# Conduit (neurotechnology company)

> Source: https://aiwiki.ai/wiki/conduit_intelligence
> Summary: Conduit (also known as Conduit Intelligence ) is a San Francisco neurotechnology startup that trains AI models to decode text from non-invasive recordings of brain activity.
> Updated: 2026-08-07
> Fact-checked: 2026-08-07
> Categories: AI Companies, Healthcare AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Conduit (neurotechnology company)." aiwiki.ai, 7 Aug 2026. https://aiwiki.ai/wiki/conduit_intelligence
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Field | Value |
|---|---|
| Company | Conduit (Conduit Intelligence) |
| Industry | Neurotechnology, [brain-computer interfaces](https://aiwiki.ai/wiki/brain_computer_interface) |
| Founded | 2024 (per press reports) |
| Founders | Rio Popper, Clem von Stengel |
| Headquarters | San Francisco, California, United States |
| Focus | Non-invasive thought-to-text decoding |
| Dataset | ~10,000 hours of neuro-language recordings (company claim, December 2025) |
| Funding | Not disclosed |
| Website | condu.it |

**Conduit** (also known as **Conduit Intelligence** [15]) is a San Francisco neurotechnology startup that trains AI models to decode text from non-invasive recordings of brain activity. The company describes its mission as building "mind-reading AI to enable high-bandwidth direct brain-to-computer communication": a user puts on a sensor headset, and software predicts the sentences they are about to type or speak from neural signals alone [1]. Conduit was cofounded by Rio Popper, an Oxford graduate, and Clem von Stengel, a Cambridge graduate who serves as CEO [3][6][7]. It drew wide attention on August 5, 2026, when [OpenAI](https://aiwiki.ai/wiki/openai) alignment researcher Naomi Bashkansky announced she had resigned to join Conduit as a founding researcher, writing that "we're training models to non-invasively read the human mind" [4][5].

## Background

Conduit was founded in 2024, according to Gizmodo [6]. Its cofounders are Rio Popper, who studied at Oxford and, per Gizmodo, went blind at age five, and Clem von Stengel, a Cambridge graduate; TechTimes identifies von Stengel as CEO [6][7]. Venture investor Rob Toews profiled the company in Forbes in December 2025 as "one young startup that exemplifies this AI-first, scaling-first approach to non-invasive BCI", noting it was "cofounded by one young Oxford researcher and one young Cambridge researcher" [3]. Bashkansky wrote that [Elon Musk](https://aiwiki.ai/wiki/elon_musk)'s 2017 Wait But Why interview on neural interfaces "inspired one of Conduit's cofounders to go into neurotech" [4].

The company's public footprint is deliberately small: a landing page recruiting paid research participants, a single technical blog post from December 2025, and the founders' quotes in the Forbes piece [1][2][3]. Conduit has not disclosed any funding round, investors, or valuation as of August 2026 [9]. Alongside the founders, the December 2025 post lists team members Nicholas Aldrich and Lydia Nottingham as co-authors, and Bashkansky wrote in August 2026 that "several others" had joined [2][4].

Press coverage frames the company's goal as a wearable "neural headband" that lets users command AI assistants by thought rather than voice or text, with Crypto Briefing noting the same decoding approach could serve people with ALS or locked-in syndrome who have lost motor output [6][9].

## Technology and data collection

Conduit's core bet is that non-invasive neural decoding is a data problem rather than a hardware problem. In a December 7, 2025 blog post titled "How we collected 10,000 hours of neuro-language data in our basement", the company said it had collected roughly 10,000 hours of neural recordings from thousands of unique individuals over six months, which it called, "as far as we know, the largest neuro-language dataset in the world" [2]. Forbes reported the same figure: over 10,000 total hours from several thousand participants expected by the end of 2025 [3].

### Sessions and participants

Participants book paid sessions at Conduit's San Francisco office ($50 for two hours, $100 for four), sit in phone-booth enclosures fitted with chinrests, and hold freeform conversations with an LLM for two hours while wearing a sensor headset, either speaking and listening or reading and typing [1][2]. Participants must be able to touch-type without looking at the keyboard, and each person is capped at 10 two-hour sessions so the dataset stays diverse across individuals rather than deep on a few returnees [1][2]. Sessions run 20 hours a day, seven days a week; recruiting has leaned heavily on daily Craigslist listings, plus paid "participant-ambassadors" who recruit others [2]. Each session produces multimodal neural data time-aligned with audio and text transcripts [2].

The conversational stack has changed over time: as of December 2025 the company used [Deepgram](https://aiwiki.ai/wiki/deepgram) for transcription, an "OSS120B" model served on [Cerebras](https://aiwiki.ai/wiki/cerebras) for LLM responses, and [ElevenLabs](https://aiwiki.ai/wiki/elevenlabs) for voicing replies, after earlier using [Gemma](https://aiwiki.ai/wiki/gemma) and [Llama](https://aiwiki.ai/wiki/llama) models on [Groq](https://aiwiki.ai/wiki/groq) [2].

### Headsets and modalities

Conduit does not disclose its full sensor configuration. The company says it tried various non-invasive modalities (its post names EEG, fMRI, fNIRS, transcranial ultrasound, and MEG as common options) and concluded "you need multiple": no single modality sufficed even with the best available headset [2]. Because no commercial multimodal headset met its needs, Conduit bought high-end single-modality headsets, disassembled them, and recombined the sensors with 3D-printed nylon parts into custom multimodal training headsets weighing about four pounds [2][6]. Forbes reported that Conduit collects primarily EEG data supplemented by other non-invasive modalities, and that decoding performance "improves dramatically" when models train on multiple sensor modalities per user [3]. Training headsets are deliberately sensor-dense and uncomfortable; the eventual consumer "inference" headset is planned as a stripped-down version, with ablation studies determining the minimal sensor set only after the full training run [2].

### Noise at scale

The post's central operational claim is that data quantity eventually swamps noise. Below a few thousand hours, the team treated denoising as mandatory: dry electrodes were spring-loaded for scalp contact instead of using conductive gel (which would have more than doubled the marginal cost per hour by slowing participant turnover), equipment ran on batteries, and power to part of the building was shut off to kill the 60 Hz mains spike [2]. Past roughly 4,000 to 5,000 hours, the company says, models saw enough people and conditions to separate signal from noise on their own, "so we turned the power back on" [2]. Conduit draws an explicit analogy to speech recognition, where models like OpenAI's [Whisper](https://aiwiki.ai/wiki/whisper) improved robustness by trading label quality for sheer volume of weakly supervised audio [2]. Storage and training moved to the Zarr 3 chunked cloud format, with real-time quality checks (modality dropout, timestamp drift, alignment jitter) that the company credits with cutting marginal data cost by about 30 percent [2].

## Reported results

Conduit has not published a benchmark paper or peer-reviewed results; its claims come from its own blog post and founders' statements to press [2][7]. The December 2025 post showed zero-shot examples in which the model, given only neural data from the seconds before a new subject typed or spoke, predicted semantically related text: for the ground truth "the room seemed colder" the model output "there was a breeze even a gentle gust"; for "do you have a favorite app or website" it output "do you have any favorite robot" [2]. The company told Forbes its model achieves roughly 45 percent semantic match with users' intended text, zero-shot on individuals it has never seen; Toews called the figure "not good enough for a mass-market product" but remarkable given the task [3].

Bashkansky characterized progress in scaling-law terms: "the cosine similarity of our latent space predictions with the target latent spaces goes up as a straight line with respect to the logarithm of the number of hours of data", adding "we're in the [GPT-2](https://aiwiki.ai/wiki/gpt_2) era" [4]. She also offered a GPS analogy for why imperfect decoding can still be useful: a noisy signal combined with a strong prior (the LLM plus context) yields accurate output, the way weak GPS plus a map yields accurate navigation [4].

## Naomi Bashkansky

Naomi Bashkansky resigned from OpenAI on July 23, 2026 and started at Conduit as a Founding Researcher the next day, announcing the move on X on August 5 alongside an essay titled "Why I'm leaving OpenAI to build telepathy" [4][5]. The announcement was covered by Gizmodo, TechTimes, Calcalist, Crypto Briefing, and others [6][7][8][9][10].

Born in 2003 and raised in Washington State, Bashkansky played competitive chess from ages 5 to 15, won the World School Chess Championship in the girls' under-13 division and a North American Women's Under-20 Championship title, and holds the FIDE Woman International Master title [4][7][8][11]. She studied computer science at Harvard (class of 2025, with concurrent bachelor's and master's degrees), where she was director of technical programs for the Harvard AI Safety Student Team [7][11]. Her published research spans [interpretability](https://aiwiki.ai/wiki/interpretability) and [AI safety](https://aiwiki.ai/wiki/ai_safety): she co-authored "What Causes Polysemanticity?" (2023) with Stanford collaborators and "Measuring and Controlling Instruction (In)Stability in Language Model Dialogs" (2024) with Harvard's Kenneth Li, Martin Wattenberg, and others [13][14].

At OpenAI she spent 18 months on the alignment team; she says her side projects included creating OpenAI's [AGI](https://aiwiki.ai/wiki/artificial_general_intelligence) onboarding presentation, helping start the company's alignment research blog, and advising the AI Resilience division of the OpenAI Foundation [4][7]. Her essay gives three reasons for joining Conduit: the vision (she estimates that "if Conduit becomes the general read and write company" it would be worth over $1 trillion), the founders, and greenfield research problems compared with working on "a narrow slice" at OpenAI [4].

### Predicted timeline

Bashkansky's essay lays out vignettes she calls "optimistic but highly plausible" rather than commitments [4]:

| Year | Predicted milestone (Bashkansky's personal projection) |
|---|---|
| 2027 | A consumer neural headband passively decodes a user's reactions and intentions during work and pipes them to AI coding agents such as [Codex](https://aiwiki.ai/wiki/openai_codex), with no deliberate subvocalization |
| 2030 | AI companies train models to interface directly with Conduit's latent representations, skipping the text decoder; Conduit expands into invasive reading and general "write" technology |
| 2035 | The interface feels like "a sixth sense and another limb", a human-in-the-loop alternative to AI replacing human work |

She frames consumer BCI as an alignment play: if AI systems are steered by passively decoded human intention, "AI directly empowers humans rather than replacing us" [4]. She also cites earlier arguments for thought-level interfaces from Elon Musk (2017), [Sam Altman](https://aiwiki.ai/wiki/sam_altman)'s 2017 "merge" essay, and Rob Toews's 2025 Forbes piece [4].

## Prior work and competitive landscape

The best published non-invasive brain-to-text results come from [Meta AI](https://aiwiki.ai/wiki/meta_ai)'s [Brain2Qwerty](https://aiwiki.ai/wiki/brain2qwerty) system, which decoded typed sentences at an average 32 percent character error rate using MEG (19 percent for the best participants) but only 67 percent using EEG, in a 35-volunteer study first posted in February 2025 [12]. Meta presented an updated version in June 2026, positioning it as a step toward restoring communication for people with ALS [6][7]. MEG machines cost hundreds of thousands of dollars and require magnetically shielded rooms, which is why Conduit's wearable-EEG-plus-other-modalities approach targets a different tradeoff: worse signal, radically more data [3][7]. Academic datasets in the field have typically involved hundreds of hours and at most hundreds of subjects, against Conduit's claimed thousands of subjects [2].

The contrast with invasive approaches defines the sector. [Neuralink](https://aiwiki.ai/wiki/neuralink) implants electrodes surgically for high-fidelity signals; Conduit cofounder Rio Popper told Forbes that "noninvasive approaches let us collect a much larger and more diverse dataset than we'd be able to if everyone in our dataset had to get brain surgery first" [3]. Toews's Forbes survey places Conduit among a wave of non-invasive ventures alongside Sam Altman's Merge Labs, Bryan Johnson's Kernel (which miniaturized fNIRS into a wearable), and the MIT silent-speech spinout AlterEgo [3]. The premise shared by these companies, and by Bashkansky's essay, is [the bitter lesson](https://aiwiki.ai/wiki/bitter_lesson): [scaling](https://aiwiki.ai/wiki/scaling_laws) data and compute beats hand-crafted decoding algorithms [3][4].

## Skepticism and privacy

TechTimes notes that researchers from the invasive-BCI tradition remain skeptical after decades of EEG language-decoding attempts falling short, and that EEG studies carry a documented risk of overestimating decoding performance because models can exploit temporal autocorrelation within recording sessions; Conduit's zero-shot cross-subject examples avoid that specific failure mode but remain semantic approximations rather than transcription [7]. Conduit has published no benchmark, and cross-participant generalization from a dataset of its claimed size has no direct academic precedent to validate against [7].

Collecting neural data from thousands of paid participants also lands in a regulatory gap. No US federal law specifically governs commercial collection of neural data from healthy adults; California's SB 1223 (2024) classifies neural data as sensitive personal information under the CCPA, with similar provisions in Colorado and other states, and UNESCO adopted a non-binding recommendation on neurotechnology ethics in November 2025 [7]. A 2024 Neurorights Foundation audit of consumer neurotech companies, cited in coverage of Conduit, found the large majority reserved rights to share brain data with third parties; Conduit's own privacy practices have not been independently audited [7].

## See also

- [Brain-computer interface](https://aiwiki.ai/wiki/brain_computer_interface)
- [Brain2Qwerty](https://aiwiki.ai/wiki/brain2qwerty)
- [Neuralink](https://aiwiki.ai/wiki/neuralink)
- [OpenAI](https://aiwiki.ai/wiki/openai)
- [AI safety](https://aiwiki.ai/wiki/ai_safety)
- [Interpretability](https://aiwiki.ai/wiki/interpretability)
- [Bitter lesson](https://aiwiki.ai/wiki/bitter_lesson)

## References

1. Conduit. Official website and research participant booking. condu.it. Accessed August 7, 2026. https://condu.it/
2. Conduit (Nicholas Aldrich, Lydia Nottingham, Rio Popper, Clem von Stengel). "How we collected 10,000 hours of neuro-language data in our basement." December 7, 2025. https://condu.it/thought/10k-hours
3. Rob Toews. "The Next Frontier For AI Is The Human Brain." Forbes. December 7, 2025. https://www.forbes.com/sites/robtoews/2025/12/07/the-next-frontier-for-ai-is-the-human-brain/
4. Naomi Bashkansky. "Why I'm leaving OpenAI to build telepathy." naomibashkansky.com. August 4, 2026. https://naomibashkansky.com/blog/telepathy/
5. Naomi Bashkansky (@NaomiBashkansky). X post announcing resignation from OpenAI and joining Conduit. August 5, 2026. https://x.com/NaomiBashkansky/status/2085043839589617918
6. Webb Wright. "This Startup Will Pay You $50 to Let AI Read Your Mind." Gizmodo. August 6, 2026. https://gizmodo.com/this-startup-will-pay-you-50-to-let-ai-read-your-mind-2000795661
7. Devin Culbertson. "Naomi Bashkansky Quits OpenAI to Build Thought-to-Text BCI on Largest Neural Dataset." TechTimes. August 6, 2026. https://www.techtimes.com/articles/323411/20260806/naomi-bashkansky-quits-openai-build-thought-text-bci-largest-neural-dataset.htm
8. "23-year-old OpenAI researcher and chess prodigy joins startup chasing AI telepathy." Calcalist / Ctech. August 6, 2026. https://www.calcalistech.com/ctechnews/article/sy8o42w8fe
9. "Former OpenAI researcher co-founds Conduit to develop mind-reading AI." Crypto Briefing. August 5, 2026. https://cryptobriefing.com/former-openai-researcher-conduit-mind-reading-ai/
10. "OpenAI researcher Naomi Bashkansky leaves company to build 'telepathy' at Conduit." American Bazaar. August 6, 2026. https://americanbazaaronline.com/2026/08/06/openai-researcher-naomi-bashkansky-leaves-company-485918/
11. Nancy Walecki. "Chess and AI: Naomi Bashkansky." Harvard Magazine. February 7, 2024. https://www.harvardmagazine.com/2024/03/chess-and-ai-naomi-bashkansky
12. Jarod Levy, Mingfang Zhang, Svetlana Pinet, Jeremy Rapin, Hubert Banville, Stephane d'Ascoli, Jean-Remi King (Meta AI). "Brain-to-Text Decoding: A Non-invasive Approach via Typing" (Brain2Qwerty). arXiv. February 2025. https://arxiv.org/abs/2502.17480
13. Kenneth Li, Tianle Liu, Naomi Bashkansky, David Bau, Fernanda Viegas, Hanspeter Pfister, Martin Wattenberg. "Measuring and Controlling Instruction (In)Stability in Language Model Dialogs." arXiv. February 2024. https://arxiv.org/abs/2402.10962
14. Victor Lecomte, Kushal Thaman, Rylan Schaeffer, Naomi Bashkansky, Trevor Chow, Sanmi Koyejo. "What Causes Polysemanticity? An Alternative Origin Story of Mixed Selectivity from Incidental Causes." arXiv. December 2023. https://arxiv.org/abs/2312.03096
15. Naomi Bashkansky. "About Me." naomibashkansky.com. Accessed August 7, 2026. https://naomibashkansky.com/