The Telepathy Ledger: Conduit's Missing Metrics and the Data-Scale Bet

Timeline data: July 23, one resignation. July 24, one onboarding. Twenty-four hours between the end of Naomi Bashkansky's 1.5-year run at OpenAI and her first day at Conduit as a founding researcher. In my corner of the industry — crypto infrastructure, Layer2 systems, capital flows that settle in blocks — twenty-four hours is also a dispute window, the time given to a challenger to prove a root state is false. I find the coincidence instructive. Bashkansky is leaving the field of AI alignment for a startup whose stated goal is telepathy: models that convert non-invasive neural recordings into text that can direct AI agents. Her essay, published August 4, predicts a headband that decodes rough intentions into prompts for an AI coding agent by 2027, AI systems consuming neural representations directly by 2030, and two-way read-write technology by 2035. The predictions are explicit. The evidence is not. Conduit's accounting, most recently in a December 2025 update, reports roughly 10,000 hours of neuro-language data collected from thousands of participants wearing multimodal headsets while typing, speaking, reading, or listening during sessions with a language model. That is the entire public evidentiary surface: a few claimed zero-shot examples. No aggregate performance metrics. No evaluation protocol. No third-party replication. In my field, this is an unaudited state root: a claim about what a system settled, with no way to verify the settlement. Trust is verified, never assumed. The technical context matters. Current thought-to-text research decodes constrained, speech-related brain activity; portable free-form communication remains unproven. The two honest reference points in the literature are clear. Meta's Brain2Qwerty, reported in June, reached 61% average word accuracy — 78% for its best participant — using magnetoencephalography on nine people while they typed sentences. Performance improved log-linearly as data increased. Nine people. A shielded room. A constrained typing task. A Nature Communications study spanning 723 participants found that decoding performance improved with more EEG and MEG data, but the study tested reading and listening, not language production, and reported only 20% top-1 accuracy in a 50-word comparison. Chance is 2%. The authors themselves called practical non-invasive brain-to-text an open challenge. Those are the binding constraints of the entire field, and Conduit's announcement does not change them. It sits somewhere between the two points: more data than Meta's nine participants, a more ambitious target than Nature's comprehension task, and no published result to show for either advantage. The core issue is a verification asymmetry. What Conduit published looks persuasive precisely because it is selected, and selection is the enemy of evidence. An evaluation protocol defines how training data is partitioned from testing data, how held-out stimuli are constructed, and how predictions are scored. None of that exists in the public record. The phrase zero-shot, used without a fixed test set and blind scoring, is a stylistic tag rather than a technical specification. I have seen this pattern before. In 2018, I spent six months auditing the 0x Protocol v2 settlement module and found seven critical reentrancy vulnerabilities. No happy-path test suite would have reached them; they existed only in adversarial orderings and abort-behavior states. When I filed the findings, nobody released a statement. The code just changed in quiet commits. That is what verification looks like: silent, specific, and in the code. The public record of Conduit's telepathy work is the inverse of that. It is loud on vision and silent on mechanism. Bashkansky's empirical argument — that Conduit's results improve with increasing training hours — deserves a precise response. The log-linear improvement seen in Meta's typing task is real. But that curve is attached to a constrained task, in a fixed sensor environment, with the participant engaged in physical typing. The Nature study's curve is attached to reading and listening, which produce stronger and more stereotyped cortical signals than intentional language generation. Neither curve demonstrates that free-form thought will behave log-linearly with data. It may. But the premises do not force the conclusion, and a fitted curve is not a causal law. In the Layer2 world, the same confusion appears whenever a transaction-throughput chart is projected onto a security assumption. Throughput and security are different planes. Beneath the hype, the logic remains static: the base-layer constraints do not vanish because the curve looks smooth. There is a deeper problem, and it is the one no press announcement will volunteer: data provenance. Conduit's 10,000 hours were collected with participants wearing multimodal headsets during sessions with a language model. The language model is not a neutral observer of the participant's thought; it is the stimulus. Participants type, speak, read, and listen to text the model produces or scaffolds. That means the neural labels may be as much a record of the model's syntactic influence as of the participant's underlying intention. In crypto terms, this is a self-referential oracle: the system's ground truth is partly generated by the system itself. Self-referential oracles are the source of the most catastrophic failures in this industry. In 2020, I spent three months stress-testing Curve's stablecoin pools against simulated oracle manipulation, building 14 liquidity fragmentation scenarios. The first lesson of that exercise was that the simulator's assumptions determine the results. If you model the oracle as the protocol expects it to behave, you prove the protocol's prior. If you model it as an adversary, you find the holes. The equivalent for Conduit would be an independent lab running the company's decoder against a fixed, pre-registered set of free-form thought prompts, scored blind, with no ability for the company to curate the outputs. Until that test exists, the 10,000 hours remain unvalidated input to an unpublished model. The strategic read is just as clear. Bashkansky describes Conduit as a greenfield alternative to the narrower alignment research available at OpenAI. Greenfield is what infrastructure companies say when they have not yet published benchmarks. It is a claim on territory, not a description of technology. The territory here is the dataset and the network effect of collection scale. I watch this dynamic in the Layer2 rollout wars daily. The decisive variable between the OP Stack and the ZK Stack was never cryptographic; it was which stack could convince more projects to deploy first, because deployment volume generates data, feedback, and default status. If you want to know which Layer2 will dominate, you watch the deployment dashboards, not the white papers. The same rule applies to telepathy: watch the aggregate metrics, when they exist, and ignore the essays. The company with the most verified decode quality will win, not the company with the best timeline. My own verification experience translated directly to this analysis. In 2024, my team audited three major Ethereum Layer2 solutions and identified a critical bug in Optimism's dispute resolution logic — a bug that could have allowed state root manipulation across roughly $2 billion in locked value. It did not surface in normal operations, published test suites, or investor demos. It surfaced only when we modeled an adversarial sequencer and a deliberately broken dispute window. We reported it to the Ethereum Foundation; the patch landed before funds moved. The ledger remembers what the code forgot. The lesson is not that Optimism is unsafe. The lesson is that a system which cannot be adversarially evaluated is a system whose safety claims are untested, regardless of how persuasive its demos appear. Conduit's telepathy pipeline is at that stage now. Now the contrarian angle. There are two blind spots worth naming. The first is distribution shift disguised as a data-collection problem. Conduit's data comes from multimodal headsets in uncontrolled environments. The best published result in the field comes from a magnetically shielded room. A decoder trained on living-room headset data may not generalize to a clinic, an office, or a moving car, because the recorded signal is a composite of the neural source and the sensor's interaction with each new environment. More hours do not fix that; you can only train within the distribution you collect. The crypto equivalent is the optimistic rollup whose fraud-proof window is shorter than the time a challenger needs to sync the chain. The theory is sound; the deployment breaks. Telepathy's deployment problem is the headband itself — the environment shifts every time the user walks through a door. The second blind spot is the moat argument. Ten thousand hours of neuro-language data is presented as an asset that becomes harder to replicate as it grows. That framing assumes a stable rule for converting raw hours into decoding ability. But liquidity is a mirror, not a moat. A liquidity pool that is deep on paper can be empty in practice exactly when it matters, because capital enters and exits along incentive gradients, not structural necessity. A dataset that is large but unverified, collected under uncontrolled conditions, behaves the same way: its size reflects the company's confidence, not its load-bearing capacity. If decode quality depends on a subset of clean hours that no one outside the company can identify, the moat is an accounting artifact. The loudest fact in this entire announcement is also the quietest one. No aggregate number was published. Silence in the logs speaks loudest. In 2018, my 0x findings were honored with quiet commits, not coverage. In 2020, two institutional funds used my oracle stress test because it was falsifiable. In 2024, the Optimism patch happened before money moved because adversarial review existed. In every case, the verification work moved before the narrative did. The narrative around Conduit has not yet reached verification. There is no protocol to review, no metric to contest, no replication to compare. A senior researcher moving from AI alignment to a telepathy startup with 10,000 hours of data and zero published aggregate performance is a press release with a 2030 vision. It is not engineering news. Engineering news arrives when the company publishes enough for an independent party to attempt a falsification. The takeaway is direct. By 2030, we will know whether the telepathy ledger holds real entries or whether 10,000 hours produced a well-lit set of anecdotes. The evidence that would change my assessment is cheap to describe: aggregate word-error rates on a portable system, decoding free-form thought, scored against a pre-registered protocol, replicated by an independent lab. Until those numbers exist, the rational position is the same one I take toward an unaudited smart contract: assume nothing settles until you can verify it. Bashkansky's 2027 headband is a prediction, and predictions are the easy part of every infrastructure story. The proof is the hard part, and the proof is the only part the market can price.