Why did Naomi Bashkansky leave OpenAI to build AI that can read your thoughts — and could brain-computer inter

🧠 Why Naomi Left: To Build Thought-to-Text AI

She resigned from OpenAI to join Conduit and help build “thought-to-text” — non-invasive brain-computer interfaces (BCIs) that translate neural signals into text — because she believes large-scale non-invasive neural datasets and new models can make telepathic interfaces practical within years, not decades [naomibashkansky][techtimes].

🔎 Things to Know

• 🧾 She left OpenAI to be a Founding Researcher at Conduit, focusing on thought-to-text models trained on non-invasive neural data [naomibashkansky].
• 🧠 Product goal: Conduit is developing a neural headband and models to map brain signals to text/commands, aiming for consumer-style telepathy experiences in coming years [gizmodo][techtimes].
• 💰 Recruiting data: The startup is collecting large quantities of neural recordings (including paid participants) to scale training data—Conduit reportedly pays participants for sessions [gizmodo][techtimes].
• 🔬 Approach vs. invasive BCIs: Conduit emphasizes non-invasive sensors (e.g., headbands) rather than implants, trading signal quality for safety and scalability [naomibashkansky][gizmodo].

🧭 Why It Matters

• ⚙️ New interface to AI: If successful, BCIs could let people control AI and compose text via thought, changing accessibility and interaction models for assistants and devices [naomibashkansky][techtimes].
• 🔐 Risks & ethics: Thought-reading raises privacy, consent, and security concerns—especially around what neural data is captured and how it’s used [gizmodo][naomibashkansky].
• ⏳ Timeline caveats: Optimistic timelines exist, but performance, generalization, and safety challenges mean widespread, reliable telepathy is still uncertain [naomibashkansky][techtimes].

👉 tl;dr: Naomi left OpenAI to help build non‑invasive thought-to-text BCIs at Conduit, betting big data + models can make practical “telepathy” interfaces but meaningful privacy, safety, and technical hurdles remain [naomibashkansky][gizmodo].

Follow-up Questions:

1. How do non-invasive BCIs compare technically to implanted systems right now?

2. What kinds of neural data is Conduit collecting and how is it labeled?

3. What privacy and consent safeguards are being proposed for thought-to-text data?

4. Which applications are most likely to arrive first (accessibility, typing, control)?

5. What performance metrics would show a consumer-ready thought-to-text system?