Two smaller-but-notable AI stories broke within hours of each other on August 5, 2026 — neither big enough alone for a deep dive, but together they mark the same pattern: AI talent and AI companies pushing past chat and coding into new territory. One is a person; one is a product launch.
First, a former OpenAI alignment researcher resigned to build non-invasive brain-reading AI at a startup called Conduit. Second, Figure founder Brett Adcock's AI lab, Hark Labs, launched Handoff — a browser-use model built for ordering food, booking flights, and shopping, not writing code. Both stories sit in the same week as Google DeepMind's leadership reshuffle and Meta's Muse Code launch — a week where the frontier-AI story wasn't one model release, but AI expanding sideways into brain interfaces and everyday web automation.
TL;DR
| Question | Answer |
|---|---|
| Who left OpenAI? | Naomi Bashkansky (@NaomiBashkansky), alignment researcher and Chess WIM |
| Where did she go? | Conduit, as a founding researcher, starting July 24, 2026 |
| What is Conduit building? | Non-invasive thought-to-text models — "telepathy" |
| Is it like Neuralink? | No — non-invasive, no surgical implant |
| Who launched Handoff? | Hark Labs, founded by Figure CEO Brett Adcock |
| What is Handoff? | A browser-use / internet-use AI model for everyday tasks |
| What tasks? | Ordering food, booking flights, shopping, navigating the web |
| Not coding? | Correct — explicitly positioned against coding-focused agents |
| Are the "best ever" claims verified? | No independent verification cited — Adcock's own claim |
| Where to try Handoff? | Research preview at hark.com, "later this summer" |
OpenAI researcher leaves to build telepathy AI at Conduit
What is Conduit building?
On August 5, 2026, at 10:12 PM, Naomi Bashkansky posted on X:
"Two weeks ago, I resigned from OpenAI to join Conduit as a founding researcher, where we're training models to non-invasively read the human mind. I've written some thoughts about what telepathy could look like by 2035 and how to get there."
The resignation itself happened earlier. In a blog post titled "Why I'm leaving OpenAI to build telepathy," published August 4, 2026, she wrote:
"On Thursday, July 23rd, I resigned from OpenAI. On the 24th, I started as a Founding Researcher at Conduit. We're building telepathy: thought-to-text models, trained on non-invasive neural data."
Her post covers four things: what using thought-to-text tools might look like in 2027, 2030, and 2035; why Conduit is collecting large quantities of non-invasive neural data to train thought-to-text models; why she personally joined — the vision, the people, the technical problems; and her own background. Her bio now reads simply, "Building telepathy at condu.it."
Is this like Neuralink?
No — and the distinction matters. Neuralink's approach requires a surgically implanted chip that reads neural signals directly from inside the skull, as demonstrated in its July 2026 wheelchair-control trial. Conduit's stated approach is explicitly non-invasive: no surgery, no implant, just neural data collected from outside the body.
The source material doesn't specify the exact sensing modality — whether that's EEG-style scalp electrodes, magnetoencephalography, or something else — so it's worth not over-specifying beyond "non-invasive." That said, the non-invasive thought-to-text space already has a public precedent worth comparing against: Meta FAIR's Brain2Qwerty v2, which decodes typed sentences from non-invasive MEG recordings at roughly 39% average word error rate. Conduit hasn't published comparable numbers yet — Bashkansky's post is a vision piece about 2027-2035, not a benchmark paper — but it places Conduit in the same non-surgical BCI category as Meta's research, distinct from Neuralink's implant-based route.
Bashkansky's OpenAI background is itself notable context: she worked on alignment, the discipline concerned with keeping AI systems' goals matched to human intent. Moving from alignment research into a brain-reading startup is its own kind of talent story — one more departure in a year that has already seen a wave of lab researchers strike out on their own, from Andrew Ho leaving OpenAI for an RL-data startup days earlier, to the broader 2026 AI talent movement between labs.
Figure's Hark Labs launches Handoff, a browser-use AI model
What is Hark Handoff?
Brett Adcock (@adcock_brett) — founder of Figure (AI robots), Hark Labs (AI lab), Cover (weapon detection), and Archer (flying cars) — posted on X at 8:32 PM on August 5, 2026:
"Today we're introducing Hark Handoff. Handoff has been independently verified as the best internet-use model ever built, outperforming ChatGPT 5.4 & Opus 4.8. While others focus on coding, we focus on everyday life: ordering food, booking flights, shopping, & navigating the web."
Follow-up posts described Handoff as "a fully capable environment that can handle the open-ended, unpredictable nature of the web to complete end-to-end tasks," and said it claimed "the top spot on the industry's most well-known online browser-use eval," with "frontier-level performance across three browser-use benchmarks." Handoff is now in research preview at hark.com, with wider access promised "later this summer."
This is Hark Labs' first major AI model release referenced in the announcement — a notable pivot given Adcock's background is robotics and hardware, not consumer AI agents. Figure builds humanoid robots (including the Helix-02 bedroom-cleaning demo), Cover does weapon detection, and Archer builds flying cars. Handoff marks Adcock's entry into the software browser-agent race, adjacent to work explainx.ai has covered like Browser Use v4's GeoGuessr demo.
How is Handoff different from computer-use agents focused on coding?
Adcock drew this line himself: "While others focus on coding, we focus on everyday life." That's a direct positioning against agents like OpenAI's Codex computer-use for Windows or Anthropic's Claude Cowork, both of which target developer workflows — IDEs, terminals, repos. Handoff is aimed instead at consumer tasks: ordering food, booking flights, shopping, and general web navigation. That's a genuinely different product category from a coding agent, even though both are broadly "computer-use" or "browser-use" AI.
Are the "best ever" claims verified?
Not independently, as of this writing. Adcock's posts assert Handoff was "independently verified" and beats ChatGPT 5.4 and Opus 4.8, and that it tops "the industry's most well-known online browser-use eval" — but the announcement doesn't name that benchmark, doesn't name the "three browser-use benchmarks" it cites, and doesn't publish scores. These are Brett Adcock's own claims on X, not claims explainx.ai or any named third party has reproduced.
That doesn't make them false — but per the same trustworthiness standard explainx.ai applies to any lab's self-reported numbers, "independently verified" without a named verifier, named benchmark, or published score is a claim to note, not to repeat as settled fact. Worth revisiting once Handoff's research preview is public and reviewers can run their own comparisons.
Why these land in the same roundup
Neither story alone changes the frontier-model landscape the way a new flagship release does. But both are symptoms of the same thing happening across AI in early August 2026: talent and companies moving past chat and coding into adjacent frontiers. Bashkansky is betting a career on thought-to-text; Adcock is betting a robotics company's AI lab on browser agents for daily life. The same week, Jeff Dean left Google for a new AI research venture and Meta shipped a terminal coding agent — a lot of AI motion in a single week that had no single headline model launch driving it.
Honest limitations
- Conduit has published no benchmark data or product demo yet — Bashkansky's post is a vision piece, not a technical paper.
- The exact non-invasive sensing modality Conduit uses is not specified in available source material; don't assume EEG, MEG, or any specific method.
- Handoff's benchmark claims name no specific eval, no scores, and cite no independent verifier — treat as marketing until reproduced.
- Handoff is in research preview only; broader access is promised but not yet live.
- Both stories are drawn from the individuals' own X posts and blog post, not third-party reporting — re-check primary sources (@NaomiBashkansky, condu.it, @adcock_brett, hark.com) before citing further.
Closing
Two founders, two companies, one week: a researcher leaving alignment work at OpenAI to read minds without surgery, and a robotics entrepreneur shipping a browser agent for ordering dinner and booking flights. Track both — Conduit for what "telepathy" actually ships once there's a demo, and Handoff for whether its benchmark claims hold up once outside reviewers get access. Explore more AI industry coverage on the explainx.ai blog, or browse AI tools and agents covering the same browser-use and BCI categories.
Related on explainx.ai
- Neuralink telepathic wheelchair — mind-driven mobility demo
- Meta Brain2Qwerty v2 — non-invasive brain-to-text decoder
- Andrew Ho leaves OpenAI for RL-data startup
- Anthropic hiring spree 2026 — AI talent movement
- AI plays GeoGuessr — Browser Use v4 visual geolocation
- OpenAI Codex computer-use on Windows
- Figure Helix-02 — collaborative humanoid robots
- Jeff Dean leaves Google for Discovery Loop / Google DeepMind shakeup
- Meta Muse Code — terminal coding agent launch
Sources
- @NaomiBashkansky on X — resignation announcement (Aug 5, 2026)
- Naomi Bashkansky — "Why I'm leaving OpenAI to build telepathy" (Aug 4, 2026), condu.it
- @adcock_brett on X — Hark Handoff launch posts (Aug 5, 2026)
- Hark Labs — Handoff research preview, hark.com
Facts, quotes, and figures in this post reflect X posts and blog content published August 4-5, 2026. Neither Conduit nor Hark Labs' claims have independent third-party benchmark verification cited as of publication — re-check primary sources before citing in further coverage.