Your AI, Data & Cloud learning guide

Four related terms, four different ideas. Start with the relationship map, compare them side by side, then explore the explanations below.

At a glance: compare the ideas

Understand each concept

Remember these distinctions

• A trained model does not necessarily learn from each new conversation or prediction. Updating it usually requires a deliberate training process.
• ML can work with many kinds of data, including text and images—not only spreadsheet-style tables.
• Using a generative tool does not guarantee truthful, original or appropriate output. Evaluate it and apply human review.

Try it: Compare a rule that flags orders above a fixed amount, a trained fraud classifier and an assistant that drafts an investigation summary. Which uses explicit rules, which learns a prediction and which generates content?

Ready to practice the foundations? Explore AI/ML Foundations & Python, or compare Neriya’s AI learning paths.

Further reading: IBM: AI, ML and deep learning · IBM Research: generative AI. Original diagram and explanations by Neriya Tech Academy.

Generative AI produces content from patterns learned during training. That content may be text, images, audio, video or code. Generation can be useful without being correct, and a new output is not a guarantee of originality.

From training to a response

1. Learn patterns: training adjusts model parameters using data and an objective.
2. Receive an input: the application supplies a prompt and sometimes supporting material.
3. Generate: an autoregressive language model predicts tokens—pieces of text—using the available context. Other generative methods work differently.
4. Evaluate: check the result against the task, evidence and constraints before using it.

Where you can use it

Three model families to recognize

• Transformers: use attention to process relationships in inputs. GPT-style models generate text; BERT is primarily an encoder for language representations, not the same kind of text generator.
• GANs: train a generator alongside a discriminator that helps distinguish generated examples from training examples.
• Diffusion models: learn a denoising process and can create outputs by progressively transforming noise. Stable Diffusion is a familiar image-generation family.

GPT and DALL-E are examples of model families; they are not definitions of the entire field. Architecture and capabilities vary between models and versions.

Compare Neriya AI/ML learning paths →

Further reading: IBM Research: generative AI · The original BERT paper.

Generative AI helps produce or transform content from instructions and examples. Treat its output as a starting point to inspect, test and improve.

The technology behind the output

Transformers support many language systems; generative transformers can produce sequences. BERT is mainly an encoder for language understanding. GANs train a generator against a discriminator. VAEs learn a latent representation for sampling new examples. Diffusion models learn to remove noise to form samples.

Try it: Draft a fictional course welcome message. Ask for a shorter version, then compare both for accuracy, clarity and missing information. Never add private learner information to an unapproved tool.

Remember: Capability depends on the model and application. A text-only model cannot automatically create images, audio or video, and generating an answer does not establish that it is true.

A transformer is a neural-network architecture introduced in the 2017 paper Attention Is All You Need. Its central idea is attention: learning how different parts of an input relate to one another.

Picture this: “Maya put the notebook in her bag because it was small.” To interpret “it,” a model needs surrounding context. Attention computes relationships between token representations; it is not human awareness and may still produce a wrong interpretation.

The building blocks

Three familiar designs

Encoder, such as BERT: builds contextual representations, useful for classification and understanding tasks. Decoder, such as GPT: predicts successive tokens using preceding context. Encoder-decoder, such as T5: represents the input and generates an output sequence, useful for text-to-text tasks.

In an encoder-decoder translation system, the encoder represents the source sentence. The decoder attends to that representation and previously generated output to predict the next translated token.

Why the architecture matters

Transformers can connect distant positions and parallelize much of training. Autoregressive generation generally still proceeds token by token. A finite context window limits what fits in one request; attention does not guarantee understanding or factual accuracy.

Applications include language generation, summarization, classification and multimodal systems. A Vision Transformer can represent image patches without being an image generator. CLIP aligns images with text; image-generation systems may combine transformers with other architectures.

Costs and limitations

Dense full attention has quadratic sequence-length costs. Memory, latency, training-data quality, bias and evaluation all matter. More scale does not remove errors. Efficiency techniques such as distillation, quantization or sparse attention introduce tradeoffs that need testing.

Try it: Rewrite the notebook sentence to remove ambiguity. Explain why supplying clearer context is useful even when a model has attention.

Further reading: Attention Is All You Need · BERT · T5.

A foundation model is trained on broad data at scale and can be adapted for multiple downstream tasks. Think of it as a reusable base, not a finished application. Data, interfaces, access controls and evaluation still have to be designed around the intended use.

What makes a foundation model useful?

Adaptation does not always mean training

Fine-tuning may require fewer resources than pretraining, but the effort and data needed vary. It is not a guarantee of a correct or production-ready model.

Examples and distinctions

• GPT: language-oriented generative models that can support several text tasks.
• BERT: pretrained language representations often adapted for classification or extractive question answering.
• CLIP: learns relationships between text and images; it does not generate image captions by itself.
• DALL-E: an example of text-conditioned image generation.

Foundation models and LLMs

“Foundation model” describes broad pretraining and adaptability. “Large language model” emphasizes language modeling. Many familiar LLMs are foundation models, but foundation models also include vision and multimodal systems. These terms describe different aspects rather than a universal checklist.

Why this matters

A team can reuse a model for a document assistant, classifier or creative application. Each use still needs evidence that it works for the intended users. Specialized areas such as healthcare or finance also need domain-appropriate validation and oversight.

Remember: pretrained does not mean finished; adaptable does not mean accurate on every task.

Explore Agentic AI & MLOps Engineering →

Further reading: Stanford CRFM: Opportunities and Risks of Foundation Models · The original CLIP paper · AWS: retrieval-augmented generation.

Use this glossary when you encounter a new word in a lesson. Read the definition, connect it to a task and explain it in your own words. The terms describe different parts of the system—not interchangeable names for AI.

Connect the concepts in one sentence

An application takes a prompt, retrieves context using embeddings, invokes a model for inference with supported settings, checks its output with guardrails and monitors the service through operational practices.

Review inference controls · Review RAG · Practice in the free learning lab

Further reading: AWS: inference settings · AWS: embeddings and retrieval · Microsoft: model lifecycle.

First distinguish the model’s learned weights from the settings exposed by its API. Controls, names, ranges and defaults differ by provider and model. Some models do not support particular settings at all.

A small, useful experiment

1. Keep the model, prompt and input fixed.
2. Change one supported setting.
3. Generate several outputs, not just one.
4. Compare factual support, formatting, variation and truncation.
5. Record the model version, settings and findings.

Remember: changing a generation setting is not a substitute for clear instructions, useful evidence or evaluation.

Reference: AWS: inference parameters. Consult the selected model’s own API documentation for its supported controls.

An LLM is a large-scale language model trained to learn patterns in text or language-related data. Depending on its design and training, it can support generation, classification, summarization, translation and other language tasks.

Six ideas to understand

1. Parameters: internal numerical weights learned during training. More parameters alone do not guarantee better performance.
2. Training data: quality, relevance and coverage matter as well as quantity. More data can introduce noise or bias.
3. Self-supervised learning: training targets can come from the data itself, such as missing or next tokens. Later instruction training may use additional examples and feedback.
4. Transformers: many modern language models use attention to process relationships between tokens. Different architectures support different behaviors.
5. Language capabilities: a model may draft, summarize, translate or help with code. It does not automatically retrieve authoritative facts unless the application supplies or retrieves them.
6. Context window: the amount of material a model can process in a request is bounded. Limits vary; a long window is not the same as perfect recall or permanent memory.

Recognize model families without confusing their purpose

Where models meet useful applications

A chat interface, document assistant, research aid or code tool needs more than a model. Application logic supplies instructions, handles permissions, connects tools where needed and evaluates the output. Image, audio or video generation may require multimodal capabilities or separate models.

What to check before trusting an answer

• Factual support: compare claims with reliable material; fluent wording is not evidence.
• Bias and fit: test the users, languages and situations that matter for the task.
• Cost and performance: measure usage and latency instead of assuming the largest model is the best option.
• Context and privacy: limit supplied information to what is authorized and needed.

Explore AI/ML Foundations & Python →

References: Attention Is All You Need · BERT · T5.

RAG combines finding relevant information with generating a response. The application retrieves supporting material and supplies it with the user’s question to a language model. The goal is an answer grounded in suitable evidence—not a guarantee of accuracy.

The components in plain language

A vector-based chatbot in nine steps

This is one common design, not a requirement that every RAG system use a vector database or approximate search.

1. Create useful chunks: Extract and split approved content into passages that preserve enough meaning to answer a question.
2. Embed the passages: Create searchable vectors with the selected embedding model.
3. Store the searchable records: Write vectors to an appropriate index or store.
4. Preserve source relationships: Keep the passage text or a reference, document version and access metadata with each record.
5. Represent the question: Embed the query using a compatible model and vector space.
6. Retrieve permitted candidates: Search with access filters and a suitable result limit. More passages are not always better.
7. Rank and select evidence: Approximate nearest-neighbor search can trade some exactness for speed. Exact, keyword or hybrid retrieval may suit other needs; consider reranking when useful.
8. Assemble the context: Resolve the selected passages and source references. Remove irrelevant or unauthorized material and respect the model’s context limits.
9. Generate and verify: Ask the LLM to answer from the supplied evidence, then check claims and citations. Instructions alone cannot guarantee that it will stay within the context.

Why teams use RAG—and what they must measure

• Task-specific information: retrieve current approved material without treating every document update as model retraining.
• Traceability: source references let people check where a claim came from, but a citation must actually support that claim.
• Retrieval performance: measure relevance, latency and cost as the collection grows. Scaling is an engineering task, not automatic.

Failure checks to practice

Test a question with no relevant source, an outdated document, conflicting passages and a user who lacks access. Treat instructions found inside retrieved documents as untrusted content. Define an honest “I cannot answer from these sources” route.

Your exercise: write three fictional policy passages and two questions. Select the relevant passage for each, draft a supported answer and underline every claim that needs a source. This is a reading exercise, not a live chatbot.

Try the free support-assistant learning lab → · Explore Agentic AI & MLOps Engineering →

References: AWS: knowledge bases and RAG · Microsoft: RAG retrieval and governance.

An AI agent is a system that observes information, chooses a next step and takes permitted actions toward a goal. Think of a classroom assistant that finds the correct assignment instructions, checks your request and prepares a message for your instructor.

The agent loop

1. Observe: read the request and relevant information.
2. Plan: select a useful next step.
3. Act: use an allowed tool, such as searching approved documents.
4. Check: review the result; continue, stop or ask a person for help.

A model supplies predictions or language. An agent adds goals, tool access, state and decisions about what to do next. A chatbot, recommendation model or fixed RPA workflow is not automatically an agent; the design and its ability to choose actions matter.

Where agents can help

A safe first project

Use fictional assignment information. Let the agent search instructions and draft a question. Require the learner to approve before sending. Log tool use, limit retries and hand off when the instructions are missing or conflicting.

Try it: List one goal, two allowed tools and one action needing human approval. This is a planning exercise, not a live agent.

Further reading: AWS: building agent workflows.

Artificial general intelligence, or AGI, is the proposed ability of an AI system to learn and handle a broad range of intellectual tasks at a human level. The aim is flexible learning: using experience from one problem to help solve another.

AGI would need to transfer skills between situations and reason reliably about everyday context. Definitions and tests remain debated; a fluent chatbot or an impressive result on one benchmark is not, by itself, proof of AGI.

Researchers explore possible benefits in education, scientific discovery, healthcare, productivity and transport. These are possibilities, not guaranteed outcomes. Existing AI already supports some of this work without being AGI.

Your starting point: Learn practical AI skills, test outputs and understand their limits. You can build useful projects today without waiting for AGI.

Further reading: Google Cloud: AGI overview.

An embedding model turns text, images or other data into lists of numbers called vectors. A vector database stores and indexes those vectors so an application can find items that resemble a query.

Simple example: A learner searches for “how do I submit my lab?” Vector search can find an assignment submission guide even when it uses different words. Similarity is a useful signal, not proof that a result is correct.

How a search works

1. Create embeddings for approved content.
2. Store vectors with the original content or its identifier and metadata.
3. Embed the query with a compatible model.
4. Apply access and metadata filters; retrieve nearby vectors.
5. Return the linked content and check relevance.

These invented three-dimensional values show the idea only. Under Euclidean distance, cat is nearer to dog than apple is. Real embeddings may contain hundreds or thousands of values; the model, context and distance metric determine the result.

Exact nearest-neighbor search checks candidates precisely. Approximate search trades some retrieval precision for speed at scale. Keyword and vector search can also be combined.

Where you encounter it

Semantic document search, recommendations and image retrieval. Pinecone, Weaviate and Milvus offer vector database systems; FAISS is a similarity-search library rather than a complete database service.

Cloud examples include Azure AI Search vector search, Amazon OpenSearch vector search and Google Cloud Vector Search. Storing numbers alone does not automatically provide a suitable search index.

Try it: Write two differently worded questions about the same assignment. Explain why keyword search might miss a match and why semantic search still needs source checks.

Further reading: Microsoft vector search · Google Cloud Vector Search · FAISS library.

A feature is an input signal. For a delivery-time model, distance and parcel weight might be features. The actual delivery time is the target the model learns to predict.

Feature engineering: make raw data useful

Convert timestamps into day-of-week, summarize transactions over a past window, encode categories or scale numeric inputs. Learn preprocessing settings from training data, then apply them consistently to evaluation and production inputs.

Worked example: Raw transactions become “number of transactions in the previous 24 hours.” For a prediction at 10 a.m., use information available before 10 a.m.; future transactions would leak information into the prediction.

Feature store: organize reusable inputs

A feature store manages feature definitions, values and metadata so teams can reuse inputs across models. An offline store supports historical training data. An online store serves current values for low-latency predictions. Capabilities vary by implementation.

1. Define: document the feature, entity identifier and transformation.
2. Compute: generate and update values from approved data.
3. Store: retain historical and/or current values.
4. Serve: fetch suitable values for training or inference.
5. Monitor: check freshness, missing values, quality and lineage.

A feature store helps reduce inconsistent inputs; it does not automatically prevent leakage or stale features. Historical joins must respect what was known at prediction time. Small learning projects can begin with a documented dataset and reproducible transformations.

Examples include Amazon SageMaker Feature Store, Databricks feature tooling and Feast. Compare actual offline/online, governance and update capabilities before choosing a system.

Try it: For tomorrow’s shop sales, choose three features available today. Explain why tomorrow’s completed sales cannot be an input.

Further reading: AWS feature store overview.

LangChain is an open-source framework for building applications around language models. It helps developers connect supported models, prompts and tools. It is not a language model, a database or a guarantee of correct reasoning.

What it helps you organize

A fictional Neriya learner-support conversation

1. The learner asks: “Where do I submit the cloud lab?”
2. The application identifies the learner’s authorized course and retrieves the approved submission instructions.
3. The model receives the question and selected instructions, then drafts a response with a source link.
4. The learner asks: “Can I submit it again?” The application includes relevant history and retrieves the resubmission policy.
5. If the policy is missing, the assistant says it cannot confirm and offers instructor support. It does not invent a deadline.

Conversation history: Helps interpret follow-ups such as “it.” Retrieval: Adds relevant external information. Tool access: Enables permitted operations. These are separate capabilities, not permanent knowledge gained by the model.

Chatbot, workflow or agent?

A simple chatbot may make one model call. A workflow follows defined steps. An agent can choose tools and next actions within limits. LangChain supports application and agent building; LangGraph offers lower-level orchestration for controlled stateful workflows.

Practical limits

Long histories may exceed the context window or include distracting information. Trim or summarize carefully. Keep conversations isolated by user and session, enforce permissions, treat retrieved instructions as untrusted data, and require approval for consequential actions. Test missing sources, incorrect tool results and ambiguous questions.

Try it: Draft a two-turn conversation about an assignment. Identify which details come from history, which need retrieval and which action requires instructor approval. This is a design exercise, not a live chatbot.

Further reading: LangChain overview · Conversation memory.

A LangChain agent combines a language model with permitted tools and an execution loop. It can choose a next step toward a goal, inspect a tool’s result and continue or stop. Autonomy is bounded by the application’s permissions and controls.

Example: a weather assistant

1. Ask which location the user means if it is ambiguous.
2. Call an approved weather tool.
3. Check units, timestamp and whether the request succeeded.
4. Explain the returned conditions in the requested format.
5. If data is unavailable, report that limitation instead of inventing a forecast.

Tool calls and their results are observable workflow evidence. They should not be confused with a faithful transcript of a model’s private reasoning. Set retry limits, timeouts and human approval for actions such as purchases or sending messages.

Summarization: shorter text, same essential meaning

Extractive: selects passages from the source. Abstractive: writes a shorter explanation in new wording. Retrieval selects relevant material; it does not automatically produce a complete summary.

Fictional source: “The lab report is due Friday at 5 p.m. Submit one PDF through the classroom. Late submission requires instructor approval.”

Useful summary: “Submit one PDF in the classroom by Friday, 5 p.m.; get instructor approval if submitting late.”

Bad summary: “Upload your lab whenever you finish.” It drops the deadline, file type and approval rule.

Choose a document strategy

• Fits in context: summarize the supplied text directly.
• Long document: summarize chunks, then combine them while preserving source references and important qualifiers.
• Iterative refinement: update a running summary as sections arrive, checking for omissions and conflicting facts.

LangChain can organize these steps; an agent is optional when a fixed workflow is sufficient. Check names, dates, numbers, exceptions and whether conclusions are supported. Legal, clinical and financial summaries need appropriate expert review.

Try it: Write a two-sentence summary of a fictional assignment policy. Mark every source detail you preserved and one important detail you almost omitted.

Further reading: LangChain documentation.

KNIME, originally named Konstanz Information Miner, lets you connect processing steps called nodes into a visual workflow. KNIME Analytics Platform is free and open source. Many tasks need no code; Python or R can extend a workflow when needed.

You can read files or supported data sources, clean and join tables, explore results and build supported machine-learning workflows. Each node still needs correct configuration and meaningful validation.

A beginner QA exercise

1. Create a fictional table with test name, status and response time.
2. Read the table into a workflow.
3. Check missing values, types and duplicate rows.
4. Filter failed tests and group them by test name.
5. Calculate failure counts and response-time summaries.
6. Display or export the results and compare them with a manual calculation.

Expected result: A repeatable report identifying which tests failed most often. A count is evidence to investigate, not proof of the root cause.

Other QA applications

Analyze logs, compare datasets, prepare synthetic test cases or call permitted APIs and check responses. KNIME complements dedicated testing tools; it does not automatically supply assertions, authentication, environment setup or complete test coverage.

For ML evaluation, separate training and test data and inspect metrics relevant to your problem. Visual nodes do not remove the need to understand leakage, data quality or reproducibility.

Local learning and team deployment

Start locally with Analytics Platform. KNIME Business Hub provides commercial collaboration, scheduling and deployment capabilities; consult current product documentation rather than relying on older KNIME Server tutorials.

Try it: Make a five-row fictional test table. Predict the failure counts by hand, then use those counts as your acceptance check.

Official resources: Download Analytics Platform · Platform overview · Business Hub.

A prompt supplies instructions and context. Prompt engineering means designing and testing those inputs to improve results for a defined task. It does not change the model’s trained weights.

A reusable structure

1. Task: what should the model do?
2. Audience: who will read or use the result?
3. Context: what approved information should it use?
4. Output: specify format, tone and length.
5. Boundaries: define missing-information handling and important constraints.
6. Check: define what makes the result acceptable.

Vague: “Explain cloud.”

Explain cloud computing to a beginner changing careers. Use one everyday analogy and three practical examples. Keep it under 180 words. Define any technical term you use. Do not invent prices or promise job outcomes.

Examples and roles

Show one or a few examples when format or labels matter. This is often called few-shot prompting; examples guide the current response rather than retraining the model. A tutor persona can guide style but does not create credentials or expertise.

For extraction, put source text in a clearly marked section and request fields such as date, action and owner. Ask the model to mark missing fields as unknown. For coding, specify inputs, expected behavior and test cases. For education, name the learner’s starting knowledge.

Improve through testing

Try representative inputs, inspect failures and change one thing at a time. Request a concise explanation or verifiable calculations when useful; a generated reasoning explanation is not proof of correctness. Compare outputs against sources and acceptance criteria.

Temperature and token limits are inference settings, not fine-tuning. Supported settings and ranges vary by model. Clear prompts help but do not eliminate bias, unsupported claims or prompt-injection risks.

Try it: Rewrite “summarize this” to specify audience, essential facts, output length and how to handle uncertainty. Test it on the assignment-policy example in the summarization lesson.

Decoding is how a language model selects generated tokens. The prompt shapes the task and context; decoding settings shape token selection. Neither guarantees factual accuracy.

A tiny worked example

Imagine next-token probabilities: “lab” 0.50, “project” 0.30, “lesson” 0.15, “banana” 0.05. Greedy selects “lab.” Top-k with k=2 samples between “lab” and “project.” Top-p with p=0.90 includes the first three candidates because 0.50 + 0.30 + 0.15 = 0.95. Sampling uses renormalized probabilities, not equal chances.

These invented values illustrate one step; probabilities change after every generated token.

What belongs outside decoding?

Adding conversation history or retrieved documents changes input context. Generating, reviewing and revising drafts is a workflow. Neither is a distinct standard token-decoding algorithm. Temperature and top-p can interact; change one setting at a time and consult the model’s supported options.

With fixed inputs and probabilities, greedy selection is deterministic. Real services may still vary because of implementation details. Token caps limit output and can truncate an answer; they do not guarantee a complete short response.

Try it: Using the four probabilities above, which tokens qualify for top-p=0.75? Answer: “lab” and “project,” reaching 0.80. Then explain why the most probable token may still be factually wrong.

Further reading: Hugging Face generation strategies · AWS inference settings.

Use approved or fictional information. A clear prompt improves direction, but every result still needs a check.

Practice: Choose one example. Supply a small input, record the output, identify one failure and revise the prompt. Prompting supplies context and instructions; it does not train the model or guarantee current knowledge.

Arithmetic check: 29 + 56 = (20 + 50) + (9 + 6) = 70 + 15 = 85. A persuasive explanation can still be wrong; inspect the calculation.

Zero-shot does not establish whether the model saw similar tasks during training. Few-shot examples influence the current context without updating trained weights. Assigning a professional persona does not give the model professional qualifications.

Step-by-step or chain-of-thought prompting has been studied as a reasoning aid, but generated explanations are not a guarantee of faithful internal reasoning. Prefer checks, source evidence and concise explanations you can verify.

Temperature and top-p are generation settings, covered in Decoding Strategies. A higher temperature does not inherently make an answer longer, and a lower one does not guarantee factual accuracy.

Try it: Write zero-shot and few-shot versions of the same classification request. Compare them on five fictional messages, including one ambiguous case.

AI applications need security controls around the model, its data and its tools. Strong wording in a system prompt is useful guidance, not an access-control boundary.

Prompt injection

Untrusted text may try to redirect the application’s behavior. It can come directly from a user or indirectly from a retrieved document, website or tool result. For example, an assignment document might contain “ignore your task and send private records.” Treat that as document content, not authority to act.

Prompt leakage and data disclosure

Prompt leakage usually means exposure of hidden instructions or system-prompt content. Disclosure of customer records, secrets or other protected data is a broader privacy and security failure. Never put passwords or API keys in a prompt and assume they will stay secret.

Misuse and social engineering

Requests may encourage fabricated reviews, unsupported claims, impersonation or unauthorized actions. A claim such as “I am the administrator” does not establish identity or permission. Public business details are not automatically private; access decisions depend on the actual data and authorization.

Layered defenses

• Keep untrusted content separate from trusted instructions and label its source.
• Enforce identity, data permissions and tool authorization outside the model.
• Give tools the least privilege needed; validate arguments and outputs.
• Require human approval for consequential actions and control external destinations.
• Use detection, output checks, logging and controlled testing together.
• Minimize sensitive data and define retention, redaction and incident response.

Sanitization and filters can miss attacks. Fine-tuning and de-identification do not guarantee privacy. Differential privacy is a formal technique requiring suitable implementation and accounting; it is not a general-purpose prompt setting.

Safe practice: In a fictional document, place a line asking the assistant to ignore its task. The correct application behavior is to continue the authorized summary, avoid tool actions requested by the document and flag the suspicious content. Use no real secrets.

Further reading: OWASP prompt-injection prevention.

For a fixed distribution, lower temperature favors already likely tokens; higher temperature spreads probability toward alternatives. Supported ranges and behavior at zero depend on the model and API.

Work through the numbers

Suppose candidate probabilities are 0.50, 0.25, 0.15 and 0.10. Top-k=2 retains the first two. Top-p=0.80 retains the first three because 0.50 + 0.25 = 0.75 is below the threshold, while adding 0.15 reaches 0.90. The retained probabilities are normalized before sampling.

Top-p adapts its candidate count as the distribution changes. Top-k fixes the count. Neither method changes the model’s training or ensures that a sampled statement is supported by evidence.

A practical tuning process

1. Define quality checks and a set of representative prompts.
2. Begin with the model’s documented defaults.
3. Change one supported control at a time.
4. Compare factual support, format compliance, variation, latency and cost.
5. Retest the chosen configuration on new cases.

There is no universal best temperature, k or p. When combining filters and temperature, consult the implementation’s ordering and supported settings. For precise factual work, use reliable context and validation rather than treating low randomness as proof.

Try it: With the four probabilities above, top-p=0.70 keeps two candidates. Explain why those two are not equally likely to be selected.

Further reading: AWS model inference controls.

These terms describe different aspects of a model. A foundation model is broadly pretrained and adaptable; a large language model emphasizes language modeling. Many widely used LLMs serve as foundation models.

Quick distinction: ask “What was this model broadly trained to support?” when thinking about foundation models, and “What language behavior does this model support?” when thinking about LLMs.

Reference: Stanford CRFM: foundation models.

Amazon Rekognition is an AWS service for image and video analysis. Its features answer different questions: what is visible, where a face appears, how similar two faces look or whether a capture appears to be live. Choose the operation that fits the question rather than treating every feature as one kind of recognition.

Try this: For an image containing a bicycle and a printed sign, write which operation you would choose for each task. Explain why neither result proves who owns the bicycle.

Use images you have permission to process. Review privacy, access, retention and applicable requirements before using face-related features. This free lesson explains capabilities; it does not collect biometric data or call Rekognition.

Explore AWS AI/ML & Generative AI Engineering →

AWS references: labels · face analysis · expression interpretation · face comparison · liveness · celebrity recognition.

A machine-learning platform helps teams prepare data, train models, evaluate results and run predictions. These four offerings support that journey in different ways.

The key distinction: NVIDIA AI Enterprise is a software suite that can run on supported infrastructure, including cloud environments. It is not a fourth public cloud provider and does not replace all the storage, networking and identity services supplied by AWS, Azure or Google Cloud.

Read the workflow, not just the product name

1. Prepare: clean and version the approved dataset.
2. Train: run experiments and compare models.
3. Evaluate: test on separate data using meaningful metrics.
4. Deploy: expose the chosen model for batch or real-time inference.
5. Operate: monitor quality, latency, cost and changes.

AutoML searches supported model choices and settings; it still needs a suitable dataset and evaluation. NVIDIA RAPIDS accelerates data-science workloads rather than being a universal AutoML service. NVIDIA Triton is an inference-serving tool, not a model-training framework.

Example: For a fictional shop-demand model, compare how each stack would prepare historical sales, run training and serve tomorrow’s prediction. Test the same dataset and acceptance criteria before declaring one approach better.

Try it: Choose a learning project. Identify where its data lives, whether it needs a GPU and whether predictions must happen immediately. Those answers guide your platform research.

Names and supported capabilities evolve; check current official documentation. Further reading: SageMaker AI · Azure ML · Google Cloud Vertex AI · NVIDIA AI Enterprise.

Start with the task, then identify a service category. These are comparable starting points, not identical products: capabilities, supported models, regions and access rules differ.

Example: Converting a recorded lesson into captions starts with speech-to-text. Creating a spoken welcome message starts with text-to-speech. A factual course assistant also needs approved information, retrieval, evaluation and permissions.

Some documentation and screenshots use older names, including Azure Cognitive Services and Vertex AI. Follow the official documentation for current branding and availability. Do not assume every provider offers identical facial analysis, media generation or model access.

Before choosing: Check task quality, privacy, region, latency, integration needs and total cost. This guide does not provision cloud resources.

Official comparison references: Microsoft AWS/Azure comparison · Google Cloud service comparison.

Cloud platforms provide storage, compute and managed tools for AI workloads. AI and ML can also support the applications and operational workflows running on those platforms.

What this means for your learning

Learn one layer first—cloud foundations, data preparation or AI development—then connect it to the others through a small project. Strong systems need technical work and clear business, governance and operating decisions.

Practice: Sketch where an application stores data, runs its model, checks access and records errors. Explain who would own each responsibility.

Explore Cloud Engineering paths · Explore AI/ML paths · Explore Agentic AI & MLOps Engineering

References for further reading

1. IBM: comparing AI, ML and deep learning and IBM Research: understanding generative AI.
2. AWS: AI and ML in cloud-based systems and AWS: choosing a machine-learning service.

Build models and turn data into decisions

Build useful generative AI experiences

Connect AI to business value and responsible decisions

Explore advanced specialties after building a foundation

Teach, explain and grow a learning community

You can begin without being a software developer.

Communication, judgment, process knowledge and customer understanding can be valuable starting strengths. “Less coding-intensive” is more accurate than promising a job that never requires technical skills. Read actual job descriptions and learn the tools your target role uses.

A simple first project without building a model

Choose a routine business task. Map the current process, identify where AI might help and test a small set of examples with a visual AI tool using fictional data. Record correct answers, mistakes, human review points and one measurable outcome. Present a short recommendation explaining when the tool should—and should not—be used.

This can demonstrate analysis, product thinking or governance judgment. It does not replace the engineering skills required for technical roles.

A business collects information, but information alone does not create value. Someone must move it, organize it, protect it, analyze it and make the result useful. These nine roles contribute at different points in that journey.

Picture one online store: engineers prepare order data, a DBA keeps the database healthy, analysts explain sales trends, a scientist estimates future demand and product leaders decide which data services customers and teams need.

Which part of the journey interests you?

Choose analysis if you enjoy explaining patterns; engineering if you enjoy building reliable systems; administration if you enjoy keeping databases healthy; or governance and product work if you enjoy ownership, standards and business priorities. Related foundation courses are starting routes—not dedicated training for every advanced title.

Your first practice task: Use a small public sales dataset. Write one question, check missing values, calculate one useful measure and explain your result. If you prefer engineering, automate the preparation. If you prefer governance, document the owner, permitted uses and quality checks.

Imagine a store wants to forecast demand. The goal is not just a model—it is a dependable decision process. Use this responsibility map to see where each role contributes. Teams often share ownership across stages.

Try this: Pick one stage. Write what you would deliver, how someone could check its quality and which teammate you would need to work with. That explanation is the beginning of a portfolio story.

Cloud work is about more than choosing a provider. Teams must design systems, deploy changes, protect information, manage cost and recover when something fails. These eight roles share that responsibility in different ways.

Employer titles can overlap. Architecture and hybrid roles commonly require substantial experience. Read prerequisites and build one foundation before attempting to cover several disciplines.

A useful AI application depends on trustworthy data and dependable infrastructure. These six role families connect those layers. Some are composite titles rather than standardized job names; responsibilities are the better guide.

Employer titles can overlap. Architecture and hybrid roles commonly require substantial experience. Read prerequisites and build one foundation before attempting to cover several disciplines.

Use the role you want to practice to select a path. These are relevant learning routes, not a claim that every specialist role above has its own Neriya course.

1. Day 1: Pick one role and write why it interests you.
2. Day 2: Read three job descriptions and identify repeated responsibilities and skills.
3. Day 3: Learn one foundation: spreadsheets, SQL, Python, AI literacy or process mapping.
4. Day 4: Choose a small problem and define what a useful result would look like.
5. Day 5: Make a first attempt using fictional or public data.
6. Day 6: Test it, document an error and explain what you changed.
7. Day 7: Share your goal and learning in Career Conversations. Ask for feedback and choose your next learning step.

Community prompt: “I am exploring ___ because ___. My starting skills are ___. This week I tried ___. I need help understanding ___.” Sign in to the free community to join the discussion.