Naomi Bashkansky resigned from OpenAI on Thursday 23 July 2026 and started at Conduit the following day. She confirmed the move on her personal blog on 4 August, with Tech Times, Gizmodo and Startup Fortune all picking up the story on 6 August. Her Conduit role is Founding Researcher.
Bashkansky spent eighteen months at OpenAI on the alignment team. Public output from that period, per her own blog, includes helping build the OpenAI Alignment Blog at alignment.openai.com, authoring OpenAI’s internal AGI onboarding presentation, and helping advise the AI Resilience division of the OpenAI Foundation. She entered Harvard in computer science at eighteen, and holds a Woman International Master title in chess.
None of Bashkansky’s OpenAI departure has been independently confirmed by OpenAI itself. All sourcing traces to her own personal blog and reporting off it. The move is credible and consistent, but should be read as her disclosure rather than an OpenAI announcement.
Who Conduit is
Conduit is a San Francisco-based non-invasive brain-computer interface company. Co-founders are Rio Popper, an Oxford graduate, and Clem von Stengel, a Cambridge graduate who is CEO. Founding date is reported as 2024 per Gizmodo (single-source). Funding raised and post-money valuation are undisclosed. There are no peer-reviewed publications and no announced institutional investors.
The company describes itself as building high-bandwidth direct brain-to-computer communication using non-invasive hardware. Its public technical disclosure to date is limited to a blog post at condu.it/thought/10k-hours and Bashkansky’s own writing.
The dataset claim
Conduit’s central technical claim is that it has assembled approximately 10,000 hours of what it calls “neuro-language data,” collected from thousands of individual participants over roughly six months. Conduit describes this as “the largest neuro-language dataset in the world,” with an explicit hedge that reads “as far as we know” in the company’s own blog. The claim is scoped to neuro-language rather than to neural data in general. It should not be conflated with Hemispheric’s Descartes model, which reports a different modality mix and different scale.
Collection cadence per company disclosure is approximately 20 hours per day, seven days per week. Sessions run in a San Francisco basement facility.
The commercial approach to data collection
Conduit pays participants $50 per two-hour session, per the company’s own blog. Session cap is ten sessions per participant. Participant recruitment channel is Craigslist per Conduit’s own disclosure; the company reports posting daily listings since April 2026 across the computer, creative and labor-gigs sections.
Session task requirement: participants must touch-type without looking at the keyboard while the headset records neural signal. The custom multi-modal recording helmet weighs approximately four pounds (1.8 kilograms), with 3D-printed housing and commercial sensors disassembled and recombined. EEG is mentioned as one modality in the mix; the company states it is not disclosing the full modality set or headset configurations.
Data storage and inference per Conduit’s own blog: Zarr 3 for data storage; Cerebras running the OSS120B model for LLM responses currently, with various Gemma and Llama models on Groq previously used. This is a standard modern AI-inference architecture rather than a bespoke BCI-decoding pipeline.
Bashkansky’s public thesis
On her personal blog, Bashkansky lays out a specific timeline for what Conduit is trying to build.
By 2027, a wearable neural headband decodes passive intent to accelerate AI coding agents, with an OpenAI Codex-style workflow named as her example. By 2030, AI companies train directly on Conduit’s latent representations, and invasive read plus general write becomes commercially available. By 2035, the technology delivers what she calls a “sixth sense and another limb.” She writes that the resulting market opportunity could exceed one trillion US dollars in valuation.
This is Bashkansky’s own forward projection, on-record on her personal blog, and reads as founder-vision statement rather than consensus market forecast.
Where this sits against prior benchmarks
For non-invasive thought-to-text decoding, the recent published benchmark is Meta AI’s Brain2Qwerty, published in Nature Neuroscience in June 2026. That system reports approximately 29 per cent character error rate with MEG. Meta’s approach uses substantially larger and more expensive hardware than a consumer headset.
Invasive comparisons are further advanced. Stanford’s Frank Willett and colleagues published BrainGate speech decoding at 62 words per minute in Nature in 2023. Ann Johnson at UC San Francisco and UC Berkeley reached 78 words per minute in the same year. Both are surgically implanted and involve neurosurgical enrollment.
Conduit has not published a benchmark comparable to any of these. It has a dataset claim, a hardware prototype, an inference architecture, and a notable hire from OpenAI.
Reading this against the recent coverage arc
The Bashkansky-to-Conduit move lands in the middle of the July-to-August financing acceleration we have been tracking. The Chinese cohort produced four operator signals in July (Neuracle first commercial prescription, NeuroXess xiangqi demonstration, StairMed return-to-work, Wang Ning WAIC wheelchair demo), followed by Active Technology’s record single-angel round of RMB 330 million on 3 August. Delian Capital partner Jiang Donghui warned in VCBeat on 2 August that Chinese BCI financing had run ahead of its translation-to-viable-product timeline.
Conduit does not disclose funding, so it does not sit directly in Jiang’s frame. What it does is add a US non-invasive operator with a notable OpenAI hire to the western cohort that also includes Hemispheric (emerged from stealth on 15 July per weekly-scan reporting), Neurable, Muse and Emotiv on the consumer end and Precision Neuroscience, Synchron and Paradromics on the invasive-clinical end.
The AI-plus-BCI thesis that Jiang argued does not automatically add algorithmic value is exactly the bet Conduit and Bashkansky are making. The near-term test is whether Conduit publishes a decoding benchmark that beats Meta’s Brain2Qwerty at consumer-scale non-invasive hardware.
What to watch
Three near-term milestones matter.
First, whether Conduit publishes a benchmark against Meta’s Brain2Qwerty numbers, or against any prior non-invasive thought-to-text baseline. Until it does, “largest neuro-language dataset” is a data-collection claim rather than a decoding-performance claim.
Second, whether Conduit discloses funding. A San Francisco company running a basement facility at 20 hours of paid data collection per day is spending real money. The identity of the funder shapes how the company reads: if crypto-adjacent capital, similar to Tether’s Blackrock Neurotech position, the reading is different from a specialist deep-tech VC or an AI-lab-adjacent operator.
Third, whether other frontier-AI-lab researchers follow Bashkansky. Whether three or more researchers from OpenAI, Anthropic, DeepMind or Meta AI make similar moves in the next quarter is the labour-market signal to watch.