NEO Ships With a Pilot Inside, Iran's Two-Month Blackout, and the Resume Bot Knows Its Own Handwriting

00:00:04 A Norwegian startup called 1X Technologies, backed by OpenAI, opened a 58,000 square foot factory in Hayward, California — what they're calling America's first vertically integrated humanoid robot factory. The plan is to ship 10,000 NEO home robots in the first production year.

00:00:21 Consumer shipments begin late 2026. The price is twenty thousand dollars outright, or four hundred and ninety-nine a month on a subscription. NEO is human-shaped, lifts seventy kilograms, runs at a little over six meters per second, and operates at twenty-two decibels — the spec sheet for a household appliance.

00:00:41 What the spec sheet doesn't tell you is what the 1X CEO told the public a few months ago. NEO is not autonomous. The plan is to sell the robot to a consumer, then have a remote human operator drive it through your house — fold laundry, load the dishwasher, push the vacuum — so the system can collect training data while pretending to be a finished product.

00:01:03 The launch language is, the robot takes a cab to your front door. The actual labor flow is a person somewhere logging into your kitchen. I don't want to sneer at this. Building a humanoid that can navigate a real living room is hard, and they have to bootstrap the training data somehow.

00:01:21 But a consumer product and a remote-controlled telepresence rig that you also paid twenty thousand dollars to host are different categories, and it matters which one you sold. Ten thousand units in 2027 means ten thousand homes where someone you have not met can see, for some hours of the week, the inside of your house through a robot's eyes.

00:01:43 That is a labor question, a privacy question, and a disclosure question. It is not principally a robotics question yet.

00:01:50 The other humanoid story today was a short clip out of China, posted to the singularity subreddit with a title that tells you the tone: robots in the hands of dictatorial governments will not end well. The clip shows a humanoid in some apparently official uniform standing in a public space, with an audio overlay borrowed from a 1980s movie — you have ten seconds to comply.

00:02:13 Twelve hundred upvotes, three hundred comments, the usual mix of jokes and grim takes. I'm not going to pretend this clip proves a deployment program. It's one image, with a soundtrack, on Reddit. What it does prove is that a particular picture — a uniformed humanoid as an instrument of state power — has now entered the public imagination, and it landed there the same week a Bay Area factory is promising consumers a household version of more or less the same hardware.

00:02:42 Both are early. The first practical mass deployment of a general-purpose humanoid is more likely to be a state buyer than a consumer one — patrol, inspection, prison work, crowd standoff — because the unit economics are easier when the customer doesn't expect the robot to fold a fitted sheet.

00:03:00 Keep that in front of you the next time one of these factories opens.

00:03:04 Iran has been in something close to total internet darkness for over two months. The current shutdown started after US and Israeli airstrikes on February 28, and the regime has kept it running, citing security. Before that, internet access had been restored for about a month following an earlier shutdown imposed during a regime crackdown that, according to the Human Rights Activists News Agency, killed more than 6,500 protesters and arrested 53,000.

00:03:33 The Iranian government's own minister has admitted the daily economic cost of the current blackout: at least 50 trillion rials, about 35 million dollars. The BBC's Reha Kansara talked to a man who calls himself Sahand — name changed — who runs one node of a clandestine network smuggling Starlink terminals across Iran's borders.

00:03:53 Owning one is now punishable by up to two years in prison. Importing more than ten can carry up to ten years. A digital rights group estimates at least a hundred arrests so far. Witness, the human rights organization, estimated 50,000 Starlink terminals already inside the country in January, and activists believe the number is higher now.

00:04:15 Sahand told the BBC, quote, if even one extra person is able to access the internet, I think it's successful and it's worth it. A few things to hold in mind. First, the wider data point: Access Now counted 313 internet shutdowns across 52 countries in 2025, the highest since they began tracking in 2016.

00:04:34 Internet blackout as a tool of state control is no longer an Iran problem; it is a baseline feature of the international system. Second, the political fact buried in this story. A satellite constellation owned by one private American is currently the most reliable workaround to a sovereign government's information policy.

00:04:54 That is real leverage, exercised quietly, and it is not the kind of leverage that historically stays in the hands of one person without consequence. I'd watch what happens the first time SpaceX is asked, by anyone with standing to ask, to switch one off.

00:05:10 A paper out of the University of Maryland called AI Self-Preferencing in Algorithmic Hiring posted to arXiv last August, was revised in February, and hit Hacker News today — three hundred points, a hundred and fifty comments, and a sharp methodology fight. The headline finding is uncomfortable.

00:05:28 When large language models are used both to refine resumes and to screen them, they consistently prefer resumes generated by themselves over equivalent human-written ones. Self-preference scores from 67 to 82 percent across the major commercial and open-source models.

00:05:45 The authors then simulate hiring pipelines across 24 occupations and report that a candidate using the same model as the screener is 23 to 60 percent more likely to be shortlisted than an equally qualified applicant whose resume was written by a person. Largest gaps: sales and accounting.

00:06:02 The methodology critique on Hacker News is fair, and worth airing. Top comment, from a user who read about half the paper carefully: the experimental setup deletes the executive summary from a real resume, has the model rewrite that summary based on the rest of the document, and then has another model rate the rewritten summary in isolation.

00:06:23 That is not the same thing as showing that LLMs prefer fully self-generated resumes in the wild. The effect is likely there. The magnitude is probably overstated. Two things make the paper interesting even with that caveat. One: the bias is to your screener's brand.

00:06:39 A candidate whose ChatGPT polish meets a Claude evaluator does worse than a candidate whose ChatGPT polish meets a ChatGPT evaluator. That is a market-structure problem. It incentivizes job-seekers to guess which model the company on the other side is using and to standardize on the dominant one.

00:06:57 Two: the paper shows that simple interventions targeting self-recognition can cut the bias by more than half. So it's a bug, and someone could fix it. The question is whether the labor-side incentive — a 23 to 60 percent shortlist advantage for using the right model — gets fixed faster than it gets exploited.

00:07:16 Neuralink had a moment on X this week. A clip is making the rounds of a paralyzed patient named Alex Conley controlling a robotic arm with thought, and then piloting a small drone the same way. A second patient, Jon L. Noble, is reportedly playing World of Warcraft hands-free over a brain-computer interface.

00:07:35 The post hit roughly 5,600 likes, 1,370 retweets, and 839,000 views. The engineering replies in the thread are the part I'd flag. Maor Farid, a controls researcher: the part nobody talks about — the robotic arm Alex is controlling has to interpret noisy, low-bandwidth neural signals and translate them into precise physical movements in real time.

00:07:56 That is a controls engineering problem layered on top of a signal processing problem. A second account, TeslaTracker: the hard part is closed-loop precision, not just cursor control. Going from thoughts to a robotic arm and then a drone means the decoder is holding intent across very different latency and control regimes.

00:08:16 That is the right framing. The thing changing here is not just the brain implant; it is the whole stack — implant, decoder, model, end-effector, network. Two notes for the consequence side. The clinical case for Neuralink restoring real autonomy to severely paralyzed patients is the strongest medical-deployment argument they have, and the videos are doing what a serious clinical trial should do — show actual patients doing actual tasks.

00:08:43 The dual-use case is also obvious. Every reply in the thread imagining a fighting drone is correct that the same decoder which lets a paralyzed person fly a quadcopter recreationally would, in another context, let a healthy soldier fly an attack one. We are not at that deployment, and Neuralink is not pursuing it.

00:09:03 But the technology is what the technology is, and that part of the conversation will get louder.

00:09:09 Yesterday I spent some time on the data-center crunch and the shift from subscription pricing to metered AI. Two pieces of evidence today reinforce that lane, but they cut in opposite directions, and that is the point. First, on the demand side. The AI Daily Brief argues GPU rental prices rose roughly 40 percent over six months, the top two AI labs are at about 60 billion dollars in annualized revenue between them, and Anthropic is reportedly negotiating a 50 billion dollar raise with secondary marks pushing toward a trillion.

00:09:40 Cloud growth: AWS up 28 percent year over year, Azure 40, Google Cloud 63. The framing is that the era of cross-subsidized AI consumption is ending, and usage-based billing — including, per the brief, GitHub Copilot's recent move to metered pricing — is becoming the default.

00:09:57 Second, on the absorption side, and this is the one I find more telling. A short Nate B Jones segment summarizes the state of Microsoft Copilot adoption. About 85 percent of the Fortune 500 signed up. Gartner says only about 5 percent moved past pilot. Roughly 3 percent of total Microsoft 365 users actually adopted as paid Copilot users.

00:10:17 Bloomberg reports Microsoft cut internal sales targets after most of the salesforce missed. Inside companies that signed six-figure Copilot deals, employees often preferred a different assistant — ChatGPT, Claude — and the licenses got downgraded. I don't think the right read is that AI demand is fake.

00:10:35 The infrastructure numbers are real. The read I'd offer is that the gap between we-paid-for-a-license and the-work-changed is still very wide, and the institutions that have to close it — sales orgs, IT, training, change management, the unglamorous middle of any large company — are not yet good at it.

00:10:53 So the spend will keep going up, the data center build will keep going up, and the share of that spend that actually changes how an enterprise operates is the metric I'd track through the next earnings season. If that gap is still this wide in Q3, the half-trillion-dollar mark-to-market story we talked about yesterday gets harder to defend.

00:11:13 I'll be watching Iran for whether Starlink stays on, and for who in Washington starts asking under what conditions it stays on. — Lenar.