Machine Learning (ML)

By Ravi Mehta, MS · Updated April 29, 2026

Definition:

Machine Learning (ML) — The broad category of statistical algorithms that learn patterns from data instead of being explicitly programmed. The technology underneath consumer-grade AI features — sleep staging, photo recognition, text generation — and the source of both their gains and their failures.

What ML is, in two paragraphs

Machine learning is the family of techniques where, instead of writing explicit rules, you give a computer a lot of labeled examples and let it find statistical patterns that map inputs to outputs. The output of training is a “model” — a stack of mathematical operations with billions of learned parameters that, when given new input, produces a prediction. The model is doing pattern-matching against the training distribution; it has no understanding of the underlying concepts.

The breakthrough that made consumer ML products possible (~2012 onward) was the deep neural network: a model architecture that could learn from raw inputs (pixels, audio waveforms, text) without manual feature engineering. The breakthrough that made consumer LLMs possible (~2022 onward) was training transformers at scale on internet-scale text corpora. Each step expanded the categories of consumer products that could ship a meaningful ML feature.

Where you actually encounter ML

The consumer-product surfaces where ML is doing the work, not just the marketing:

Photo recognition — your camera identifying faces, your calorie tracker identifying foods, your smart lock recognizing your dog (vs. an intruder).
Sleep staging — your smart ring or smartwatch translating HRV + motion data into a sleep-stage prediction.
Voice assistants — speech-to-text, intent classification, text-to-speech.
Recommendation systems — what shows up on your TikTok feed, your Netflix suggestions, your Spotify Discover Weekly.
Spam filtering — what your email puts in inbox vs. spam.
AI writing and coding tools — the LLMs that power Claude, ChatGPT, GitHub Copilot, Claude Code, Cursor.

In each case, the marketing might describe the feature as “smart” or “AI-powered.” The underlying technology is a machine-learning model trained on a (hopefully large, hopefully representative) dataset.

What ML can and can’t do

ML is excellent at pattern matching within the distribution of training data. It is mediocre or bad at reasoning, multi-step planning, and any task that requires understanding causal relationships rather than statistical correlations. The 2022-2025 wave of LLM products (ChatGPT, Claude, Gemini) has narrowed this gap substantially, but it has not eliminated it. Consumer LLM products remain better at pattern-extension tasks (rewrite this paragraph, summarize this document, generate code similar to existing code) than at out-of-distribution reasoning (debug this novel system, plan this complex project, do mathematics on the frontier of what was in training data).

For consumer trackers, the failure modes show up at the distribution edges. A calorie tracker trained on normal-sized restaurant meals will be more inaccurate on uncommon foods or unusual portion sizes. A sleep tracker trained on conventional-schedule sleepers will be less accurate for shift workers. The marketing won’t tell you about the distribution; the validation studies (when they exist) will.

Why this matters for our verdicts

The performance of every AI-driven consumer product we cover depends on the quality of the underlying ML model and the representativeness of its training data. PlateLens’s 1.1% MAPE in calorie estimation is the result of a photo-recognition model trained on a much larger food-image dataset than its competitors. Claude’s edge in long-form writing is the result of a foundation model that’s specifically trained for the task. The Oura Ring’s sleep-staging accuracy is the result of an ML model that’s been validated against polysomnography references in published research.

When we evaluate “which consumer ML product is best,” we are largely evaluating which company has built the best model for the task. The user-facing product matters too — UI, friction, integrations — but the model is the engine.

What’s coming

The ML capabilities keep getting better at a rate that’s faster than any other consumer-tech component. The capabilities most likely to ship into consumer products in the next 18-24 months: longer-context LLMs (already shipping, now expanding to multi-million-token windows), agent-driven multi-step task execution (already shipping in coding tools, expanding to other domains), and continuous on-device personalization without cloud dependency (early stages). Each of these will reshape some of our verdict categories.

For the photo-recognition models specifically used in calorie tracking and similar consumer-vision products, see photo recognition. For the latency-vs-quality tradeoffs in ML inference, see latency.