A departing OpenAI engineer wants to build telepathy without implants
Naomi Bashkansky says she is leaving OpenAI to build telepathy. The HN thread immediately shifted from the provocation to the unresolved questions: non-invasive signal limits, privacy, consent, and whether this is communication or a new interface for existing work.
The announcement offers a deliberately expansive ambition—telepathy—without a technical specification in the submission. HN readers asked the questions that will decide whether it is a new interaction method or a slogan: which signals are measurable non-invasively, how much noise remains, and what reaches a model.
Privacy was the most durable concern. A tool that can infer intent must make a hard boundary between deliberate input and private thought, not merely promise a useful transcription experience.
The word telepathy does not answer the interface question
In her announcement, Naomi Bashkansky says she is leaving OpenAI to build telepathy. The title is intentionally provocative, but it also makes the technical and ethical requirements unusually clear: a system that claims to mediate thought cannot be judged like an ordinary text box.
The HN discussion asked first about the channel. One reader asked whether the work is specifically non-invasive, noting a likely noise barrier. Another imagined brain transcriptions as high-signal preference data for model training. Neither is a product description, but together they identify the gap between an aspiration and a workable interface.
Consent must be a mechanism, not a setting
The most important response asked how private thoughts would be prevented from reaching an LLM. That concern is not science-fiction ornament. If a system tries to infer communication from cognitive signals, it needs an observable activation boundary, clear data handling, and a way for a user to know what was captured.
Other replies made the useful distinction between transformative telepathy and a tool that makes existing work more convenient. That distinction should govern the evaluation. A system that helps a person intentionally express a message could be valuable without claiming access to unspoken thought; a system that cannot demonstrate that boundary should not earn the stronger label.
Bashkansky has set a high bar by choosing the word. The next credible update is not a broader promise—it is a precise account of the input signal, the consent model, and the failure cases.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
AI is making progress on Erdős problems; verification is the bottleneck
A Quanta report describes AI-assisted advances on problems associated with Paul Erdős. HN’s response was less triumphal: mathematical results only become useful when people can verify, understand, and build on them.