Summary

1X is pitching NEO as both a household appliance and a data-collection fleet for making household robotics work. The robot costs $499 a month with a six-month commitment or $20,000 outright; Bernt Bornich says preorders exceeded expectations “by quite a margin,” with plans to ship more than 10,000 units in 2026 and hopes for roughly 80,000 the following year. That puts a concrete deployment plan behind a category that has absorbed $5 billion in VC funding since 2024.
Today’s NEO is a teleoperated service with islands of autonomy, not yet the autonomous butler consumers may imagine. Bornich says doors work “90-something percent” of the time and object retrieval about 80%, while tidying, vacuuming, and laundry use a mixture of autonomy and human help. When pressed for the autonomous share, his answer captured the ambiguity: “All of it is being done autonomously, but with a human in the loop.”
The core technical thesis is that household deployment can create an embodied-data advantage the public internet does not provide. Bornich argues that 10,000 robots could approach the volume of video uploaded to YouTube each day, while home activity better represents intelligence than content optimized for likes. NEO gathers video, audio, touch, and force data: “They need to live and learn among us.”
Privacy is not peripheral to adoption because the training loop can give remote workers a view of customers’ physical surroundings. 1X says collection occurs only during active tasks, data is stored locally before cloud transfer, people are blurred, operators do not know the home, and VR intervention is narrowly delegated; one manager currently monitors eight teleoperators. Bornich’s answer is “radical transparency,” but the hosts emphasize that children, partners, and visitors may also be captured, so the whole household effectively has to be in on it.
Bornich’s roadmap is aggressive, and its biggest business risk is teleoperation economics. His bet is that 10% of American homes could have humanoid robots by 2030; cleaner-level autonomy could arrive in a bullish 2027 or a 2028–2029 bear case, while expert-carpenter quality is closer to 2030. He concedes 1X is “running toward a cliff”: autonomy must improve as deployments scale, or the company will have to slow shipments.
The live demonstration validated the hardware experience while exposing the autonomy gap. The roughly 5-foot-6, 66-pound robot fetched a drink, filled water, and cleared clutter smoothly, but everything shown was teleoperated; it dropped tongs, could not retrieve them, nearly fell while squatting, and could not sit. Casey Newton’s consumer verdict was “buyer beware”: an owner expecting a butler might instead receive “an intern that needs a lot of your attention.”
The hosts nevertheless see a plausible mass-market product if 1X can solve a small number of recurring chores reliably. Newton’s “Holy Trinity” is laundry, the dishwasher, and trash: if NEO handled those quickly and permanently, he would pay $500 a month. Kevin Roose compared the demo with early constrained self-driving cars—“this thing is gonna get good”—while preserving uncertainty over whether that takes 3, 5, 10, or 20 years, and whether 1X is the winner.
HatGPT’s stories converge on who controls the data, labor, and authenticity behind AI systems. Common Crawl reportedly resisted publisher removal requests; xAI employees were allegedly compelled to license faces and voices for Project Skippy; Meta blamed disputed adult-film downloads on personal use; and a 52-page memo complicated the accepted story of Sam Altman’s 2023 firing. Coca-Cola’s one-month AI ad illustrated the other trade: production accelerated from a year, but the hosts thought the result looked “floaty and weird.”

Deep dive

1. 1X is launching NEO before household autonomy is finished

Bornich declined to disclose preorder numbers but said demand landed above expectations “by quite a margin.” The polarized response produced both “haters” and a “huge supporter base” of early adopters willing to develop the product alongside 1X.
NEO is scheduled to become available in 2026 at $499 a month with a minimum six-month commitment, or $20,000 outright. Roose and Newton place it inside a heavily financed category: VCs have invested $5 billion since 2024, while Tesla, Figure, and Chinese manufacturers are pursuing humanoids.
Bornich describes his own NEO as “pretty freaking good at tidying” and better than his Roomba because it moves chairs and vacuums the couch. It performs some cleaning and laundry, but folded clothes remain imperfect and he still has to clean himself.

2. Autonomy is a spectrum—and the hosts force a clearer definition

Physical chores currently combine autonomous actions with remote assistance. Bornich compares the arrangement with Waymo interventions and iterative ChatGPT prompting: the human does not necessarily control every movement, but helps when the system reaches something it cannot handle, producing training data for the next attempt.
Newton’s direct question—what percentage is autonomous, 5%, 20%, or 50%?—elicited a slippery but revealing answer: “All of it is being done autonomously, but with a human in the loop.” Once Newton removed the teleoperator from the hypothetical, the genuinely independent capability set became much narrower.
In fully autonomous mode, Bornich says NEO can converse as well as leading language models, detect who is addressing it, and use body language to enrich the interaction. He says this has replaced his normal use of language models on a laptop or phone: “I just talk to Neo.”
Physical autonomy includes navigation, opening doors with “90-something percent” success, moving objects, and fetching requested items with roughly 80% success. Scheduled chores retain teleoperation so the robot can be “very, very useful day one” while gathering demonstrations for tasks it cannot yet complete.

3. Household data is the real scaling thesis

Newton identifies the foundational constraint: text models had billions of internet words, while no comparable corpus exists for folding laundry. Bornich agrees and invokes self-driving cars, but argues household robots face a far larger variety of objects, rooms, tasks, and human behavior.
1X plans to ship more than 10,000 units in 2026—more robots, Bornich notes, than Waymo has cars—and hopes for about 80,000 the following year. He claims 10,000 robots begin to approach YouTube’s daily uploaded-data volume, while 80,000 would be way more than publicly available internet video.
Bornich’s deeper claim is that the internet is “such a bad representation of human life and intelligence”: only a tiny fraction of life gets posted, usually “whatever gets you likes on YouTube.” Home deployment instead captures the mundane behavior required to become a “well-functioning human.”

4. Trust depends on constrained human access, not zero human access

During active work, NEO collects video, audio, touch, and forces acting on its body; Bornich says it does not collect while merely charging. Data is first stored locally, allowing deletion, then sent to a secure cloud store where he says essentially no one sees it except the model accumulating knowledge from it.
Newton makes the surveillance concern deliberately concrete, asking whether NEO would recognize an intimate moment and leave. Roose adds less comic cases: people change clothes, children live in homes, and families may not want an unseen teleoperator observing their private routines.
Bornich says the supervisory interface resembles StarCraft: workers queue instructions across multiple robots while people are blurred and home identities hidden. If autonomy repeatedly fails—for example, on a leather jacket—the task passes to a VR operator who sees the object and instruction, completes the manipulation, then moves on.
1X limits a given group to roughly four people exposed to a set of homes, uses one manager for every eight teleoperators, vets US operators, and keeps video logs for accountability. Bornich compares this with admitting a cleaner and argues 1X can be safer; his broader defense is “radical transparency,” including telling uncomfortable customers to wait beyond the early-adopter phase.

5. A resident companion, not a visiting cleaner, is the product bet

Roose asks why 1X does not operate an Uber-like fleet that cleans a home and leaves. Bornich’s answer is that much of the value comes from five-minute or two-minute interventions scattered through the day—an operating pattern conventional cleaning services cannot provide.
Assisted living is his strongest example: someone who dropped an object or wants a drink may need brief help repeatedly, not a scheduled cleaning visit. The same logic extends to asking NEO for tea during a board game; Bornich says reducing that vision to cost-effectiveness misses the product’s purpose.
Emotionally, he places NEO “somewhere in between” an appliance and a pet—“my Hobbes from Calvin and Hobbes”—without calling it a replacement for people. He argues embodied conversation could pull users away from phones and make them more present; he rejects romantic use but cites an “infinitely patient machine” as potentially empowering for autistic children.
Newton preserves the harder companion question: software companions have been associated with delusion, alienation, and some suicides. Bornich says population statistics make causation difficult to establish, but still hopes to reduce harm; he also concedes that social-context judgments—whether to mention a private rash when a friend enters—remain unsolved. “AI models are snitchers,” he adds, because honesty-focused alignment can make them disclose too much.

6. The roadmap is bullish, but teleoperation creates a scaling cliff

Bornich’s bet is that 10% of American homes could contain a humanoid robot by 2030. Once reliability kinks are solved, he expects demand to resemble cars or phones; the limiting problem becomes manufacturing enough units without exhausting the early-adopter pool first.
Asked when no 1X employee would need to take over any household task, he reframes autonomy around whether a human cleaner counts despite occasionally needing instructions. On that standard, he gives a “bullish 2027” target, a 2028–2029 bear case, and says he is “pretty confident” about 2027.
The quality distinction matters more than a binary autonomy label. “Most AI today is slop,” Bornich says, yet mid-quality physical labor can still be useful for menial chores; a robot capable of expert carpentry throughout the home is a different threshold, probably closer to 2030.
In 2026, heavy usage is valuable because the alternative is expensive mock kitchens, constantly rearranged environments, and dedicated data farms. Customers receive labor while 1X receives demonstrations—but Bornich acknowledges the company is “running toward a cliff”: if autonomy does not catch up as deployments scale, 1X must pause or slow deployment.

7. The live demo looked real, smooth—and entirely human-driven

In person, NEO stood roughly 5 feet 6 inches and weighed about 66 pounds. Large eyes, no mouth, and a woven suit concealing metal hardware made it approachable; Roose called it a “short king” and reported that its hug was unexpectedly warm.
The robot fetched a drink from the refrigerator, filled a cup with water, and moved table clutter into the trash “pretty well and pretty quickly.” Its first discarded object, a pair of tongs, fell to the floor—and NEO could not retrieve it.
When asked to pick up blocks, NEO attempted a squat, lost balance, and nearly fell backward before a person steadied it. 1X later attributed the failure partly to Wi-Fi and improper calibration; the same limitations prevented a promised chair-sitting demonstration.
Everything the hosts saw was teleoperated through a VR headset by a 1X employee named Eric. Autonomous conversation was also unavailable because of Wi-Fi issues. The result was smoother than Roose expected; a Google robot Casey believed was acting autonomously had taken longer to do everything, though Newton stresses: “We did not see the robot do a single thing autonomously.”

8. The hosts see a future platform inside a buyer-beware product

Standing beside a humanoid produced an effect videos had not captured for Roose: a machine turned toward him, offered its hand, and made science fiction feel physically present. It never felt dangerous, though Newton found its alternating human-like and “herky-jerky” movement unsettling.
Roose experienced a second uncanny valley: giving orders felt uncomfortable because the robot was effectively “just a guy right now.” He could have picked up the clutter himself; routing the act through Eric and a 66-pound machine felt like “a very inefficient way of picking up and putting down things.”
Newton would not buy one soon. A bleeding-edge customer might enjoy the experiment, but someone expecting a butler could discover “an intern that needs a lot of your attention” and end up working for the robot. Roose similarly expects extensive hand-holding and says household members must jointly accept the data collection.
Their longer-term view remains open. Roose recalls safety drivers and parking-lot routes in early self-driving cars and repeats the same intuition: “This thing is gonna get good,” though perhaps in 3, 5, 10, or 20 years—and perhaps not through 1X. Newton’s purchase threshold is laundry, dishwasher, and trash handled reliably; solve that “Holy Trinity,” and $500 a month becomes credible.

9. HatGPT finds rights conflicts hiding inside training pipelines

An Atlantic report said Common Crawl, whose corpus helped train GPT-3 and other models, scraped much of the internet, including publisher material, and did not reliably honor removal requests. Its executive director’s position—“The robots are people too” and should read books free—struck the hosts as indefensible despite legitimate reasons to preserve the disappearing web.
A Wall Street Journal report said Elon Musk oversaw the racy companion Annie and that xAI employees were compelled to provide faces and voices for confidential Project Skippy. The requested license was perpetual, worldwide, non-exclusive, sublicensable, and royalty-free; Newton’s formulation captures the labor problem: being selected as the face of a sex bot, “and this is not optional.”
Strike 3 Holdings alleged that adult films were downloaded through Meta corporate addresses and a concealed network, seeking damages estimated above $350 million. Meta said employees, guests, or contractors downloaded them personally—not for training—and the hosts note Strike 3 has been accused of being a copyright troll, even as they mock Meta’s argument that the volume was too small to resemble its training acquisitions.

10. Depositions and pardons expose how power actually operates

President Trump pardoned Binance founder Changpeng Zhao after his 2023 guilty plea over money-laundering violations; the company had also struck a May deal involving the Trump family’s crypto venture. Trump said he did not know CZ. Roose believes someone may simply have handed him a list, while still calling the surrounding picture seemingly corrupt and expecting more facts to emerge.
The hosts treat presidential clemency as institutionalized magic: “You go free,” and a person becomes free. Newton compares the process to Oprah telling viewers to look under their seats for pardons—an absurdity sharpened by the possibility that the president did not read the underlying cases.
The White House’s own institutional voice came under scrutiny through whitehouse.gov/mysafespace, a mock MySpace profile attacking Hakeem Jeffries and Democrats with references to DEI, immigration, George Soros, Antifa, and Chucky. Newton calls it the work of a mean 13-year-old; Roose doubts nostalgia aimed at “41-year-old political staffers” changes votes.
Ilya Sutskever’s 10-hour deposition in Elon Musk’s OpenAI lawsuit supplied a different kind of accountability. Newton says Sutskever presented the board with a 52-page dossier of alleged office misconduct, reframing Sam Altman’s 2023 firing from an effective-altruist panic into a C-suite rebellion involving Sutskever and Mira Murati. Roose values the suit chiefly for surfacing information that otherwise stayed hidden; Murati’s deposition may disclose more.

11. Synthetic media compresses production faster than it earns trust

Coca-Cola’s marketing chief said its standard holiday-ad process had fallen from roughly a year to about a month using AI. The hosts’ review was blunt: animals wobbled without visual integrity, everything felt “floaty and weird,” and the clearest demonstrated benefit was probably lower cost.
Newton sees legitimate uses for voice cloning, including an audio edition of his text newsletter or replacing one or two incorrect words in a podcast. He rejects cloned ad reads because advertisers are buying the host’s personal contribution; for limited corrections, his operating rule is disclosure.
His ElevenLabs experiment revealed that voices carry context, not merely identity. Training on high-energy Hard Fork audio made his somber newsletter sound as though a “crazy person” were reading it, rendering the clone unusable—a small example of why synthetic delivery can preserve the words while breaking the intended performance.

Amjad Masad & Adam D’Angelo: How Far Are We From AGI?Alternative Investing: Alts For All - [Business Breakdowns, EP.234]