Machine Learning Applications: Where ML Earns Its Keep

TL;DR: Machine learning applications pass or fail a single test: build models from patterns in data, then use those models to predict on new data, and this article runs that test over the famous claims, from fraud flags to robotaxis.

• The test comes from Amazon’s own developer guide: patterns in, models built, predictions out. The bullet “find all possible solutions to a problem” describes search and predicts nothing.
• Money learns with limits. Fraud classifiers return an indication of likelihood while human analysts make the call, and the trading edge on record is processing speed and volume.
• Waymo’s robotaxi service is the real thing: Level 4 vehicles, no safety drivers aboard, carrying paying riders since October 2020, about 500,000 rides a week as of June 2026. The consumer Tesla stays at Level 2 assistance under continuous supervision.
• Watson is the counterweight: MD Anderson shelved Watson for Oncology in 2017, roughly $65 million down, and IBM sold Watson Health in 2022 for about $1 billion against $4 billion invested.
• The skipped sectors hold court-tested wins: US courts accepted predictive coding in 2012, tutoring systems have adapted pace since the 1970s, and Left 4 Dead’s AI Director tunes difficulty in real time.

Every machine learning applications list makes the same promise: learning inside your photos, your bank, your feed, even your car. Some entries pass a real test; others are marketing with a model drawn on top. This companion to the Whiteboard Architect episode Where ML Earns Its Keep keeps the ten chapters and the fact-checks, plus the tables and source links a video cannot hold. Every verdict traces to a checked reference; where the record cuts against the story, the record wins.

Board 1 of 17: the app-list promise, genuine versus marketing.

The one test

Amazon’s Machine Learning Concepts guide gives the loop three moves. Algorithms find patterns in data. They build mathematical models from those patterns. The models then generate predictions on data they have never seen, the whole basis for calling this learning.

The tutorial pages slip in a bullet that fails their own definition: “find all possible solutions to a problem”. That describes search, machine learning’s traditional opposite number. Walking a list of known candidates says nothing about tomorrow, while a prediction stakes that claim.

What shapes can a prediction take on this map?

Three shapes cover most of the genuine entries. Classification sorts a new item into a labeled bucket: this transaction looks fraudulent, this scan looks like retinopathy. Scoring attaches a number: the likelihood a customer buys again, the risk a machine fails this week. Ranking orders a list: which song, which job, which video goes first. All three stake a testable claim about something that has not happened yet.

That is the line the test draws. Retrieval and enumeration stay on one side, because a stored league table of popular queries predicts nothing. Anything that outputs the next transaction’s class, the next word or the next click has crossed to the other side, whatever its label says.

Recognition and language

Speech and images are the easiest wins to check, because you use the systems yourself. The voice assistants turn a spoken sentence into a command and answer in a synthesized voice: Siri shipped in 2011, Alexa in 2014, Google Assistant followed, and some still update from each interaction.

Search autocomplete is the borderline case: suggestions rank what everyone else typed, which is lookup, but they also learn your habits and weigh the likeliest next words with a language model. One of the three ingredients ranks stored popularity; the other two learn.

Which named tasks does the language side run on?

The natural language side runs on a set of named tasks, and the middle column is the test applied: each cell is a prediction about audio or text that does not exist yet.

Is autocomplete machine learning or not?

The ingredients can be counted. Popular-query ranking reads a stored league table of what the crowd typed; nothing in it models you. Personal-history adaptation models your patterns as you type: learning. Next-word weighing is a language model scoring continuations: learning. A suggestion list drawing on all three sits at roughly two parts learning to one part lookup, which is why the board writes “partly” and moves on.

This is also the entry that teaches the test. Once you can name which ingredient predicts and which only retrieves, every other entry on the map sorts the same way, one component at a time. One search box spans the whole axis of this map, plain lookup at one end and full learning at the other, which is why the video gives it a board of its own.

Where money learns

Supervised learning at industrial scale is what fraud detection runs: analysts label samples of past transactions as fraudulent or clean, a classifier learns the boundary between the two, and every new transaction crosses it. Credit card fraud detection has used Bayesian neural networks on this job, and the signatures update as new fraud patterns arrive.

Two limits come attached to that win. The classifier’s output is an indication of likelihood, a number for a person to weigh, never a verdict on its own. The analysts stay essential, because the labels come from people and so does the final call.

Trading and advertising run the same loop. Algorithmic trading leans on predictive models to initiate trades while reinforcement learning tunes the policy, but what the record supports is speed and the volume of processing, not superhuman insight. Online advertising runs self-learning systems that work out what converts; privacy and bias concerns attach to every impression.

The recommender era

Board 8 of 17: the recommender era.

The origin story is a competition Netflix ran itself: the Netflix Prize, from 2006 to 2009, offered $1 million to any team that could beat Netflix’s own Cinematch rating predictor by 10 percent. The lesson still fits in four words: everything is a recommendation.

The pattern spread to every feed you open. Spotify’s Discover Weekly has generated a weekly playlist from your listening history since 2015. LinkedIn recommends jobs, people and opportunities while training models on its own knowledge graph. YouTube and Amazon run session-based recommenders, which weigh the current visit rather than the whole history.

Segmentation needs no labels at all. In customer segmentation, clustering algorithms such as k-means surface groups the marketing team never named in advance: the unsupervised family this series introduced earlier, applied to a billing database instead of a classroom example.

How does a recommender combine your history with everyone else’s?

Two inputs, one box, matching the envelopes on the board. Your past behavior is one envelope: what you watched, searched, skipped and replayed. Users like you is the other: people whose histories resemble yours, contributing their outcomes as evidence about yours. Collaborative filtering mixes one user’s history with outcomes from similar users, and Netflix itself runs a hybrid of watching and searching habits.

The output is a ranked list with your next click at the top, refreshed as often as the feed. That refresh is the prediction under test: every click or skip becomes new labeled data for the next round.

Inside industry

A camera turns into a prediction problem once its pixels are data. Computer vision reads images and video as high-dimensional data and produces decisions, from detecting objects and classifying scenes to recognizing faces and segmenting regions. Convolutional neural networks, today’s strongest algorithms here, read medical images and contribute to vehicle steering.

Factories run the quieter win. In predictive maintenance, the model learns which sensor patterns come before failure, so a stop that would have landed mid-shift is scheduled in advance; the repair crew books a slot instead of reacting to an alarm.

Medicine’s record now carries dated numbers, and one caveat travels with the imaging claims: in the studies comparing deep learning with health professionals on medical images, accuracy landed at roughly the professionals’ level, and few of those studies were externally validated: strong at the source hospital, fewer proven beyond it.

Which dated markers does medicine’s record carry?

External validation means testing the model on patients from institutions that did not supply the training data: different scanners, protocols, populations and prevalence. A model that scores as well as the radiologists at its own hospital has cleared one bar. Keeping that score across hospitals clears a higher one, and the literature itself reports that few studies in this family did.

The FDA column answers a separate question: whether a device works as submitted.

Driver assistance is not driverless

Two different machines share this chapter. One is a service: Waymo’s Level 4 robotaxis have run with no safety drivers aboard since October 2020, across ten US metro areas, at about 500,000 paid rides weekly as of June 2026. You cannot buy that car; you book it.

The other machine sits in driveways. The Tesla you can buy delivers SAE Level 2 assistance: the system steers and brakes, a driver supervises continuously, and neither fact moves with the branding. Camera-only Tesla Vision and the end-to-end networks shipped since version 12 still sit inside that envelope. Driverless miles in a Tesla come from the separate Robotaxi service, a different fleet under a different rulebook.

The safety record is measured, weaknesses included. A 2024 meta-analysis rated autonomous vehicles safer across most circumstances and clearly safer for pedestrians, with crash risk at dawn and dusk five times higher. A July 2026 IIHS study found Waymo crashing 68 percent less per vehicle-mile than human drivers.

What do the SAE levels actually promise?

Level 2 means the automation steers and accelerates while responsibility stays with the human: eyes on the road and hands ready for the whole trip. Level 4 means the system handles everything inside its operating domain, a mapped area and conditions it was approved for, without needing a human fallback there. The number between them answers one question: who takes over when it goes wrong.

A Level 2 system with confident lane changes still assigns its failures to the driver, while a Level 4 service operating in its domain carries them itself.

When marketing outran the record: Watson

Board 14 of 17: Watson, decision support against the record.

Watson for Oncology is the story most app lists skip. It shipped as decision support: it suggested treatment options for a physician to consider, and never diagnosis. Real product, real learning inside.

The record then ran the other way. MD Anderson shelved its Watson program in 2017, about $65 million deep. The assessment now recorded in Wikipedia’s machine learning article is blunt: the system “failed to deliver even after years of time and billions of dollars invested”. IBM offloaded Watson Health in 2022 for roughly $1 billion, having spent about $4 billion building the division.

The board’s caption reads far short of the story. Nothing in that record says the learning was fake. It says the marketing claimed more than the product delivered.

The sectors the app lists skip

Three sectors almost never appear on the lists, and one holds the most court-tested win anywhere on the map.

Law’s entry is technology-assisted review, also called predictive coding, and it works like the fraud chapter with documents instead of transactions: expert reviewers code a seed set for relevance, and supervised learning extrapolates those judgments across the collection. US courts accepted the technique in 2012, treated it as settled by 2015, and by 2016 were calling it the “best and most efficient search tool” for document review.

Education’s claim is older and more modest: intelligent tutoring systems have adapted pace since the 1970s, adaptive textbooks and exercises carried it forward, and large-language-model tutors such as Khanmigo are the newest generation. The research record carries a caveat: the effort went into the technology rather than the teaching, so the adapting is proven; tutors out-teaching the lecture is not.

The tour ends with games: dynamic difficulty adjustment reads player ability in real time and tunes the challenge to match. Left 4 Dead’s AI Director is the named example; it paces the hordes and the breathing room against how your run is going.

Three checks

Check the verb. Customer-service chatbots are pitched as replacements for whole call centers; what is deployed handles simple questions as a first point of contact. The replacement is a proposal, and the page’s verb tells you which one you are reading.

Check the date. Uber still turns up on lists of self-driving developers, though it handed its ATG unit to Aurora in December 2020, a $4 billion deal that left Uber a 26 percent stake in the buyer. Aurora has run driverless trucks between Dallas and Houston since April 2025; Uber books partner-built robotaxis.

Then run the test: an entry that predicts nothing about the future has already failed it.

In short

Board 17 of 17: the recap.

A single test sorts the machine learning applications map, patterns in and predictions out. Passing it as genuine learning: recognition, recommenders, fraud flags, document review and game difficulty. Autocomplete passes it halfway, half learning and half lookup. Robotaxis add heavy engineering to learning, and driverless exists only as a service; the consumer car stays supervised at Level 2. Watson’s marketing ran ahead of its record, by billions. The next episode covers where AI and ML hit their honest limits, and the AI in Industry episodes then take the sectors one at a time: health and banking first, then sales, factories and cars.

Continue learning

• Machine Learning: The Honest First Lesson: how AI, machine learning and deep learning nest, and where the learning loop and the four families fit.
• The Branches of AI (and Where ML Sits): where machine learning sits in the AI tree, and which branches are fields rather than labels.
• What Artificial Intelligence Actually Is: the episode this map builds on, with the checked definitions later lessons keep returning to.

Disclosure: this article was drafted with AI from the script of my video lesson, then checked against its sources before publishing.

This lesson first appeared on iqraa.tech.

More whiteboard lessons on the Whiteboard Architect YouTube channel.