QWE AI Academy Most people still type or talk to ChatGPT like it’s 2023. An OpenAI alignment researcher just quit with the line “I’m leaving OpenAI to build telepathy” – and the post lit up HN, Reddit, and X after it landed.
Naomi Bashkansky left OpenAI on July 23, started the next day as founding researcher at Conduit, and published the full write-up August 4, 2026. Conduit trains thought-to-text models on non-invasive neural data so rough intentions can become prompts without fingers or voice. They’re already paying people $50-$55 per completed 2-hour session (or $100-$110 for four hours) to wear a headset and produce language with LLMs in San Francisco. Check only. Caps apply.
Prompting friction is real. You know the half-formed idea that dies while you hunt for the right words. Better system prompts, voice mode, fancy autocomplete – still the same bottleneck. You have to turn intent into language before the model can help.
Why high-bandwidth work still stalls on today’s interfaces
The bottleneck isn’t model IQ. It’s the seconds you spend packaging a thought. Invasive BCIs (Neuralink-class) aim for higher fidelity but need surgery; most people won’t do that. Academic EEG/MEG sets are still small – hundreds of hours, not tens of thousands.
Meta’s public bar is useful context. Brain2Qwerty v2 (MEG, non-invasive) reported ~61% average word accuracy, 78% on the best participant. Solid lab result. Still not something you book on a Tuesday.
Conduit’s bet is scale plus Sutton’s bitter lesson: gather far more non-invasive neuro-language data and let LLM priors clean the mess. Their own zero-shot sketches (new subjects, no per-person fine-tune) show semantic near-misses an LLM can repair – e.g. a “room felt colder” latent steered toward breeze/gust wording. Noisy signal plus a strong language prior. Same shape as a jittery GPS trace dropped onto a map.
HN skepticism is fair. Scalp sensors are a parking-lot listen on a conversation indoors. Pre-speech intent data is a mental-privacy landmine. A useful consumer headband by 2027 is an aggressive clock. None of that changes the fact that the data flywheel is already running.
Your actual move: get paid to be in the training set (or prep like it)
Consumer headbands aren’t shipping. The concrete path today is Conduit’s research sessions. English speakers who can touch-type without looking at the keys. In-person, San Francisco only (as of the current booking page).
- Open condu.it/booking and grab a free slot.
- Block is 2 hours (or stacked 4). You wear a custom multimodal headset – training rigs are about four pounds, dry spring-loaded electrodes, padding cut to keep contact. No gel. Switch time between people is ~3 minutes; gel setups can eat 30 minutes and dry out, which is why they abandoned them for throughput.
- Freeform chat with an LLM (speak or type) or light read/type tasks. Goal: maximum coherent language. Signals from the seconds before you produce words get aligned to what you actually said or typed – pre-speech / pre-type semantic intent, not a full private monologue dump.
- Pay is by check: $50-$55 for two completed hours, $100-$110 for four. Hard limits (as of booking copy): around 10 two-hour sessions total per person – some older participant pages have mentioned figures like 32, but the live booking flow lists 10 – and at most two sessions in any 24-hour window. They score token quantity and quality for return invites. Company can decline any booking.
Read the ops post before you go: How we collected 10,000 hours. As of that December 2025 write-up, Conduit claimed ~10k hours across thousands of people – their “largest neuro-language set” line. Twenty-hour ops days, personalized LLM chats to keep people talking, ventilation tricks, battery power to cut electrical noise. Past roughly 4-5k hours they argue raw quantity beats most clever noise-reduction work.
Pro tip: Spend 20 minutes a day touch-typing clean sentences on technical or personal topics before you book. Early participants crashed software with wild key combos; keyboards got physically simplified. Looking at the keys or dumping incoherent mash tanks your quality score – and future slots. The model needs real semantic signal.
Not in SF? You can’t sit a session yet. Watch the site. Until then, treat every ChatGPT or Claude thread as interface practice: notice the half-second when intent is clear and words aren’t. That gap is the product target.
What a real session loop looks like
Booth. Headset on. Software live. You talk or type with the LLM about your day, a bug, whatever holds a thread.
Several sensor modalities stream at once – single-modality consumer headsets weren’t good enough on their own, per their ops notes. Data is time-aligned to your output. Training target: map pre-production neural patterns to semantic content. At inference, those noisy latents ride a strong LLM the way a messy GPS trace rides map constraints.
Naomi’s longer roadmap sketches (2027 agent handoff while you make coffee, later latent image transfer, eventual write paths) are vignettes, not a shipping SKU. Exact live accuracy and full hardware stack stay undisclosed. That gap matters when you compare to Meta’s published numbers.
| Aspect | Conduit (as of latest public info) | Meta Brain2Qwerty v2 |
|---|---|---|
| Access | Non-invasive custom headset | Non-invasive MEG (lab) |
| Data scale | ~10k hours, thousands of people (as of late 2025 claim) | Large per-subject; smaller total public corpus |
| Reported accuracy | Not fully public; log-linear scaling claimed; self-described “GPT-2 era” | ~61% word acc avg, 78% best participant |
| Goal | Consumer thought-to-AI + LLM loop | Sentence decoding research |
| You can join? | Yes – paid SF sessions if you pass screening | No |
Numbers and sensor details can move. Conduit stays deliberately vague on the exact suite. No independent public benchmark of their live decoder yet – cosine-similarity-vs-log-hours is a claimed trend line, not a peer-reviewed error rate you can re-run.
Practical prep you can do this week
- Audit one real workday for “intent bottlenecks” – moments you dropped a thought because typing it cost too much.
- Run parallel days: voice-only vs pure text. Which half-formed ideas survive?
- If you code, push partial-intent tools (Cursor, Aider-style agents) and watch how much scaffolding the model still needs when your prompt is mushy.
- Read the primary source for the three-axis diagram (access / direction / scope) and the long-horizon sketches: Why I’m leaving OpenAI to build telepathy.
One open question I keep turning over: once decode quality is good enough that the AI feels like a limb, how do we keep a real “off” switch? The blog frames the project as giving people more agency. Plenty of HN threads jump straight to adtech and compliance nightmares. Both futures are plausible. Who ships the stack decides which one you get.
FAQ
Can I try Conduit thought-to-text right now without flying to SF?
No. No consumer device. In-person paid collection in San Francisco is the only public path. Watch condu.it.
Is this reading my private thoughts or just speech preparation?
Per their write-ups and examples, training targets semantic content in the seconds before you type or speak – intent about to become language – not a full internal monologue export. An LLM turns the noisy latent into text. That still raises mental-privacy issues (UNESCO and others are already drafting frameworks). Assume anything under the headset could eventually be inferred. Don’t treat a research booth like a diary vault.
How does the $50 session help me use AI better today?
It doesn’t hand you a headband. It funds data. The side effect people actually report: two hours of forced clean language output makes sloppy prompting feel obvious. You also see the unglamorous ops layer – spring tension, quality scores, check payment, the ~10-session cap – that pure hype threads skip. If the scaling curve holds, early participants are part of the pre-training distribution for an interface a lot of us may use later.
Local and already a touch-typist? Book at condu.it/booking. Everyone else: start a one-week log of intent-to-prompt gaps. The resignation post is new. The flywheel isn’t.