Naomi Bashkansky resigned from OpenAI on July 23. She joined Conduit the next day as a founding researcher. The media immediately framed this as another AI safety leadership exit. Crypto audiences should read it differently. This is a bet that non-invasive brain-to-text can be cracked by data volume, not by a novel algorithm. Bashkansky's Aug. 4 essay predicts a headband that can decode rough intentions into prompts for an AI coding agent by 2027. By 2030, AI systems would consume neural representations directly. By 2035, two-way read-and-write technology. Those are bold claims. But the only number that matters in the story is the one we do not have: Conduit's 10,000 hours of neuro-language data, published with zero aggregate performance metrics.
For readers who live in blockspace, the term "brain-computer interface" might feel adjacent to the chat. It is not. Non-invasive BCI uses EEG, MEG, or fNIRS to record cortical activity and decode it into text. Existing systems work in constrained environments. A person can spell by attending to a specific stimulus or by imagining a movement. Free-form thought-to-text — thinking a sentence and seeing it on a screen — is not proven. Conduit wants to change that. The company says it has collected 10,000 hours of neuro-language data from thousands of people wearing multimodal headsets. Participants type, speak, read, or listen while paired with a language model. The premise is that the model learns to map neural patterns to intended language. Bashkansky, who spent 1.5 years at OpenAI, is staking her reputation on this premise. She also acknowledges that the timeline is optimistic and includes no launch commitment.
That is where I start paying attention. In my years as a DeFi yield strategist, I have seen too many projects claim a massive data asset. Yield farms can cite total value locked. DAOs can cite governance participation. But when the underlying asset is a black box, the claim is not verifiable. The code does not lie, only the audits do. Until Conduit releases an evaluation protocol, its 10,000 hours are a marketing number, not a technical milestone.
Let us examine the scaling curve that exists today. Meta's Brain2Qwerty study, reported in June, used magnetoencephalography to record nine participants actively typing sentences. The system reached 61% average word accuracy, with 78% for the best participant. The researchers observed log-linear improvement in performance as data increased. Log-linear scaling is the single best argument in favor of a data-heavy pipeline. Unlike some neural decoding efforts that plateau, Brain2Qwerty suggests that adding hours makes the model strictly better.
But nine participants is not a population. A 78% best-participant number reveals a wide distribution. It tells me the model can be personalized, but not that it will generalize. For a headband to work in 2027, the model must accept a first-time user who has never been inside a MEG scanner. That requires zero-shot capability. Conduit claims a few zero-shot examples, but without an aggregate metric, we cannot judge whether those examples are the top of a noisy distribution or the median.
There is also the Nature Communications study with 723 participants. The authors found that performance improved with more EEG and MEG data. But the study tested reading and listening, not language production. The top-1 accuracy was 20% in a 50-word comparison — 10 times better than chance, but far below any practical threshold. The study's own authors wrote that practical non-invasive brain-to-text remains an open challenge. Current thought-to-text studies decode constrained speech-related brain activity. Portable, free-form communication remains unproven despite forecasts extending to 2035.
So the field presents three data points: a small study with high per-person accuracy, a large study with weak aggregate accuracy, and one startup with a large private dataset. The startup could hold the missing piece. But the missing evaluation metrics are a red flag. In 2017, I manually audited early-stage smart contracts during the ICO boom. I found critical reentrancy vulnerabilities in two major projects. The code did not lie, but the fundraising material was carefully worded. The lesson is that a claim without a reproducible test is just a narrative. A claim without a metric is a thought, not a result.
Let me add a cost perspective. For a BCI company, data collection is not free. Ten thousand hours of neural recordings requires a sustained participant pipeline. If a session lasts 30 minutes, that is 20,000 sessions. You need recruitment, compensation, operator time, equipment amortization, data cleaning, and labeling. Realistically, the capital deployed is in the seven-to-eight-figure range. That is a serious commitment. But capital does not solve the fundamental problem: neural data is noisy, non-stationary, and heavily participant-specific. A dataset with thousands of people at five hours each is different from a dataset with hundreds of people at fifty hours each. Conduit has not disclosed the per-participant distribution. Breadth improves generalization; depth improves personalization. For a log-linear scaling law, both matter.
In DeFi, we deal with a similar problem when pricing oracle data. A price feed must be transparent about its sources, aggregation method, and latency. An oracle that hides its individual data points is not auditable. Conduit's 10,000 hours is an oracle with a hidden ledger. That is why my operational framework for autonomous AI agents always includes a manual kill-switch. In 2026, I managed an AI-agent trading system that executed 10,000 micro-transactions a week. The system needed real-time data feeds, but every single action had a human review layer for abnormal exposure. The same principle applies to brain-controlled agents: you need a hard stop between thought and action.
Now consider the biometric layer. Neural recordings are indirect identity data. A model trained on your brain patterns can potentially identify your cognitive states, emotional responses, and even latent intentions. That is a liability if the data is processed off-device. If a headband streams recorded signals to a cloud server, the server holds a biometric key. An attacker with that data can train a model that simulates your decision-making tendencies. In crypto, we would call that a private key compromise. The BCI industry has not yet demonstrated an equivalent to self-custody for neural data. If your thoughts can be replayed, your intent is no longer yours.
There is another execution risk. A headband that decodes rough intentions into coding-agent prompts is not just a text input device. It is a command interface. If the decoder produces a wrong token, the coding agent may take a different action. That is a trigger order with no slippage protection. A single mis-signed transaction can liquidate a position. A single mis-decoded intent could cause an agent to push a malformed change to a codebase. At 61% average word accuracy, the error rate is one in thirty or worse. That is unusable for production engineering. Smart contracts execute logic, not intentions. With a brain-controlled agent, the logic is only as good as the decoder.
Bashkansky described Conduit's focus as a greenfield alternative to the narrower research she could pursue at OpenAI. That is a significant signal. OpenAI has a safety mandate that constrains how aggressively models can incorporate neural inputs. Conduit, a small team backed by data, can move faster. But speed without auditability is just acceleration without direction. The company's next evidentiary hurdle is to establish public aggregate performance by linking its 10,000-hour dataset to a portable system that decodes free-form thought. The company has not yet shown that result.
Let me be precise about what would change my mind. Conduit does not need to release its full dataset. It needs to publish a summary of aggregate performance on a fixed test set, with enough detail to allow independent replication. That means specifying the task, the number of participants, the recording hardware, the preprocessing filters, and the decoding model. If the median word accuracy exceeds 70% with a small variance, 2027 becomes plausible. If the result is nearer the 20% top-1 accuracy from the 723-participant study, then 10,000 hours are simply a starting point, not a breakthrough.
The hardware gap is another risk. The headband Bashkansky imagines is not the same as the million-dollar MEG system used in Meta's study. MEG machines require magnetically shielded rooms. EEG headsets are cheap but have poor spatial resolution. A portable headband with dry electrodes is a significant downgrade from the lab. A scaling law learned in a shielded room may not hold in a coffee shop. In my trading operations, a strategy that works in a backtest often fails in an adversarial environment. Neural decoders will face the same backtest-to-live divergence.
Why does this matter for a crypto publication? Because the neural data economy will eventually need an auditable infrastructure. If a model can decode thoughts, the thoughts themselves become an asset. You need to prove where the data came from, how it was labeled, and who has access. That is a data provenance problem. Hashed records and public ledgers are the natural solution. A neural dataset with each session hashed and each participant's consent recorded on-chain would be a legitimate foundation for this new asset class. Conduit has not proposed such a system. Without it, the dataset is a trust-dependent, single-point-of-failure.
Risk Exposure, specifically: 1) Replication risk — no aggregate metrics, no third-party replication. 2) Hardware transfer risk — lab-grade MEG to consumer headband. 3) Biometric data leakage risk. 4) Misclassification risk in intent decoding. 5) Regulatory risk — neural data is increasingly treated as sensitive personal data under GDPR and emerging neurorights laws. A dataset collected from thousands of participants without a clear consent framework is a liability.
Human oversight protocols are mandatory in any AI-related deployment. In my experience with autonomous AI trading, a system without a manual kill-switch is a disaster waiting to happen. Brain-controlled AI agents magnify that need. If the decoder is wrong, the agent could change code, transfer funds, or sign a message. The human must have the ability to veto before execution. The headband itself should be the kill-switch: a button that prevents any decoded prompt from reaching the agent. That is not optional.
The conventional takeaway from this story might be that brain-controlled coding agents are a decade or more away. I think the more dangerous, more contrarian position is that 2027 is too late. By 2027, coding agents will likely have adopted lightweight eye-tracking, gesture recognition, and ambient voice input. A headband that decodes rough intentions may be redundant for routine tasks. However, the underlying dataset — 10,000 hours of thought-language pairs — is a unique asset that goes beyond the interface. The value is not in the headband. The value is in the proprietary data and the alignment it enables.
That is why I see a parallel to Tether's $200 million investment in BCI technology. The bet is not purely about hardware. It is about the neural data layer and the protocol around it. Whoever controls the lifecycle of neural data — collection, labeling, verification, and inference — will set the standard for how intelligence is represented. In crypto, standardization happens through open protocols. The question is whether Conduit chooses to become a closed lab or an open protocol that allows third parties to audit and contribute.
Retail attention will focus on telepathy. Smart money will focus on data provenance and evaluation protocols. The code does not lie, only the audits do. We have to demand the audit before we underwrite the narrative.
The next six months are the tell. If Conduit releases a reproducible evaluation protocol and honest aggregate metrics, the 2027 headband deserves serious consideration. If the company stays quiet, treat its 10,000 hours as a claim, not a fact. I have spent the last nine years learning that markets reward transparency and penalize obfuscation. The neural data economy is about to become as opaque as the early ICO era. That does not mean dismissing the technology. It means asking for the evidence. The code does not lie. Neither does the signal. But both need to be measured.
