She Left OpenAI to Build Telepathy. The Machine Won’t Read Your Mind. It’ll Finish Your Sentences.

Mon, 14 Sep 2026 15:13:23 +0000

Two weeks ago, a 23-year-old alignment researcher named Naomi Bashkansky resigned from OpenAI. The next morning she started as a Founding Researcher at Conduit, a San Francisco startup with a mission statement that reads like a rejected Black Mirror pitch: thought-to-text. Telepathy. No implants, no surgery, just a headband that reads neural signals and types what you’re thinking.

LinkedIn did its thing. BREAKING. Someone is building it NOW. Is BCI the next platform?? Sirens, reposts, a thousand hot takes about Professor X.

And look, when someone walks out of the most valuable startup on Earth because she thinks the keyboard is the bigger problem, attention is warranted. Also, delightfully, she’s a Woman International Master in chess, so she has professional experience thinking many moves ahead while everyone watches her hands.

But I read her actual essay, and the interesting story isn’t “mind reading is coming.” The interesting story is a technical confession hiding in her own explanation of how it works, one that makes the whole thing simultaneously less spooky and far more likely to succeed. Let me show you.

The confession: it’s not a mind reader. It’s a world-class guesser.
Here’s the analogy Bashkansky herself uses, and it deserves to be famous. A noisy GPS signal alone can’t pin down your location. But combine that noisy signal with a map and a navigation route, and suddenly it’s remarkably accurate. In her system, the noisy GPS is your brain signal. The map is an LLM plus your context, what’s on your screen, what you were just doing, what a person like you would plausibly say next.

Sit with what that means. The headband doesn’t decode your thoughts with precision. It catches a fuzzy, static-ridden whisper of them, and then a language model, armed with context and its priors over what humans say, guesses the rest. It’s not telepathy in the comic-book sense. It’s autocomplete with a neural nudge. Your brain provides the hint; the LLM provides the sentence.

Deflating? Slightly. But here’s the twist: this is exactly why it might actually work.

Fifty years of failure, dissolved by an accident
Brain-computer interfaces are old. EEG decoding has been “ten years away” since before I was born, and the villain was always the same: signal-to-noise. Reading through the skull is like eavesdropping on a conversation through a hotel wall, during a wedding, in the monsoon. So for five decades the field did the only sensible thing, it tried to clean the signal. Better electrodes, better filters, better maths. And it kept hitting the same wall.

Then LLMs arrived, for completely unrelated reasons, and quietly changed the question. Because if you have a model with world-class priors over human language, you no longer need a clean signal. You need just enough signal to pick between the guesses.

The bottleneck wasn’t broken. It was made irrelevant, by a technology built for chatbots. Nobody in neurotech ordered this; it showed up anyway, like a neighbour’s Wi-Fi that happens to reach your house. Conduit’s founding bet is essentially: the field hasn’t noticed its oldest problem is gone.

The bet, quantified: 10,000 hours in a basement
So if clean signal isn’t the bottleneck, what is? Data. Which is where this story acquires my favourite detail: over six months, Conduit collected roughly 10,000 hours of paired brain-and-language recordings from thousands of volunteers, in a basement studio in San Francisco, the largest dataset of its kind ever built. Somewhere in that sentence is a documentary I’d watch.

Bashkansky’s essay applies the Bitter Lesson, AI’s most reliable heuristic: stop hand-crafting clever algorithms, throw orders of magnitude more data at a big model. She reports that decode quality is rising as a clean straight line against the logarithm of data hours, the same scaling behaviour that took language models from party trick to industry. Her exact phrase for where they are on that line:

”We’re in the GPT-2 era.”

Now, founders’ scaling curves shown without axes deserve the same polite squint as market-size projections. But take the framing seriously for a second, because it’s doing two jobs. GPT-2, you’ll recall, was the funny little model that wrote incoherent news articles in 2019. Nobody’s boss cared about GPT-2. Three doublings of scale later, everybody’s boss cared about nothing else. Saying “we’re in the GPT-2 era” is the most honest possible way to say the demo is unimpressive and the trajectory is the entire company. The roadmap isn’t decoration on this pitch. The roadmap is the pitch.

The part I can’t stop thinking about, and it’s not the headband
Her essay’s 2027 vignette is charming: she makes coffee while vague thoughts about plot labels get decoded into crisp instructions for her coding agent. She stresses she isn’t shouting words in her head; she’s just thinking, and it feels like magic.

I believe her. And precisely because I believe her, my architect brain snags on the supply chain behind the magic. Follow it: the product only works because a model holds priors over what you would say, fed by context about what you’re doing, correlated against recordings of what your brain does while you think. Every hour of use makes the model better at predicting you, specifically.

We have spent the last two years, on this very publication, watching companies fumble the governance of chat logs. Prompts, of all things, became discovery material in lawsuits. Now extend the line: the training corpus here is the correlate of inner monologue. There is no more intimate dataset that can exist. “Non-invasive” is doing heroic work in that sentence, no scalpel touches you, true, but invasiveness was never really about the skin.

None of this makes Conduit sinister; by all accounts they’re building carefully, and Bashkansky came from alignment, of all teams. It makes something else true, the thing regular readers can recite by now: when a capability this powerful becomes real, the technology is the easy half. The company that wins won’t be the one that decodes thoughts best. It’ll be the one people trust to hold them. Consent architecture, deletion guarantees, audit trails for the thought-stream, boring, unglamorous, and worth more than the decoder. The premium migrates to trust; this time it’s just wearing a headband.

The cool-down
Three takes to carry out of here.

One: the transferable insight is bigger than BCI, LLM priors can rescue any noisy sensor. Messy handwriting, degraded audio, cheap medical sensors, garbled field data: if a strong prior exists, “too noisy” stopped being a verdict in 2023. Audit your own dead projects with this lens.

Two: watch data-per-rupee, not demos. Conduit’s moat, if it materializes, is the dataset flywheel, cheap headbands generating hours nobody else has. The demo will look bad right up until it doesn’t; that’s what GPT-2 era means.

Three: if this works even partially, the interface layer of computing changes owners, and the trust layer becomes the product. Position accordingly, whichever side of that trade you’re on.

A researcher left the company building artificial minds to join one learning to listen to real ones, and her honest pitch is that the machine won’t read your thoughts so much as finish them.

Half of me finds that reassuring.

The other half notes that my phone’s autocomplete already knows me embarrassingly well, and it’s only ever seen my thumbs.

-Hardik

I’ve spent about two decades building data and AI systems, and I’ve learned that every “impossible signal” problem eventually loses to someone with a bigger dataset and a better prior. I write about AI’s real mechanics, its hype cycles, and the fine print under the magic. If you’d rather understand the trick than just applaud it, subscribe. My autocomplete predicted you would.

Sources / further reading: Naomi Bashkansky, “Why I’m leaving OpenAI to build telepathy” (naomibashkansky.com, Aug 4, 2026), the primary source for the vignettes, the GPS-and-map analogy, the Bitter Lesson argument, and the scaling claims; read it in full, the 2027 vignette is worth your five minutes. Conduit’s own post on collecting the 10,000-hour neuro-language dataset (condu.it). News coverage of the departure and Conduit’s founding team, Rio Popper and Clem von Stengel (Calcalist, ynet, TechTimes, Aug 2026). Background: Rich Sutton’s “The Bitter Lesson” (2019); Tim Urban’s Neuralink deep-dive on Wait But Why (2017), cited by Bashkansky as an origin story for one of Conduit’s co-founders. Commentary and any leaps beyond the sources are mine.

Coder Legion - Recent posts tagged arti

She Left OpenAI to Build Telepathy. The Machine Won’t Read Your Mind. It’ll Finish Your Sentences.