent-style work, lighter Flash-Lite options, and specialist image, speech, and live models. Check Google’s live catalog because preview and availability status can change.
Functions explained
- Chat and multimodal input: work with text, supported images, audio, video, and documents in supported models.
- Google Search grounding: retrieve current web information and show grounding sources.
- Deep Research: plan a broad search and synthesize an organized report.
- Image generation/editing: create and revise images with the Nano Banana model family.
- Audio and Live: transcribe, synthesize speech, or support real-time voice interaction with specialist models.
- Built-in tools: supported API models can use services such as Google Search, Maps, URL Context, File Search, and Code Execution.
- Function calling: return structured arguments for your application to execute; the model does not itself run your custom function.
- Large-context document work: process long inputs when supported; always check model and product limits.
When it fits
Try when your task combines several media types or needs Google Search/Maps and related tools. Confirm model availability in the Gemini app, AI Studio, or API separately; they do not expose identical features.
MODEL FAMILY / 03
DeepSeek
The API currently lists deepseek-flash (V4.1-Flash-1017) and deepseek-v4-pro (V4-Pro-0813). Flash supports vision; V4-Pro does not in the current table. Both list thinking modes, JSON output, tool calls, and Responses API compatibility. “Using a hosted API” and “self-hosting open weights” are different choices.
Functions explained
- Chat: ask questions or provide text in the DeepSeek chat product.
- Thinking mode: spend more effort on multi-step tasks; the API may return a separate reasoning-content field.
- Vision: analyze supported image inputs with the Flash model; do not assume every model accepts images.
- JSON output: constrain the response to a machine-readable JSON shape.
- Tool calls: request a declared function with structured arguments; your application performs the call and returns its result.
- Search and Fetch: use DeepSeek’s search/fetch API for current pages where configured; verify sources.
- API compatibility: use supported Responses or Anthropic-style interfaces with the published model IDs.
- Self-hosting: only use weights and terms that the current release license permits; budget for infrastructure and operations.
When it fits
Try for structured coding, API workflows, or a comparison of reasoning effort and cost. Check the exact model, image support, data handling, license, and Khmer quality before use.
MODEL FAMILY / 04
Claude · Anthropic
The current lineup lists Fable 5.1 for demanding reasoning and long-horizon agents; Opus 5.5 for long-running coding and knowledge work; Sonnet 5.5 for a speed/intelligence balance; and Haiku 5.5 for low-latency, high-volume tasks. Feature support and model IDs can vary by platform.
Functions explained
- Chat and vision: write, analyze, and interpret text and images; current models support multilingual output and tool use.
- Extended/adaptive thinking: spend more inference effort on complex work; choose effort based on your own evaluation.
- Projects and RAG: keep instructions and files together and retrieve relevant passages from project knowledge.
- Web Search and Research: find current pages and synthesize a multi-step report when enabled.
- Artifacts: create an interactive document, visualization, or small app that can be inspected and revised.
- Claude Code: work with software repositories through a coding agent and its approved tools.
- Tool use and computer use: request APIs or UI actions through schemas; the host runs client tools and returns results.
- Haiku / Sonnet / Opus / Fable: trade response speed and cost against capability; compare on your actual tasks.
When it fits
Try for long documents, complex coding, careful research, or multi-step knowledge work. Start with an appropriate speed/capability tier and compare output quality, latency, and cost on a representative task.
PRACTICE / SIDE-BY-SIDE
A fair, hands-on model comparison
This site does not send your text to AI providers or use their API keys. Copy one task, run it yourself in each service, then paste the answers here. Your notes stay in this browser.
Use the same rubric for every model: factual accuracy, completeness, language quality, editing time, privacy fit, and cost. Score from 1 (weak) to 5 (strong).
Responses are saved only in this browser’s local storage. Do not paste private or confidential information.
Official source
- OpenAI · models ↗
- OpenAI · ChatGPT capabilities ↗
- OpenAI · tool use and function calling ↗
- Google · Gemini API model catalog ↗
- Google · Gemini API tools ↗
- Google · function calling ↗
- DeepSeek · models and features ↗
- DeepSeek · thinking mode ↗
- DeepSeek · tool calls ↗
- Anthropic · current models ↗
- Anthropic · tool use ↗
Checked October 29, 2026. Availability and features change by product, plan, country, and account.
