Something peculiar is happening to Neuromancer. William Gibson’s 1984 novel, one of the founding texts of cyberpunk, is finally becoming a television series. Apple has just shown the first teaser; the ten-part adaptation arrives in January 2027. This is good timing, because some of Gibson’s more extravagant ideas are beginning to look less like science fiction and more like slightly inconvenient product roadmaps.
In Neuromancer, Case enters cyberspace through machinery. The same basic assumption survived for decades. In The Matrix, humans plug cables into sockets in their skulls. Neuralink approaches the problem with considerably better engineering and less black leather, but the principle is still recognisable: if brains and computers are to exchange information at high bandwidth, we need a better physical connection. Neuralink’s N1 implant records neural activity through electrodes on flexible threads placed in the brain. The interface is the difficult part.
Now consider a rather different proposal.
Naomi Bashkansky has just left OpenAI to join Conduit, a company that describes what it is building with a word normally best kept away from venture-capital decks: telepathy. The company is training thought-to-text models on non-invasive neural recordings. No socket in the skull. No chrome plug. No cranial USB-C port. A headset watches the faint signals produced while a person thinks, speaks or types, and a model tries to recover the semantic content.
At first glance, this sounds like Neuralink without the surgery.
That understates the idea.
The clever part is that Conduit does not necessarily need to read the brain very well.
Modern language models already know an absurd amount about what humans are likely to say next. They may know the conversation, the document on the screen, the code being edited, the previous commands and the task at hand. The neural signal does not have to contain a pristine sentence. It may only need to provide enough information to choose between a small number of plausible intentions.
Bashkansky offers a useful analogy: GPS. A noisy position signal is not very useful by itself. Put it on a detailed map, however, and a surprisingly vague coordinate can tell you which road you are on. In this picture, the brain signal is GPS and the LLM is the map.
That turns the traditional brain-computer-interface problem inside out. Instead of asking, “How can we extract every word from this dreadful signal?”, we ask, “How little signal do we need before a sufficiently informed model can guess the rest?”
Conduit has taken the scaling approach with enthusiasm. Over six months, it says it collected roughly 10,000 hours of neuro-language data from thousands of people, using several configurations of non-invasive sensors. Participants converse with language models while speaking or typing, producing neural recordings aligned with language. Conduit reports zero-shot semantic predictions for people the model has never encountered before and says performance continues to improve as the dataset grows.
There is independent evidence that the broad direction is not nonsense. Meta’s Brain2Qwerty research has decoded typed sentences from non-invasive brain recordings. The peer-reviewed system reached a character error rate as low as 18 percent for its best MEG participant; its newer successor reports 61 percent word accuracy on average and 78 percent for its best participant.
That is not quite mind-reading over breakfast. MEG equipment is hardly something one casually wears to the supermarket. But the experiments demonstrate that useful linguistic information really is present in non-invasive measurements and that modern neural networks can extract some of it.
And here we arrive at the necessary bucket of cold water.
Conduit’s most impressive claims are still Conduit’s claims. The company has published intriguing examples, but not yet the sort of detailed model publication, independent benchmark and adversarial evaluation that would justify throwing away the keyboard. This matters especially because language models are magnificent guessers. Give one a weak signal and rich context, and it can produce a beautifully plausible sentence that the brain never supplied.
A recent fMRI decoding study demonstrated the danger almost comically well. Its LLM-based decoder appeared to perform impressively—until the researchers replaced the brain input with zeros and obtained nearly the same result. Much of the apparent achievement had come from the language prior rather than the neural signal. In neural decoding, therefore, “the sentence sounds right” is not enough. One has to establish that the information came from the head rather than from the model’s expectations about the head.
Yet this weakness may also be the technology’s strength.
For controlling an AI, perfect transcription may be unnecessary. If I look at an ugly graph and think something approximately equivalent to “those labels are terrible,” my coding agent does not need a court-certified transcript of my inner monologue. It needs enough evidence to infer that I want the labels fixed.
Intent is a much lower-bandwidth object than prose.
This is where the idea becomes more radical than Gibson’s jack. Gibson imagined increasing the bandwidth of the connection. Conduit is effectively proposing something else: reduce the amount of information that has to cross it.
The intelligence on the other side reconstructs the rest.
Bashkansky extrapolates aggressively. In her 2027 scenario, vague thoughts become instructions to AI agents. By 2030, AI systems consume neural latent representations directly, bypassing text. By 2035, she imagines bidirectional systems in which AI is experienced almost as another sense or cognitive limb.
The first step is ambitious but imaginable. The second is speculative but conceptually coherent. The third contains an enormous hidden difficulty: reading and writing are not symmetric. Inferring that I am thinking about a red apple is one problem. Causing my brain to experience an artificial red apple—or an abstract mathematical intuition supplied by a machine—is quite another.
There is another uncomfortable implication. The better such systems become, the less obvious the boundary between decoding and prediction becomes. If the machine knows my work, habits, conversations and immediate context, and receives only a faint neural nudge, did it read my thought or merely predict me exceptionally well?
From the user’s perspective, the distinction might eventually become irrelevant.
From everyone else’s perspective, it probably should not.
Still, Gibson would recognise the direction of travel. Cyberpunk assumed that connecting humans to machines would require us to build a better jack.
The stranger possibility is that we may instead build a machine that needs only a hint.
If the model already knows where we are, what we are doing and what we are likely to want, perhaps the brain does not have to shout.
Perhaps it merely has to whisper.
No comments yet