A few years ago, a machine translation system knew one thing about your sentence: the sentence. No speaker. No purpose. No setting. No memory of the previous turn. Statistical mapping over decontextualised strings, and whatever quality you could squeeze out of that.
Today I can hand a live interpreting system the domain, the topic, the names and roles of the people in the room, the institution, a paragraph describing what the meeting is about, a terminology list, and forced renderings for specific terms — and it will actually use all of it. That is not a better version of the old thing. It is a different thing.
The examples are almost embarrassingly simple, which is precisely the point. English “you” becomes German “Sie” or “du”, and nothing in the sentence tells you which — only the setting does: a courtroom and a kitchen table give opposite answers. “I’m tired” is one word longer in Spanish depending on who is speaking, and no amount of source-side analysis will tell you whether it is “cansado” or “cansada”. In a legal hearing, “discovery” is not a discovery. “Charge” in the same conversation may be a fee, an accusation, or an instruction to the jury, and the deciding evidence was three turns ago. A speaker says “right” and means the opposite of yes.
Each of these used to be a known, unfixable failure, worked around with glossaries and post-editing. Now they are handled by the same mechanism that handles everything else: the system knows what is going on.
Understanding, in the ordinary sense of the word, is the connecting of knowledge to context: grasping how things relate and applying that relation to a situation you have not seen before. That is exactly the capacity a configuration like the one above is buying. You cannot make use of the information that the speaker is a physician addressing a frightened patient unless you can work out what follows from it. Decades of machine translation could not use that information — not because nobody thought to supply it, but because there was nothing in the system that could do anything with it once supplied.
This is what I spend my days on: systems that do not translate and stop there, but that model the whole communicative situation — in live machine interpreting, and in high-quality dubbing, where a line that is accurate but tonally wrong is simply a bad line. Setting, roles, purpose, conversational history: these used to be what you stripped away to get to the sentence. Now they are half the signal.
Those examples are kindergarten level. Multiply the difficulty a hundredfold and frontier models still get them right. The reason is straightforward. Frontier LLMs have started to understand language as communication rather than as strings, and understanding is exactly what context requires — a system that cannot tell why something is being said has no way to use the information that it is being said in a hospital, by a physician, to a frightened patient. Applications that were unthinkable eighteen months ago are now engineering problems. In a next post, in preparation now, I will focus on why LLMs DO understand, albeit in a different way then humans, and that translation and interpreting is the perfect ground for such “alien” understanding to thrive.
I have been studying speech translation for years. This is the first time the ground has moved this fast.
The pic shows a typical “initial prompting” of modern translation pipelines, as used in machine interpreting and dubbing. It is used for example in www.interpreter24.com, but many API providers are increasingly offering them. This initial prompting can be also updated in realtime to take into consideration the evolving nature of communication.