Machine learning in everyday life

The keyboard that finishes your sentences

Autocomplete on a smartphone keyboard is a language model inference problem running on a device with a few watts of power and a few gigabytes of RAM. Every tap updates a probability distribution over the next word given everything typed so far. The model was trained on billions of words of text and compressed - quantised, pruned, distilled - to fit in an embedded processor. It runs at tens of milliseconds per prediction, invisibly, on every device in your pocket.

The same technology, scaled up by many orders of magnitude, produces the large language models that can draft emails, write code, and hold extended conversations. The keyboard autocomplete and the conversational AI are the same algorithm operating at different scales and with different training budgets. Daily exposure to the smaller version has quietly habituated users to the capabilities of the larger one.

Spam filters, navigation, and power grids

Spam filters were among the first consumer-facing ML applications, and they remain a remarkable example of the technology at scale. Gmail processes roughly 15 billion messages per day; more than 99 percent of spam is blocked before reaching inboxes. The classifier must generalise across every language, every domain, every evasion tactic spammers invent - and it must do so in milliseconds, at essentially zero marginal cost per message. Bayesian classifiers gave way to gradient-boosted trees gave way to neural networks as the adversarial landscape evolved. The current state is a continual arms race that users never see.

Navigation is ML applied to graph search and demand prediction simultaneously. The traffic conditions displayed on Google Maps are not polled from fixed sensors - they are inferred from the aggregated GPS traces of millions of phones, processed through models that distinguish slow traffic from a red light that will clear in forty seconds from a motorway incident that will persist for two hours. Estimated arrival times are the output of regression models trained on historical journey data stratified by time of day, day of week, and weather conditions. The map is a live prediction.

Power grids represent perhaps the least visible and most consequential application. Electricity demand forecasting - predicting load at five-minute intervals across an entire national grid - directly determines how much reserve capacity operators must hold online. Overestimate demand and you burn fuel unnecessarily; underestimate and you risk brownouts. ML-based forecasting models, trained on historical demand, weather, economic indicators, and calendar effects, have measurably improved grid efficiency in countries where they have been deployed, translating to both cost reduction and emissions reduction at scale.

Supply chains, fraud detection, and search

The supply chain disruptions of the early 2020s exposed the cost of brittle inventory management. Modern supply chain systems use demand forecasting models that ingest sales history, supplier lead times, logistics costs, and macroeconomic signals to recommend order quantities and flag potential shortfalls weeks before they appear in stock levels. The model does not eliminate disruption; it gives operations teams lead time to respond.

Fraud detection is a classification problem operating under extreme class imbalance - fraudulent transactions account for a tiny fraction of total volume, but their cost per incident is high. Models trained on transaction graphs, device fingerprints, and behavioural patterns can flag anomalies in milliseconds, pausing suspicious transactions for review before they complete. The false positive rate is as important as the true positive rate: too many blocked legitimate transactions and customers switch to a competitor. The model must be accurate in both directions under real-time latency constraints.

Web search moved from keyword matching to neural ranking a decade ago and has continued evolving since. Query understanding, document representation, and relevance scoring are all now neural, trained on hundreds of billions of search-click pairs. The ten blue links returned for any query are the output of a model that has seen more text than any human will read in a lifetime, compressed into a ranking function that answers in under 200 milliseconds.

Translation, recommendations, and what connects them

Neural machine translation has democratised multilingual communication in a way that is difficult to fully appreciate until you travel somewhere you do not speak the language and navigate entirely through a phone. The models behind real-time translation - trained on billions of parallel sentence pairs across hundreds of language pairs - compress the statistical structure of human language well enough to produce fluent, contextually appropriate text across domains.

Recommendation systems are, in aggregate, among the highest-impact ML deployments in existence. The fraction of content consumed on YouTube, Netflix, Spotify, and TikTok that was suggested rather than actively sought varies by platform, but is consistently above 70 percent. These systems do not simply match users to content they have previously enjoyed; they model the relationship between the two, predicting engagement with content the user has never encountered. The downstream effects - on culture, on information diet, on social cohesion - are significant and still being understood.

What connects all of these systems is not the specific algorithm - the architectures vary enormously - but the underlying philosophy: learn statistical patterns from large quantities of observed human behaviour and use those patterns to make decisions or predictions that serve human goals. The technology is, at bottom, a very powerful mirror. What we see in it depends on what we have put in front of it.