Somewhere along the way we stopped asking the machine to talk to us.
I want to be careful with that sentence because it sounds nostalgic, and I am not nostalgic. I am not saying the old chatbots were better, or that the conversational turn was a golden age worth returning to. It was stilted and brittle and mostly embarrassing. What I am saying is that the whole enterprise was aimed at something — conversation, understanding, exchange — and at some point the aim quietly changed, and almost nobody admitted it, and now we are paying for the detour in the only currency that matters, which is the years we will not get back.
Think about what the Turing test actually asked. Not whether a machine could optimise a supply chain. Not whether it could call an API and file a report. It asked whether a machine could hold a conversation well enough that you could not tell it from a person. Conversation. The original question about machine intelligence was a question about language between two parties. That is what we were trying to build. And now look at us.
A language system that we only ever showed the text half of
Large language models are, in the plainest possible reading, language systems. They were trained on text because text was what we had in bulk, and text was the first interface, and the first interface has a way of becoming the only interface in the minds of the people building the second version. But there is nothing in the architecture that says language is a thing which exists only as characters on a screen. Language is sound before it is writing. Every language a human being ever spoke was spoken for tens of thousands of years before anybody thought to scratch it into clay. We taught the machine the transcript and wondered why it never learned the room.
And here is the part that keeps bothering me: the capability is already in there. It is not missing. You can hear it, if you listen — the pacing, the register, the way a good language model can put a halt in a sentence at the correct place, or take the measure of what you have not said. Those are conversational competencies and they arrived, more or less by accident, inside models that were never asked to develop them. They are sitting there, dormant, like an entire wing of a house that we built and then locked and forgot about.
The things we chased instead
Meanwhile, the emergent abilities that did get attention were spectacular and undeniably useful. Code generation. Structured output. Tool use. The agentic harness — the scaffolding around the model that lets it plan, call, retry, inspect, and act.
And look, I understand the appeal. I do. The implicit theory of the last two years has been something like this: if we can only give the model enough limbs, enough instruments, enough memory and enough loop, then the harness itself will converge on the thing we have always said we wanted — the automation of high level work. Not a tool that helps a professional, but a system that performs the professional’s job. That is the bet. That is the whole bet.
Whether that bet is sound is a discussion for another day. It might be. There are serious people who think the scaffolding is the actual innovation and the model is just the engine, and they could be right. But it is worth noticing, in the meantime, that it is a bet — not a demonstrated outcome — and that we have arranged an entire global industry around it with the intensity of people who have stopped asking whether they are winning the right game.
Because the resources are not infinite. Compute is finite. Talent is finite. Attention is finite. The best language researchers on earth are, right now, building orchestration layers. The cheapest and most impressive demos are task-completion demos. Every funding conversation I have seen begins with how many steps of an agentic workflow you can run unsupervised, and almost none of them begin with how well the thing can hold a conversation.
What we left on the table
Now let me do something unfashionable and sketch the other history. The one we could have had. Six years, say, of the same intensity, pointed slightly differently.
We could have started with data. Not a bigger crawl of the same English web, but deliberate collection of every language spoken by man. Not as an afterthought, not as a fine-tune on top of an English engine, and not as a marketing footnote for the launch video — but as the founding dataset of the enterprise. Every language, every register, every accent, every regional variant, collected with consent and paid for properly.
Talk about inclusivity. Talk about the common good. There is no more convincing statement of purpose available to this industry than that, and it would have cost a fraction of what one frontier training run costs. Instead the low-resource languages of the world remain a rounding error, and the people who speak them remain, in the most literal sense, out of distribution. That is not a technical shortcoming. That is a choice about who the technology is for, made by default, by people who never had to make it explicitly.
And with that data — with the whole linguistic inheritance of the species, not the profitable ninth of it — we could have pursued genuine duplex audio. Systems that listen while they speak, that handle interruption, that signal understanding with a breath rather than a turn, that know when to be silent. Available to everyone at prices ordinary people could pay. Again, talk about what that unlocks. The education pipeline alone. A patient tutor in every language on earth, not a chatbot that can answer questions but a conversational partner that can actually teach, that can hear confusion in a pause and go back. The medical pipeline. Triage, intake, explanation, counselling — the parts of medicine that are already, in practice, mostly conversation, delivered by systems that cannot hold a conversation.
That is the version of this technology I wanted to live in. It is not technically far-fetched. Every element of it was on the table within the last few years.
What is actually stopping it
If cost is not the obstacle, it is worth being precise about what is. Because the honest answer is less comfortable than the conspiracy, and it is this: the systems we have are not conversational. They are turn-taking machines wearing a conversation’s clothes, and nearly every limitation they show comes from that one architectural fact.
Start with the relay. A conventional voice agent is a pipeline — something transcribes the audio, something thinks about the transcript, something speaks the answer. Three models and two seams. Each seam is a place where the interaction stops being a conversation and becomes a handover. You speak, the system waits for you to finish, it transcribes, it thinks, it speaks, you wait. That is not dialogue. That is a postal service. Real conversation has overlap: people begin before the other has finished, hold the floor with a syllable, yield with a falling tone. None of that survives a design in which the stages take strict turns.
Then the information loss, which is worse. At the first seam, audio becomes text. Everything that is not a word is discarded — pitch, pace, hesitation, breath, the emphasis that changes a sentence’s meaning, the tremor that says the sentence should not be trusted. By the time the language model sees the utterance, it is reading a transcript of a conversation it never heard. And here is what people miss: the model is a language model. Prosody is part of language. We built something that understands language, and then deafened it at the door.
Then latency, which is a distribution and not a number. Three stages in series means three tails in series. Each stage’s average is unremarkable; each stage’s worst case is not; and the worst cases add. This is why a benchmark can report an excellent median while the product feels broken. The median is not what a person experiences — the tail is. A pause of four seconds in the wrong place does not read as slowness. It reads as the machine having lost the thread. And then the person starts talking again, and now there are two speakers and no protocol for that.
Then turn-taking itself, which is barely engineered at all. Backchannels — the small noises that mean I am still here — are almost never produced, and rarely interpreted. Graceful yielding is absent: systems either stop dead or talk over you. Silence is treated as failure and filled. Repair, which is most of what human conversation actually consists of — the half-sentence abandoned, the clarification, the self-correction two words in — is not modelled, because the pipeline has no concept of an utterance in progress. It only has finished ones.
Then memory and identity across a session. Audio is costly to keep in context, so conversational history is truncated far sooner than it should be, and a voice that was recognisable at minute two can drift by minute twenty. Multi-party conversation — three speakers or more — is close to unsolved: diarisation and turn attribution fail exactly where real rooms become interesting.
Then evaluation, which is the quiet disaster. We measure word error rate, median latency and task completion. All three can improve while the experience gets worse, because not one of them measures whether a conversation happened. There is no widely used benchmark for whether a person felt understood, whether the repair worked, whether the pacing was bearable. What is not measured is not funded, and what is not funded does not get built.
And under all of it, coverage. Acoustic modelling is long-tailed, so the further you get from the well-resourced centre of the data distribution, the worse everything performs — recognition, synthesis, prosody. The gap compounds, because poorer recognition yields worse transcripts, which yield worse replies, which yield worse prosody. It is a spiral, not a slope.
I have assembled enough of these systems by hand to know that the wall is always in the same place, and it is never the model. It is the seam between the parts. Which is, incidentally, the one place almost nobody is funding a solution, because it is unglamorous plumbing and it does not demo well.
None of these are mysteries. They are engineering problems with known shapes and obvious tests. They are unsolved not because they are impossible but because they are boring — and the industry has mistaken the interesting problem for the important one.
Why did we lose the ball
Because the industry optimises for what it can show. An agent that books a meeting and writes the notes is a demo you can put in a slide. A model that speaks to an elderly woman in the language she has spoken her whole life, at the right pace, with the right pauses, so that she is not afraid of it — that is not a demo. Nobody claps. It does not compress to a chart.
Because agentic capability fits neatly into a procurement conversation. It maps onto a market that already exists and already has a budget line. Conversation maps onto a market that is diffuse: education, healthcare, public service, accessibility — sectors that are enormous but slow, and which reward patience rather than heroics.
And partly, I think, because conversation is harder to claim victory over. You can declare an agent benchmark passed. You cannot declare that a machine now converses like a human being, not really, not convincingly, not without someone in the room saying — no, not quite, there is something missing, there is a flatness in it. Conversation is the one capability where the test never ends.
That is the irony. We built machines that can do the hard analytical things and still hesitate in the easy human one. The Turing test was never passed. It was abandoned. And we called the abandoning progress, and we wrote papers to justify it, and the justification was that the test was never a good measure anyway — which might be true, but is also exactly what you would say.
What would it take to pick it up again
Three things, and none of them are miracles.
Put language data at the centre, deliberately, for all languages — funded, consented, paid for, and treated as core infrastructure rather than charity.
Make duplex the default rather than the premium feature. A model that listens and speaks at the same time, with the tail latency engineered down rather than averaged away, at a price that makes education and health deployment thinkable in the places that need it most.
And measure the conversation itself. Not task success. Not token cost per objective. Whether a person who does not know they are talking to a machine comes away feeling respected by it. Whether a patient with a bad question got a good answer. That is a research programme. It is a harder programme than building another harness. It is also the programme we started.
I am not saying the agentic work is wasted. I have used it. It works. I am saying it stopped being a means and became the point, and that in becoming the point it absorbed the resources that belonged to the original question. We asked for a machine that could talk to us about anything. We are getting a machine that can do things for us in exchange for being spoken to in an unnatural, overspecified, machine-readable way. Those are not the same product, and only one of them was ever the ambition.
Language is the capability that made us, and the capability we handed to these machines first. We had it in our hands. We have spent a fortune learning to give it hands instead.
Bbumba Emmanuel Ezekiel