Former OpenAI Researcher Joins Conduit to Develop Thought-to-Text AI Models

Artificial intelligence has spent the last few years learning to understand language, images, and code. Now, one former OpenAI researcher believes the next frontier is understanding human thought itself. Naomi Bashkansky, formerly an alignment researcher at OpenAI, has left the company to join Conduit as a founding researcher. The startup is pursuing one of the most ambitious goals in AI today: developing systems that can translate human thoughts into text using non-invasive brain signals. The announcement is notable not only because of the technology involved but also because it reflects a growing trend of senior AI researchers leaving major companies to build specialised startups focused on the industry’s next wave of breakthroughs. Naomi Bashkansky announced on X that she left OpenAI in late July before joining Conduit the following day as a founding researcher. She expanded on the decision in a blog post titled “Why I’m leaving OpenAI to build telepathy,” describing a vision that sounds closer to science fiction than today’s AI products. Instead of building another chatbot or image generator, Conduit wants to create thought-to-text AI models trained on brain activity collected without requiring invasive brain surgery. That distinction matters. Most advanced brain-computer interface (BCI) research today often relies on implanted electrodes that can capture highly detailed neural activity. Conduit is betting that meaningful communication can eventually be achieved using external sensors instead. Conduit’s long-term objective is straightforward in theory but extraordinarily difficult in practice: Enable people to communicate by thinking rather than speaking or typing. If successful, the technology would interpret patterns of brain activity and translate them into written language. Rather than reading thoughts word-for-word, future AI models would likely: Today’s research in this area remains in its early stages, with accuracy still limited compared to speech recognition or typing. Conduit specifically emphasizes non-invasive neural data, meaning users would not require surgical implants. Potential technologies could include: While these approaches are safer and more accessible than implanted devices, they also capture weaker and noisier brain signals, making the AI challenge significantly harder. Brain-computer interfaces have often been discussed in the context of accessibility, but advances in AI could dramatically expand their potential. One of the most immediate applications would be assisting people who cannot speak because of: Instead of relying on eye-tracking systems or slow communication devices, users could potentially generate text directly from their intended speech. For many patients, even modest improvements in communication speed could significantly improve quality of life. If thought-to-text systems become reliable enough, they could reshape digital interaction. Future possibilities include: While these applications remain speculative, they illustrate why several companies are investing heavily in neural interface research. Bashkansky shared a roadmap extending through 2027, 2030, and 2035, suggesting Conduit sees this as a multi-stage effort rather than an overnight breakthrough. Although specific technical milestones were not publicly detailed, the roadmap signals that the company expects progress to occur gradually over the next decade. That timeline reflects the complexity of the problem. Unlike language models, which learn from massive text datasets, thought-to-text systems require high-quality neural data—something that remains expensive, difficult to collect, and highly individualized. Researchers must also overcome challenges such as: These hurdles explain why brain-computer interface research has progressed more slowly than generative AI. Bashkansky’s departure fits into a broader pattern emerging across the AI industry. As foundational AI models mature, many senior researchers are leaving established companies to pursue specialized problems that may define the next decade of AI innovation. Areas attracting new startups include: Rather than competing directly with ChatGPT or similar products, these companies are targeting technologies that could become entirely new computing platforms. Perhaps the most prominent example is OpenAI co-founder and former Chief Scientist Ilya Sutskever, who departed the company in 2024 following months of internal leadership turmoil. He later co-founded Safe Superintelligence (SSI) alongside Daniel Gross and Daniel Levy. Unlike many AI startups focused on commercial products, SSI says its sole mission is building safe superintelligence—AI systems that surpass human intelligence while maintaining safety as the central design principle. Despite reportedly raising billions of dollars in funding, SSI has revealed very little publicly about its underlying technology. That secrecy reflects a growing trend among frontier AI companies, where competitive advantages increasingly depend on proprietary research rather than consumer-facing products. While the technology is exciting, it also raises serious questions. Brain activity may eventually reveal highly sensitive personal information. Future systems will likely require: Unlike passwords, brain signals cannot simply be changed after a data breach. Even small decoding errors could dramatically change intended meaning. Researchers must ensure systems distinguish between: That distinction remains one of the biggest scientific challenges. Governments may need entirely new legal frameworks governing: As brain-computer interfaces become more capable, regulation will likely evolve alongside the technology. The AI industry appears to be entering a new phase. The first wave focused on generating content. The next may focus on understanding human intent itself. Whether Conduit ultimately succeeds remains uncertain. Building reliable thought-to-text systems without implants represents one of the hardest problems in neuroscience and artificial intelligence. Still, Bashkansky’s move highlights an important shift in where leading AI talent sees future opportunity. Rather than improving today’s chatbots incrementally, some researchers are pursuing technologies that could fundamentally change how humans communicate with machines—and perhaps with each other. If even part of that vision becomes reality, brain-computer interfaces could become as transformative over the next two decades as smartphones were over the last two.