The anomaly is not the technology. It is the silence. Conduit, a neuro-decoding startup, claims a dataset of roughly 10,000 hours of neuro-language data, harvested from thousands of human participants. Yet after months of processing and a multi-billion-dollar narrative cycle, they have published only a handful of "zero-shot examples." There is no aggregate top-1 accuracy. No information transfer rate. No evaluation protocol. No third-party replication. No baseline audit.
Meanwhile, Meta released Brain2Qwerty. An average word accuracy of 61% across participants. A best-participant accuracy of 78%. Nine people. Magnetoencephalography. A transparent statistical appendix. The contrast is stark: a giant claiming a slim margin with complete methodological rigor, versus a startup claiming 10,000 hours of data while providing zero verifiable output. This is the classic signature of an unevaluated claim. In my years of tracing on-chain data, the pattern is unmistakable. When a project sits on a mountain of evidence but refuses to show the registry, it is because the audit trail likely breaks the story.
Let the ledger speak. Logic is the only audit that never expires.
Context: The Silent Migration of Smartest Money
On July 23, Naomi Bashkansky resigned from OpenAI. On July 24, she joined Conduit as a founding researcher. One day. There is no gap for reflection. That is the migration speed of conviction. She spent 1.5 years at OpenAI working on AI alignment, a field that is internally unstable, mathematically undefined, and heavily conflicted. In an August 4 essay, she laid out a roadmap: a 2027 headband decoding rough intentions into prompts for an AI coding agent; 2030 with AI systems consuming neural representations directly; and 2035 with two-way "read and write" neural interaction. She calls these "optimistic vignettes." They are not commitments. They are projections. There is no launch date. There is no testnet. There is only narrative.
The institutional translation is crucial here. In the crypto world, we track capital flows to measure conviction. In the AI world, the capital is human. The migration of AI alignment researchers from the established citadel of OpenAI to a "greenfield" startup called Conduit is not a career choice. It is a liquidity event. When the marginal researcher decides that the safety constraints of a central lab are a tax on innovation, they move to the unregulated sector.
The broader market context reveals the same pattern. Tether invested $200 million into brain-computer interface technology in early 2024, aiming to place computers into the human brain. Meta published Brain2Qwerty. The narrative of "telepathy" is being financed by the same speculative overdrive that powered the ICO boom and the DeFi yield fantasy. The difference is that the physical world of neurons cannot be forked. The promises are larger, and the evidence is thinner.
Bashkansky’s essay admits that current thought-to-text studies decode constrained speech-related brain activity. Portable free-form communication remains unproven, even by her own timeline forecasts extending to 2035. Yet she has staked her career on a startup that has not demonstrated a top-1 accuracy score for a fixed vocabulary. This is the "greenfield" argument: She can join a small lab where the research direction is broad and uncharted, rather than the "narrower" work at OpenAI. In 2017 I spent three months cross-referencing ICO wallets, tracing 450,000 ETH against known exchange addresses. The founding team wanted to tell a "decentralized community" story. The ledger told a different story: 68% of early token holders were interconnected entities. The difference between the narrative and the data is the difference between conviction and dogma.
Core: The Evidence Chain and the Missing Registry
The forensic analysis begins with the published baselines. Meta’s Brain2Qwerty is the closest thing we have to a scientifically auditable non-invasive brain-to-text benchmark. The experiment was simple: nine participants typed sentences while wearing a magnetoencephalography (MEG) scanner. The decoder mapped MEG signals to the typed characters. The result was an average word accuracy of 61% and a best participant accuracy of 78%. The study showed that performance improves log-linearly as data increases. That last point deserves scrutiny. A log-linear scaling function is brutal for practitioners. It means that to increase accuracy from 61% to 90%, you need exponentially more data, not linearly more. If nine participants yield 61% average accuracy, then 900 participants might yield 71%, and 90,000 might yield 79%. The asymptote is far from perfect. The "best participant" at 78% is simply the outlier who produced a particularly clean signal. In my NFT wash-trading audit of 2021, I identified 450 interconnected wallets that executed circular trades to inflate floor prices by 40%. The "best participant" in a neural experiment is akin to the "largest trading volume" on an NFT collection: it is often the noise, not the signal.
Meta’s protocol uses MEG, a device that costs between $2 and $3 million, requires a magnetically shielded room, and demands that the participant stay nearly perfectly still. The magnetic fields it detects are measured in femtotesla, roughly a billion times weaker than the Earth’s magnetic field. This is not a headband. The 61% accuracy was achieved under laboratory conditions with millisecond-level temporal resolution and centimeter-level spatial resolution over the cortical surface. Even with those resources, the system misinterprets nearly four words out of ten.
Now let’s turn to the Nature Communications study. A broader study of 723 participants examined EEG and MEG data from people reading or listening, not producing language. They reported a 20% top-1 accuracy in a 50-word comparison. Random chance would be 2%. So 20% is ten times better than chance, but it is still fundamentally insufficient for communication. The authors explicitly said that practical non-invasive brain-to-text remains an open challenge. That is the baseline of the entire field: 20% accuracy on a 50-word vocabulary for comprehension tasks, and 61% accuracy for motor typing with a $2 million MEG machine.
Conduit walks into this landscape claiming a dataset of roughly 10,000 hours of neuro-language data from thousands of people. Let me perform a simple arithmetic audit. If Conduit has "thousands of people," let’s generously assume 2,000 participants. 10,000 hours / 2,000 people means 5 hours of data per participant. That is a wide, shallow pool. Deep decoding of free-form thought requires long-tail articulation, not average peak activity. The data must include sufficient samples of rare intentions, unusual phrasings, and internal mental states. Five hours of typing and speaking with a multimodal headset in a supervised setting is not a deep brain map. It is a surface scan.
The data collection protocol itself raises red flags. Participants wore multimodal headsets continuously while typing, speaking, reading, or listening during sessions with a language model. That is a supervised, task-oriented environment. The model is likely correlating neural signals to the programmatic interface of typing and speaking, not to what we would call "thought formation." The distinction is crucial. When you type the sentence "I want a coffee," your motor cortex generates signals for each keystroke. A decoder can map those signals to text. But if you merely think about wanting a coffee without typing or speaking, there is no motor signal. There is only an abstract semantic representation. No large-scale study has yet decoded abstract semantic representations from non-invasive recordings. The 723-participant study’s 20% top-1 accuracy confirms this: comprehension is easier than production, but even comprehension remains hopelessly noisy.
Where current systems fall short is the gap between movement and thought. Typing and speaking are motor actions. They engage the premotor cortex, the primary motor cortex, the supplementary motor area, and the cerebellum. They also generate massive electromyographic (EMG) artifacts—muscle activity from the jaw, face, and tongue—that can dominate the neural signal. Any MEG or EEG headset picks up a hybrid of neural and muscular signals. The decoder might not be reading thoughts at all; it might be reading jaw clenches, lip movements, or tongue pressure. The 61% accuracy from Meta’s Brain2Qwerty may largely be a measure of motor decoding, not thought reading. The 20% top-1 accuracy from the 723-participant study may be dominated by stimulus-carrier correlations. None of this is true telepathy.
Let’s apply the pre-mortem framework that saved my portfolio during the LUNA collapse. In early 2022, I built a real-time dashboard tracking TerraUSD’s liquidity depth relative to its market cap. My model flagged a critical divergence when stablecoin reserves fell below 60% of circulating supply. I published a warning three weeks before the crash. The divergence was specific: the narrative projected a growing ecosystem, but the ledger showed a shrinking war chest. The same pre-mortem logic applies to Conduit. The thesis is that 10,000 hours of neural data will unlock free-form communication. To invalidate that thesis, we need to see evidence that the decoder fails to generalize beyond motor-command artifacts. That evidence exists. The Nature Communications study showed that even with data from 723 participants, comprehension accuracy stayed at 20%. If Conduit’s dataset is weighted toward typing and speaking—two heavily motor-based activities—then the decoder is learning to map motor commands to text, not mapping abstract thoughts to text. The 2027 headband timeline presupposes that the decoder can identify intent from a portable EEG signal. But portable EEG headsets have terrible signal-to-noise ratios. EEG measures electrical potentials in microvolts. It is heavily contaminated by EMG, eye movement (EOG), and environmental electrical noise. The spatial resolution is 2-3 centimeters—too coarse to distinguish adjacent functional columns in the cortex. A headband cannot see the deep structures involved in semantic memory, such as the hippocampus or the prefrontal cortex.
Let me continue with the math. Meta observed that performance improved log-linearly with data. If Conduit has published zero-shot examples, they exist only as cherry-picked samples. These examples are likely the equivalent of the four best hours from the best participant. They mean nothing. If Conduit had real aggregate performance metrics, we would see a white paper, not an essay. The absence of a technical report with a cohort size, an Evaluation Protocol, and false-positive rates is alarming. It is not a matter of time; it is a matter of structural integrity.
The "data-scale bet" is an old trick. In DeFi, we saw it during the ICO era. Projects would claim enormous "community" sizes and massive GitHub repositories filled with empty repositories and unmerged pull requests. The "data" was quantity, but the signal was noise. Basansky argues that Conduit’s results improve with increasing training hours. That’s a machine learning truism. Any model improves when you feed it more of the same distribution. The real question is whether the distribution is broad enough to cover free-form thought. The answer is no. The 723-participant study, with far more diverse data than Conduit likely possesses, settled exactly this question: non-invasive neural decoding of natural language reading/listening tops out at 20% top-1 accuracy. That 20% does not provide usable communication. It provides a proof of concept. It is the equivalent of a smart contract that only allows you to mint for the first block of each month. It is a toy.
To extrapolate to 2030, when Basansky predicts AI systems consuming neural representations directly, we need to understand what that actually means in the absence of invasive electrodes. The brain does not present a human-readable text interface. It presents a stream of action potentials, synaptic potentials, and high-dimensional dynamical states. Even if you had a perfect non-invasive sensor with zero noise, you would still need to solve the inverse problem: mapping 3D intracranial electrical sources from 2D scalp recordings. This problem has no unique mathematical solution. It is ill-posed. Any decoder is, therefore, a learned approximation, and its errors are systematic.
Contrarian: The Wash-Trading of Neural Signals
The contrarian angle is not that telepathy is impossible. It is that we are confusing a trained decoder with true intention extraction. Correlation is not causation. When Meta reports 61% average accuracy on typing, the decoder is likely using information theoretic priors—the statistical structure of English sentences—to fill in the blanks of noisy neural input. It is not reading minds; it is autocorrecting neural noise with language model priors. Conduit’s "zero-shot examples" may be nothing more than the model successfully guessing the most likely word based on the participant’s limited context window. This is the wash-trading of neural signals.
The best participant at 78% accuracy is the outlier, the excellent user, the one who happened to produce a clean EMG profile. This is akin to identifying a whale wallet that exhibits unusually low transaction latency and using it to prove that the entire network is centralized around fast nodes. It is a sampling bias. The variance between participants is enormous. In any EEG-based study, roughly 30% of participants produce unusable data due to high impedance, hair thickness, excessive muscle tension, or cognitive strategy variance. Conduit’s claim of "thousands of people" does not tell us the rejection rate. If 60% of participants were rejected, the effective dataset size is only 4,000 hours. And even those 4,000 hours are dominated by motor artifacts.
Let me also address the 2035 prediction of two-way read-and-write technology. Writing to the brain involves neurostimulation. Invasive implants, such as those used in deep brain stimulation for Parkinson’s disease, require surgical placement and produce localized effects. Non-invasive transcranial magnetic stimulation (TMS) has poor spatial resolution and can cause seizures at high intensities. The technology to "write" fine-grained semantic information into a specific cluster of neurons through a headband is nowhere in the research pipeline. It is not a data problem. It is a biophysics problem. You cannot train a transformer to overcome the laws of electromagnetism. The mean spatial resolution of portable EEG is insufficient to target a single cortical column. The signal-to-noise ratio is insufficient to bypass the skull’s low-pass filtering effect.
Most interestingly, the narrative of "telepathy" serves a purpose. It attracts capital. The same capital flow that Tether allocated to brain-computer interface research is flowing to Conduit. The narrative is beautifully simple: a headband that reads your thoughts and sends them to an AI agent. It promises a future where you can think about a code command and see it appear on the screen. But this is fundamentally a UX problem, not a human-computer interface breakthrough. Large language models already excel at taking rough, noisy, ambiguous instructions—even ones written in broken English—and turning them into useful outputs. The actual bottleneck is not the interface; it is the instruction. Conduit is trying to eliminate the middleman of typing, but the model still needs to make sense of your often incoherent intention signal.
Furthermore, the alignment community that Bashkansky belongs to is inherently optimistic about capability scaling. They assume that if you feed enough data into a sufficiently large model, it will bridge the noise gap. This is the same false optimism that led LUNA’s design team to believe that an algorithmic stablecoin could withstand a bank run because the market cap would always exceed the available liquidity. The data disagreed. In my three weeks of pre-mortem analysis, I noticed that the underlying protocol had no circuit breaker. Similarly, Conduit has no circuit breaker for the physical noise in their dataset.
The Hard Truth of Scaling Laws
Let me get deeper into the math for a moment. Meta’s log-linear finding is the most dangerous piece of information for the telepathy thesis. The function is roughly: accuracy = a log(b + participants) + c. If we assume that scaling from 9 participants to 90 participants gives a 10% accuracy gain, that is only 61% to 71%. Scaling from 90 to 900 gives another 10%, to 81%. Scaling from 900 to 9,000 gives to 86%. Conduit claims "thousands" of participants. If they have 2,000 participants, they might achieve an average accuracy of maybe 70-75%, assuming perfect data quality with massive filtering. But that is not enough for free-form communication. It is not enough to replace a keyboard. Compare this to invasive BrainGate implants: those have reached 97% accuracy on a small vocabulary. But invasive devices are permanently inside the skull. They are not headbands. The gap between non-invasive and invasive is a chasm of physics, not data.
Conduit’s multimodal headset—combining EEG, EMG, and EOG—appears to integrate multiple noisy signals. The integration can help, but it also introduces contradictory noise sources. EMG from facial muscles might correlate with the emotional state of the participant, adding a variable that has nothing to do with the semantic content. EOG eye movements might correlate with screen attention, adding another confound. The decoder might learn to rely on eye tracking to predict which part of the screen the user is looking at, rather than the thought itself. In one of my audits of an NFT market, I noticed that circular trading volumes were highly correlated with certain wallet clusters. Knowing that a trade happened between a known whale and a known Bot is not the same as knowing why the trade happened. The cause of the trade was likely market manipulation, not organic demand.
The Pre-Mortem Framework Applied to Conduit’s Hurdle
The essential next step is to establish public aggregate performance statistics that link the 10,000-hour dataset to a portable system. Conduit must show that decoding free-form thought can be done with a reasonably sized cohort, and they must publish top-1 accuracy across a large vocabulary, like 1,000 words. If they cannot do that by 2026, the thesis is dead. I will define the specific invalidation threshold: if Conduit fails to publish a top-1 accuracy exceeding 40% on a 100-word vocabulary using a portable EEG headband without reliance on language model priors, then the 2027 prediction should be treated the same way I treated the LUNA stablecoin after the liquidity divergence. The 40% threshold is a step above the 20% from the Nature study and below Meta’s 61% with MEG. A headband that cannot hit 40% is not a telepathy device; it is a parlor trick.
A further invalidation metric: if Conduit claims a 10,000-hour dataset but does not release even a partial neural decoder checkpoint to the scientific community for replication, then it is a protocol with an unaudited contract. I have spent enough time in the DeFi trenches to know that unaudited code is the first thing that goes rogue. The Aave v1 audit I did in 2020 simulated over 10,000 liquidation events and found a critical edge case in the utilization rate formula that would have led to $2.4 million in unsustainable debt. I found this by modeling stress cases, not by reading the marketing docs. Conduit needs to be stress-tested in the same way.
Institutional Translation and the Human Capital Market
The talent migration to Conduit is analogous to early defectors from Ethereum to BSC after 2021. The "narrative" is that Ethereum—like OpenAI—is too restrictive for true innovation. The "data" suggests that the new chain is releasing token economic models that do not survive market stress. When I analyzed the BlackRock ETF and IBIT inflows in the first 100 days, I found that 72% of daily inflows were retained by the custodian—clear evidence of institutional long-term holding. This is what real institutional behavior looks like: accumulated and retained. In contrast, Bashkansky’s move is a capital outflow from a large cap and an inflow to a micro-cap. The "smart money" in AI may be chasing a headband, but the underlying technology has not passed a basic audit.
I would argue that the researcher migration is itself the tradeable signal for the AI sector. When senior safety researchers start leaving the safety lab to found a telepathy startup, it indicates that they have become more confident in capability jumps and less concerned about alignment. This is a classic "flight to alpha" that we saw in the crypto space when DeFi developers left audited platforms to build yield farms on unaudited chains. The results were almost always catastrophic. The structural problem is that the incentive to create a narrative outpaces the incentive to verify the technology. Basansky’s essay calls these predictions optimistic, but the essay itself is a form of marketing. The most efficient audit is to ignore the essay and analyze the data. The data is missing.
Takeaway: The Next Signal to Watch
What I will be observing is not the next essay or the next funding round. I will be watching for a published technical benchmark. Specifically, I want to see top-1 accuracy on a fixed vocabulary of at least 1,000 words, using a portable EEG headband, with an explicit count of participant rejections, and a comparison against the 20% baseline from the Nature study. If Conduit publishes that number and it is above 60%, the 2027 thesis becomes plausible. If they publish only more zero-shot examples, or if they release a vignette about 2035, then they are following the path of every under-collateralized project that I have audited. The market is forgiving of failure, but it is not forgiving of opacity. s silence. Let the ledger speak.
