The Telepathy Tipping Point: An OpenAI Researcher's Exit and the 10,000-Hour Bet on Brain Tokens

July 23 was not a dramatic day in AI. No frontier model was released; no congressional hearing was scheduled. Yet on that Tuesday, Naomi Bashkansky, a former OpenAI alignment researcher, quietly resigned. The next morning, she started a new job at Conduit as founding researcher. The move took 24 hours; the implications will take a decade to unpack. Bashkansky isn't joining a competitor in the LLM arms race. Conduit is attempting something weirder and far more intimate: decoding non-invasive neural recordings into text that can direct AI agents. She calls it "telepathy." A headband that translates rough intentions into prompts for an AI coding agent by 2027. AI systems consuming neural representations directly by 2030. Two-way "read and write" by 2035. Those dates are optimistic, she admits. They are also a bet on one of the most audacious data pipelines in human history—and one that is remarkably short on public evidence. This is not a story about one researcher changing jobs. It's a story about what happens when brain data becomes training data, and whether we are building a mind-reading interface or an expensive autocomplete. The timeline itself tells a story. 2027: headband decodes rough intentions into prompts for an AI coding agent. 2030: AI systems consume neural representations directly. 2035: two-way read and write. This is not an incremental roadmap; it is a declaration that thought will become the next user interface. In Web3, we learned that interfaces determine power. Whoever controls the interface controls the relationship. If the mind becomes an interface, we are entering an entirely new category of centralization risk. Let's start with the numbers, because that's where the signal lives. Conduit claims to have gathered roughly 10,000 hours of neuro-language data from thousands of people. Participants wore multimodal headsets while typing, speaking, reading, or listening during sessions with a language model. That's a big dataset by any standard. But public proof? The company published only "a few claimed zero-shot examples," without aggregate performance metrics, without an evaluation protocol, and without third-party replication. In crypto terms, this is the equivalent of a protocol releasing a beautiful litepaper, a stunning dashboard, and zero on-chain verification. The market may pump on narrative; the protocol remains unproven. Bashkansky argues that Conduit's results improve with increasing training hours. She calls the work a "greenfield alternative" to the narrower research she could pursue at OpenAI. That is a reasonable argument—volume matters in machine learning. But the available evidence from the broader field suggests that volume alone is not the bottleneck. Meta's Brain2Qwerty experiment reached 61% average word accuracy, and 78% for its best participant. Performance improved log-linearly as data increased. That sounds like a clear scaling law. But the experiment recorded only nine people, all actively typing sentences while their brain activity was captured with magnetoencephalography. Nine people. Typing. That is not free-form thought; that is a constrained motor task with a known output. A Nature Communications study covering 723 participants found that performance improved with more EEG and MEG data. But those participants were reading or listening, not producing language. The reported accuracy was 20% top-1 in a 50-word comparison. Twenty percent. On a vocabulary the size of a cocktail party guest list. The authors concluded that practical non-invasive brain-to-text remains "an open challenge." So when I look at Conduit's 10,000-hour claim, I see both promise and a familiar pattern. Based on my years in Web3—where I've seen protocols with six-figure Telegram communities and no functioning mainnet—data claims are seductive until you ask for the accountant. The key missing piece is not the hours. It's the label. What makes Meta's result meaningful is that the task was typing: the model knew the expected output was text. The brain signal was tied to a physical action, and that action had a known ground truth. Conduit wants to move beyond that. Free-form communication from neural recordings is a category leap, not a linear extrapolation. Think of it this way: a model can learn to recognize the neural pattern for pressing the letter "A" if thousands of people press "A" while wearing a headset. It can even learn to predict which word someone is about to type. But decoding "I'm thinking about whether to switch my DeFi position to a safer chain, but I'm worried about the gas fees" is a different species of problem. It requires enormous context, intention, and semantic grounding—none of which can be captured by watching someone type. There is also a deeper epistemological problem. Thought is not a single signal. It is a construction. If you train a model on text produced after headset-recorded neural activity, the model may not be reading thoughts at all. It may be reading the neural signature of attention, confidence, or confusion, and using language-model priors to fill in the rest. That's not telepathy; that's stochastic autocomplete with a brain-shaped graph. Let me be clear: I believe in the vision. The convergence of AI and brain-computer interfaces is one of the most exciting areas in technology today. I spent 2022 studying ZK-rollups because I believed privacy needed a technical foundation; the same instinct tells me that brain data needs an ethical architecture before it needs more megabytes. But the enthusiasm gap between what Conduit is claiming and what it has demonstrated is the largest risk in this entire story. This is not a criticism of Conduit's mission; it is a demand for evidence. Here is where the contrarian angle cuts against mainstream AI commentary. Most observers will say: "The company needs more data, more compute, more training." That is the standard scaling narrative. My read is the opposite: the bottleneck is not data quantity but data trust. In an industry where vibes easily outrank algorithms, the only antidote is measurement. You cannot scale a reliable telepathy system on noisy, unverified neural data. If 10,000 hours of headset data is collected from thousands of people in uncontrolled environments—people scrolling, blinking, fidgeting, half-watching a video—then more data might simply make the model more confidently wrong. In machine learning, garbage in, gospel out. The credible path requires a public evaluation protocol, third-party replication, and aggregate performance metrics on a constrained task. Without that, the 10,000-hour dataset is an inventory, not an asset. It's a warehouse of unlabeled promises. And there is a second contrarian point: the rush to neural interfaces may be accelerating for the wrong reason. The market value of "brain data" is growing, but so is the potential for a catastrophic privacy breach. Once you connect a headset to a large language model, you are generating the most sensitive dataset ever created. Not just your words, but your pre-verbal intentions. The alignment community at OpenAI spent years worrying about AI systems that might deceive humans. Conduit is building a system that reads humans directly. Which risk deserves more attention? Bashkansky's own background is in AI alignment, and she calls this work "telepathy" with a hopeful tone. But the ethical questions are not future hypotheticals. By 2030, if her own optimistic timeline is right, AI systems could be consuming neural representations directly. Who owns the model that reads a human brain? Who controls the training data? What happens when the user's intention is ambiguous, and the model silently chooses an interpretation? Code is law, but people are truth. If we build brain-reading systems without transparent governance, we will have encoded the biases of a few companies into the hardware of human thought. I keep returning to Conduit's public stance: no aggregate metrics, no eval protocol, no third-party replication. This is precisely the kind of pattern Web3 has taught me to treat with suspicion. A project can have the right team, the right mission, and the right narrative, and still be years away from a working product. The difference between a cult and a community is the willingness to show the ledger. Conduit has a ledger. It just hasn't shown us the entries. So what should we watch for over the next 12 to 18 months? Three things. First, a public evaluation benchmark. If Conduit can demonstrate that its 10,000-hour dataset improves free-form brain-to-text in a controlled setting, with a clear baseline and error analysis, the world should pay attention. Without that, the scaling law is speculative. Second, replication. Meta's Brain2Qwerty result was log-linear with data, but it was also small-scale and task-constrained. Conduit needs independent teams to reproduce its results using different hardware, different protocols, and different participants. Third, an ethical framework for informed consent and data ownership. Neural data is not like browser cookies. It is the electrical signature of a person's inner life. If Conduit can lead on data governance, it will do more for telepathy than a thousand additional hours of training data. The 2027 headband prediction is exciting. But the real question is not whether we can decode rough intentions. It's whether the infrastructure surrounding that technology—the storage, the access rights, the provenance, the ability to opt out—can be built before the first massive leak. In Web3, we say "build in public, live in truth." That phrase has never felt more urgent. Naomi, if you're reading this, I'm rooting for you. Conduit has a rare chance to prove that non-invasive telepathy is not just a data collection fantasy. But 10,000 hours of neuro-language data is a beginning, not a demonstration. The field is full of technologies that work in laboratory conditions and fail in the human world. We have two decades to decide what brains-in-the-loop AI means for privacy, identity, and autonomy. The decisions made at Conduit over the next 18 months will shape that conversation more than any speculative roadmap. Embrace the volatility, find the signal. And please, show us the ledger. The prediction I feel most confident about is not 2035. It's that within the next two years, someone will start a decentralized project to let people own their neural datasets as programmable assets. It will probably fail, likely succeed enough to make a point, and then the industry will have to confront the question Conduit has so far avoided: if thought is the ultimate content, who gets to mint it? That is the real alignment problem. And it is exactly why an OpenAI alignment researcher just staked her career on a company that wants to read your mind.