Eight external learning resources with SohoBook commentary. Choose a learning path or an individual course; preparation advice and exercises connect the material with the textbook chapters.

Provider curriculum summaries are separate from SohoBook’s own study guidance and exercises. Descriptions are available in three languages; check the provider for the course language, registration, lab access and certificate conditions. Original links are preserved; courses.nvidia.com addresses redirect to the learning platform.

Source information checked on 15 October 2026.

Suggested learning sequences

AI foundations: Generative AI explained A neural network: a “brain” in ten minutes
RAG and language models: Augmenting a language model with RAG Building RAG agents with language models
Applied analytics: Accelerating data science workflows without code changes Recommender systems: analysing competition solutions
Infrastructure: Introduction to computer networking AI infrastructure and operations fundamentals

NVIDIA

Generative AI explained

Generative AI Explained

Provider’s curriculum

This introductory NVIDIA course presents the principles of generative artificial intelligence, examples of its applications, capabilities and limitations. It provides an overview of technologies that create new content and the conditions under which they can be used. Treat it as a conceptual introduction before studying individual model architectures or practical development.

SohoBook study guidance

Learning objective

The learning objective is to explain what problem a generative model solves and why its output differs from a verified fact. Compare generation with classification and search: these tasks can coexist in one application, but require different evaluation criteria.

Recommended preparation

No specialised mathematical preparation is needed for an initial introduction. First read the SohoBook introduction and clarify the meanings of data, models and training. While watching, build a glossary that connects each definition to a concrete example of a system in use.

What to focus on

Pay particular attention to the dependence of an answer on its prompt and context, plausible errors and limits of applicability. Discuss why coherent prose does not establish factual accuracy. An interesting question is how a demonstration becomes a working process with sources, quality criteria and human accountability.

Practical value

The material supports informed selection of automation tasks, including drafting, text transformation and assistance with information analysis. For each case, specify acceptable errors, a verification procedure and requirements for input data. This helps define the problem before selecting a particular tool.

Self-check exercise

Choose three academic or professional tasks. For each, describe the expected result, one likely failure and a verification method. Then explain where ordinary search or an algorithm is sufficient and where generation provides additional value. Support your conclusions with examples rather than general judgements about AI capabilities.

Related SohoBook chapters

Open learning resource

Back to collection contents

NVIDIA

A neural network: a “brain” in ten minutes

Building a Brain in 10 Minutes

Provider’s curriculum

This short NVIDIA learning resource introduces how a neural network learns from data and the mathematical operations behind an artificial neuron. The official catalogue lists Python among its tools. The title describes a compact introduction; a full understanding of the mathematics requires further independent study.

SohoBook study guidance

Learning objective

The objective is to connect input features, weights, bias and model output into one computational scheme. Distinguish evaluating an already specified computation from adjusting its parameters during training. This distinction supports later study of the perceptron, loss functions and optimisation methods.

Recommended preparation

Review addition, multiplication, functions and vector notation. Familiarity with variables and simple Python expressions will help you read examples. If the notation is unfamiliar, work through a two-input calculation on paper first, then compare it with the corresponding program expression.

What to focus on

Track which quantities come from observations, which are chosen by the developer and which are determined during training. Examine the role of nonlinear transformations and the meaning of prediction error. Discuss the limits of the biological analogy: an educational mathematical model does not reproduce the entire nervous system.

Practical value

The practical value is being able to understand a small example in full. Instead of memorising library names, explain the purpose of each operation and the change you expect in the output. This approach helps detect dimension errors and understand why changing one weight affects a prediction.

Self-check exercise

Specify two inputs, two weights and a bias, then calculate the weighted sum. Change an input and a weight separately and compare the results. State what additional information training requires, including a target answer and an error measure. Distinguish manually changing a parameter from an algorithm that selects an update using data.

Related SohoBook chapters

Open learning resource

Back to collection contents

NVIDIA · Coursera

AI infrastructure and operations fundamentals

AI Infrastructure and Operations Fundamentals

Provider’s curriculum

The supplied introduction-to-AI-in-the-data-center link leads to NVIDIA’s updated Coursera course, AI Infrastructure and Operations Fundamentals. Its curriculum covers AI, machine learning and deep learning foundations, software stacks, computing platforms and infrastructure operations. Topics include GPUs and CPUs, networking, storage, cloud deployment, monitoring and orchestration.

SohoBook study guidance

Learning objective

The objective is to understand the relationship between a machine learning task and the system that executes it. Explain why model specifications alone are insufficient for choosing hardware, and why training and serving user requests impose different demands on resources.

Recommended preparation

The provider classifies the course as introductory. Basic familiarity with operating systems, files, memory and network connections is useful for understanding the material. Read the SohoBook introduction first; record unfamiliar infrastructure terms separately from the names of algorithms and models.

What to focus on

Follow data from storage through computation to delivery of a result. Pay particular attention to bandwidth, latency, available memory and device utilisation. Compare accelerating an individual operation with improving overall system performance: a fast processor cannot eliminate transfer delays or insufficient memory.

Practical value

The material helps you discuss project architecture with infrastructure engineers and formulate deployment requirements. For a student project, compare a local server with a cloud environment in terms of resource availability, manageability and workload control. Consider observability separately: which measurements would reveal declining performance?

Self-check exercise

Draw a service that accepts a document and returns a model’s processing result. Identify computation, memory, storage, networking and monitoring. For training and inference, name a likely bottleneck and a measurement that could test your hypothesis. Justify your conclusions by tracing data movement rather than listing hardware.

Related SohoBook chapters

Open learning resource Current course page

Back to collection contents

NVIDIA

Augmenting a language model with RAG

Augment Your LLM Using Retrieval Augmented Generation

Provider’s curriculum

This NVIDIA introduction covers Retrieval Augmented Generation: generating answers using retrieved external material. It presents a way to supply a language model with current domain-specific data without retraining it. The course introduces RAG fundamentals and the retrieval process used to provide information for answering a user’s question.

SohoBook study guidance

Learning objective

The learning objective is to explain retrieval and generation as distinct, cooperating stages. Understand that a knowledge source can exist outside model parameters. Compare RAG with an ordinary model prompt and with fine-tuning, identifying what changes in each approach.

Recommended preparation

Begin with an introduction to generative AI and the basic operation of language models. Familiarity with text documents, search queries and datasets is useful. Advanced mathematics is not required for an initial conceptual overview; study implementation details after understanding the overall information flow.

What to focus on

Examine the sequence of document preparation, retrieval of suitable passages, addition of context and answer generation. Distinguish a similar-looking text from a passage that actually supports a claim. Cases where the sources lack an answer or contradict one another are especially instructive.

Practical value

The approach is useful for reference manuals, technical documentation and educational materials. When designing an assistant, assess retrieval quality separately from answer accuracy. Connecting a document collection does not guarantee correctness: verify that the conclusion follows from the retrieved source.

Self-check exercise

Take a small collection of public documents and write five questions, including one the documents cannot answer. Identify a supporting passage or the absence of evidence for each question. Describe the assistant’s expected behaviour and explain how to distinguish retrieval errors from generation errors when checking the result.

Related SohoBook chapters

Open learning resource

Back to collection contents

NVIDIA

Building RAG agents with language models

Building RAG Agents with LLMs

Provider’s curriculum

This practical NVIDIA course addresses language-model systems that use documents and information retrieval to support informed dialogue. Topics include composing interacting components, representing documents, using embeddings and implementing RAG agents. The provider’s description also highlights efficient query execution and scaling systems to meet user demand.

SohoBook study guidance

Learning objective

The objective is to move from a general RAG diagram to understanding a working application. Explain the role of document loading, retrieval, context construction, model calls and response handling. Viewing the system as connected operations makes it easier to validate behaviour and locate faults.

Recommended preparation

Recommended preparation includes confident use of basic Python, functions and data structures, plus the introductory RAG material. Reading error messages and working with notebooks are useful skills. Before running examples, check the requirements of the current course environment and its libraries.

What to focus on

Focus on the relationship between document representation and retrieval quality, dialogue state management and component responsibilities. Reproducibility is particularly interesting: a fixed set of evaluation queries makes system variants comparable. Distinguish successfully starting an application from demonstrating that its answers are supported by sources.

Practical value

The material is useful for building an assistant for technical instructions or academic literature. Begin with a small public corpus whose answers can be checked manually. As load increases, measure retrieval time, generation time and total response time separately, so architectural changes can be justified by observations.

Self-check exercise

Design a RAG application and six evaluation queries, including ordinary questions, a follow-up question and a query with no answer in the documents. Describe each expected result. Propose two system changes and specify measurements that would reveal whether they improved quality or merely changed speed.

Related SohoBook chapters

Open learning resource

Back to collection contents

NVIDIA

Accelerating data science workflows without code changes

Accelerate Data Science Workflows with Zero Code Changes

Provider’s curriculum

This NVIDIA resource introduces GPU acceleration for data processing and machine learning through familiar programming interfaces. Its catalogue entry lists RAPIDS tools—cuDF, cuML and cuGraph—alongside pandas, scikit-learn and NetworkX. The central topic is moving suitable computations to an accelerator without completely rewriting an existing analytical workflow.

SohoBook study guidance

Learning objective

The objective is to understand when compatible interfaces can accelerate computation while preserving familiar program structure. Interpret the title in relation to supported operations and a specific environment. It does not mean that every Python program automatically becomes faster or that environment configuration is unnecessary.

Recommended preparation

Basic Python and experience with table filtering, grouping and joins are recommended. Understand the purpose of model training and result validation. Before practising, distinguish data volume, operation count and available memory, since these characteristics influence the choice of computing resources.

What to focus on

Examine data-type compatibility, operation support and transfer costs between CPU and GPU. Compare the complete task runtime rather than a single accelerated segment. Small datasets are especially instructive: overhead can outweigh the benefit of parallel computation.

Practical value

The approach is useful for repeated analytics and machine learning experiments. A justified migration begins with measuring the existing workflow and checking result agreement. Record library versions, hardware specifications and input size; without these details, performance comparisons are difficult to reproduce and interpret.

Self-check exercise

Choose a table operation and design a CPU–GPU comparison at three data sizes. Specify whether loading is included in runtime, how warm-up is handled and how results will be validated. If no GPU is available, produce an experimental protocol and hypotheses without presenting expected values as measurements.

Related SohoBook chapters

Open learning resource

Back to collection contents

NVIDIA Developer · YouTube · Class Central

Recommender systems: analysing competition solutions

Mastering Recommender Systems · Grandmaster Series

Provider’s curriculum

This is an NVIDIA Developer video seminar from the Grandmaster Series, published on YouTube and listed by Class Central. Participants examine solutions to an e-commerce recommendation competition. Listed topics include a two-stage model, candidate generation with co-visitation matrices, feature engineering for reranking and model ensembling.

SohoBook study guidance

Learning objective

The objective is to understand why recommendation is divided into selecting a limited candidate set and ranking it more accurately. Connect the available user-interaction events to the signal a model uses to predict the next action.

Recommended preparation

Study this material after introductory statistics, machine learning and tabular data processing. Review features, training samples and quality evaluation. Examine a small user-event log beforehand so that the seminar’s terminology relates to a concrete structure of input data.

What to focus on

Distinguish products appearing together in sessions from causal relationships or stable user preferences. When examining a feature, ask whether that information was available at recommendation time. Comparing a competition metric with system quality for a real user is particularly valuable.

Practical value

The seminar illustrates engineering analysis of a complex problem and comparison of strong teams’ solutions. When transferring ideas, consider catalogue size, changing inventory and new users. Competition success is a result under specific data and evaluation conditions, rather than a universal guarantee.

Self-check exercise

Create five short product-viewing sessions and count pairwise co-occurrences manually. Propose candidates for the next recommendation, then describe two features for ranking them. Explain how to evaluate results on future events and why using future information during prediction produces an inflated quality estimate.

Related SohoBook chapters

Open learning resource

Back to collection contents

NVIDIA · Coursera

Introduction to computer networking

Introduction to Networking

Provider’s curriculum

This NVIDIA course on Coursera introduces the purpose of networks, their components and requirements for networking solutions. The curriculum includes the OSI model, TCP/IP protocols, Ethernet fundamentals, addressing and frame forwarding through switches. A separate part covers data-center requirements and the design of the corresponding network environment.

SohoBook study guidance

Learning objective

The objective is to explain how data travels between applications across multiple networking layers. In AI, this supports understanding dataset loading, remote model requests and distributed infrastructure. Network transfer contributes its own portion of task execution time.

Recommended preparation

Confident computer use and a general understanding of servers and client applications are sufficient for getting started. Distinguish a file, a process and a network connection. While studying, create a table connecting each network layer to its function, addresses and an example of the information it carries.

What to focus on