How it works
Every connection between layers has a weight — a number the network learned during training. A translation model has billions of them, but this is the whole idea.
A network is layers of learned numbers
Strip away the mystique and a neural network is arithmetic: numbers flow in, get multiplied by learned weights, summed, passed through a simple non-linear function, and flow on to the next layer. No single weight means anything by itself — the behavior lives in the pattern of billions of them acting together.
What makes networks remarkable is that nobody programs those weights. The network starts random and learns them from examples — for translation, from millions of sentence pairs translated by humans.
Embeddings: meaning as geometry
A network cannot read letters, so the first step turns each word — actually each sub-word piece — into an embedding: a list of a few thousand numbers that acts like coordinates in a 'meaning space'. Words used in similar ways end up near each other: 'doctor' sits close to 'physician', far from 'bicycle'.
These spaces capture surprising structure — the classic demonstration is that the direction from 'man' to 'woman' roughly matches the direction from 'king' to 'queen'. Translation works because a model can learn to map the geometry of one language's space onto another's.
Attention: reading everything at once
Older networks read a sentence one word at a time and had to squeeze everything into a running summary — by the end of a long sentence, the beginning had faded. Attention removed that bottleneck: every word can look directly at every other word and decide how relevant it is.
When translating 'the bank was steep and muddy', attention from 'bank' to 'steep' and 'muddy' tells the model this is a riverbank, not a place with ATMs — so it picks the right word in the target language. Stack dozens of attention layers and you get a transformer, the architecture behind every serious translation system today.
The strength of each attention link is itself a learned number. 'Bank' attends strongly to 'steep' and 'muddy', so the model translates it as riverbank.
Training vs. inference
Training is the slow, expensive part: weeks on thousands of specialized chips, adjusting weights over billions of examples. It happens once, in a data center. Inference — actually using the model — is the fast part that happens every time you hit translate, and it can be small enough to run on your phone.
That split explains offline packs: vendors compress a trained model (quantization shrinks each weight from 32 bits to 8 or 4) into something a phone can run. You trade a little accuracy for privacy and independence from the network.
Why bigger models translate better — and their limits
More weights, more layers, and more data mean the network can store finer-grained patterns: rare idioms, technical vocabulary, subtle grammar. This is why cloud translation usually beats offline packs, and why LLM-based translation handles context better than compact dedicated models.
But scale does not fix everything. A model can only reflect its data — languages with little digital text remain weak regardless of model size. And because networks are trained to be fluent, their failures are quiet ones: a confident sentence that reads perfectly and means something slightly different. Good translation apps add guardrails — glossaries, back-translation checks, human review for high-stakes text — precisely because the model itself never signals when it is guessing.
Frequently asked questions
Do translation apps use the same technology as AI chat assistants?
Increasingly, yes. Dedicated translation models and chat assistants are both transformers. Dedicated models are smaller and faster; LLM assistants are better at context and instructions. Many apps now combine both.
Does a neural network understand what I'm saying?
Not in a human sense. It has learned statistical structure so rich that it usually behaves as if it understands — but it has no goals or awareness, and its mistakes (confident mistranslations) reflect that.
Why do offline packs translate worse than the cloud?
Offline packs are heavily compressed versions of the cloud models, so they run on a phone without a connection. Fewer effective weights means coarser patterns and slightly lower accuracy.
