• Last class closed the module with retrieval-augmented generation
• Four stages: chunk → embed → retrieve → generate
• The whole pipeline was about 50 lines of Python, on your own laptop
• Fine-tuning changes the model’s weights, never its context
• Use it for behaviour and style. Facts that change belong in retrieval
• The ladder to take with you: prompt → RAG → fine-tune
• Climb a rung only when the one below fails

• We start with the money: what a company pays for when it owns servers, and what changes when it rents them
• Two newsroom stories from 2008 make the difference concrete
• Then the vocabulary you will meet everywhere: IaaS, PaaS, SaaS, FaaS, and the virtual machines underneath
• After that, a tour of the AWS catalogue, from compute to storage to databases
• We look at what happens when AWS breaks, because it did, for fifteen hours last October
• You open your own account, choose the free plan, and set a budget that emails you
• We finish by handing a scanned document to Amazon Textract and reading what comes back

• Cloud computing is the on-demand delivery of computing resources over the internet, with pay-as-you-go pricing
• The machine sits in someone else’s data centre, and you rent it by the second
• Four service types: SaaS, PaaS, IaaS, and FaaS
• The “Big Three” are AWS, Microsoft Azure, and Google Cloud
  • AWS holds about 28% of the market, Azure 21%, and Google Cloud 14% (Synergy Research, first quarter of 2026)
  • The world spent $129 billion on cloud infrastructure in those three months, 35% more than a year earlier
• AWS is Amazon’s most profitable division
  • 21% of sales in the second quarter of 2026, and 60% of the company’s operating profit ($16.6 billion in three months)

• Animoto: Lets users create videos from their own photos/music
• Built using Amazon EC2 (Elastic Compute Cloud) + S3 (Simple Storage Service) + SQS (Simple Queue Service)
• Released a Facebook app in mid-April 2008
• More than 750,000 people signed up within 3 days
• EC2 usage went from 50 machines to 3,500 (x70 scalability!)
• No way they could have done this with traditional infrastructure!

• Hillary Clinton’s White House schedule was released to the public in 2008
• 17,481 pages of low-quality PDF, none of it searchable
• Reading it by hand would have taken hundreds of hours
• Peter Harkins, a senior engineer at the paper, asked: “Can we make that data available more quickly, ideally within the same news cycle?”
• He tested optical character recognition programs, then launched 200 EC2 instances
• Nine hours of waiting, 1,407 hours of machine time, $144.62
• The results were online 26 hours after the release

• Middleware is software that connects two separate applications
• The name is literal: it sits between the operating system and the application
• We rarely see it, and the internet stops working without it
• It handles authentication, encryption, and the drivers that talk to databases
• A message queue holds work until a machine is free, so neither side waits for the other
• Animoto used SQS that way: the queue held the video jobs while EC2 grew to 3,500 machines
• Apache Kafka and RabbitMQ do the same job outside AWS

• Suppose Alice has a machine with 4 CPUs and 8 GB of memory, and three customers:
  • Bob wants a machine with 1 CPU and 3 GB of memory
  • Charlie wants 2 CPUs and 1 GB of memory
  • Daniel wants 1 CPU and 4 GB of memory
• What should Alice do?
• Alice can sell each customer a virtual machine (VM) with the requested resources
• From each customer’s perspective, it appears as if they had a physical machine all by themselves (isolation)

• AWS is a collection of cloud-based services
• It’s a very large platform
• Let me say it again: a VERY LARGE ONE 😂
• They offer a wide range of services:
  • Compute, storage, databases, analytics, machine learning, AI, IoT, security, and more
• Many companies use AWS, including Netflix, Airbnb, and NASA
• The most widely used services are EC2, S3, and RDS, for computing, storage, and databases, respectively
• Let’s look at them in more detail

• Amazon EC2 is a web service that provides resizable compute capacity
• It’s designed to make web-scale cloud computing easier
• You can launch instances with a many operating systems (mainly Linux)
• Which means that you can use Bash and run any software you want from the command line
• You can also use a GUI, but it is not necessary
• You can use EC2 to host a website, run a database, run a machine learning model, etc
• https://aws.amazon.com/ec2/

• However, please be mindful of the costs!
• You pay for the instances you use, and the costs can add up quickly
• Per-second (or per-hour) billing
• Data transfer is not included!
• Persistent storage is not included!
• Scaling is not included!

• Scaling is the ability to increase or decrease the compute capacity of your application
• You can do it manually or automatically
• Useful when you have fluctuating workloads, such as a website that gets more traffic during the day, or a machine learning model that gets more requests at certain times
• You set a minimum, desired, and maximum number of instances, and AWS keeps the group inside those bounds
• A rule triggers the change, usually average CPU or the number of requests
• You can also set up load balancers to distribute traffic across multiple instances

• ELB is a service that distributes incoming application traffic across multiple targets
• It handles the varying load of your application traffic in a single availability zone or across multiple availability zones
• ELB is a key component of auto scaling
• As with EC2, you pay for the data transfer and the number of requests
• And it can be quite expensive too!
• It can detect unhealthy instances and reroute traffic to healthy instances, so your application is always available
• Health checks run every few seconds, and a failing instance stops receiving traffic until it recovers
• The load balancer also holds the HTTPS certificate, so your instances do not each need one

• Amazon has three main storage services: S3 (Simple Storage Service), EBS (Elastic Block Store), and EFS (Elastic File System)
• S3 is a web service that provides scalable object storage for data backup, archival, and analytics
• You can store data in buckets, which are like folders
• Data are distributed across a minimum of three availability zones
  • 99.999999999% durability (eleven nines!)
• You can set up lifecycle policies to automatically move data to cheaper storage classes
  • Standard, Intelligent-Tiering, Glacier, Glacier Deep Archive
• You can also set up versioning to keep multiple versions of an object (works like git)

• EBS is a web service that provides block storage volumes for use with EC2 instances
• The difference between S3 and EBS is that EBS is block storage, while S3 is object storage
  • Block storage is like a hard drive, while object storage is like a file system
• Why would you use it? Mainly for SQL files
  • Low-latency, high-performance storage for frequent access
• You can attach EBS volumes to your EC2 instances
• You can also take snapshots of your volumes for backup (like git again!)
• Up to 64 TiB per volume

• EFS is a web service that provides scalable file storage for use with EC2 instances
• EFS is mainly used for shared file storage
  • Multiple EC2 instances can access the same file system
  • Useful for big data and analytics
  • Also for content management and web serving
• You can create file systems and mount them to your EC2 instances
• An EBS volume attaches to one instance at a time, while many instances mount the same EFS file system
• EFS grows and shrinks on its own, so there is no volume size to choose
• As with S3 and EBS, you pay for the storage you use

• Amazon RDS is a web service that makes it easy to set up and operate a relational database
• You can choose from seven popular database engines:
  • Amazon Aurora, PostgreSQL, MySQL, MariaDB, Oracle, SQL Server, and Db2
• It’s a managed service, so you don’t have to worry about backups, patches, or updates
• It works well with another AWS service called Athena, which allows you to query data in S3 using SQL
• The database runs on a machine you never log into. You get a hostname and a port, and no ssh
• Multi-AZ keeps a standby copy in another availability zone and fails over on its own
• You can also use Redshift for data warehousing and business analytics, and DynamoDB for NoSQL databases

• Finally, let’s talk about Amazon SageMaker AI
• SageMaker AI is a fully managed service that allows you to build, train, and deploy machine learning models at scale
  • AWS renamed the original service in 2025, so tutorials still say “SageMaker”
• It has several pre-configured models, such as TensorFlow, PyTorch, and Scikit-learn
• You can also bring your own models too, and use Jupyter notebooks to train them
• Lots of other models available in the AWS Marketplace, and SageMaker Unified Studio is the single workbench for data and AI projects
• Some application services
  • Comprehend (for NLP)
  • Rekognition (Visual Analysis)
  • Translate
  • Polly (text-to-speech)

• Two of DynamoDB’s DNS programs updated the same record at once, and the stale one finished last
• That is a race condition: two processes touch the same thing at the same time, and the result depends on who wins
• The address for dynamodb.us-east-1.amazonaws.com was left empty, so nothing could find the database
• The failure spread to EC2 launches, load balancer health checks, and signing in to the console
• The region took roughly fifteen hours to recover
• Snapchat, Fortnite, Duolingo, Signal, Ring doorbells and several banks went down with it, and 17 million people reported problems
• Analysts put the losses above $550 million

• The account starts with $100 in credits, and the console offers up to $100 more for trying key services
• That is $200 across the six months of the free plan
• The plan ends after six months or when the credits run out, whichever comes first
• AWS puts it plainly: you will not be charged unless you convert to a paid plan
• The free plan blocks some services, and workloads stop at the credit thresholds
• Everything this course asks of you fits inside those limits

• The budget appears in the list with a Healthy status
• Billing and Cost Management home repeats it in the Cost monitor panel
• Billing and Payments, then Credits is where you watch the $200 drain

• Layout: titles, paragraphs, headers, lists, page numbers, and the reading order
• Forms: key and value pairs, the way a paper form pairs a label with its box
• Tables: cells, rows, and columns
• Queries: put a question to the page, such as who signed it
• Signature detection: where a signature sits

Plain text detection is the cheap option. Everything above it costs more per page

• Textract returns blocks in reading order, each labelled by type
• Page 8 came back as a header, a list, and eleven text blocks
• The header reads “SCHEDULE FOR HILLARY RODAHM CLINTON”
• The typo is in the 2001 document. Textract copied the page faithfully
• Anyone searching these files for “Rodham” would miss this page
• Scores on this page ran from 28% to 92%. They rate how sure Textract is about the type of block, not the spelling inside it
• The Raw text tab gives the plain transcript to copy
• The appendix runs an audio transcription job using Amazon Transcribe

• Why cloud computing is important and how it can help your business
• The Big Three cloud computing platforms: AWS, Azure, and GCP
• The main types of cloud computing services: IaaS, PaaS, SaaS, FaaS
• The most popular AWS services: EC2, S3, RDS, SageMaker
• How to create an AWS account and set up a billing alarm
• How to create an S3 bucket, upload an audio file, and transcribe it with Amazon Transcribe
• And that’s just the beginning!
• In the next sessions, we will learn more about AWS services, and how to use them to build and deploy applications
• Remember to close your EC2 instances and S3 buckets when you’re done!

• Imagine that you have a task similar to that of the Washington Post journalists
• You found a series of audio files that you need to transcribe, and you have to do it quickly
• So let’s create an S3 bucket (folder), upload an audio file, and transcribe it
• The file we will use is available here: transcribe-sample.mp3
• Select AWS Management Console to open the console, then search for S3 in the search bar
• Click on Create bucket, give it a name, accept the permissions, and click Create bucket

• Now let’s go to Amazon Transcribe
• Just search for Transcribe in the search bar
• Click on Transcription jobs (left sidebar), then click on Create job
• Give it a name, select the language, and choose the S3 bucket where the file is located
• Click Create