> ## Content Index > Fetch the complete content index at: https://www.siliconsnark.com/llms.txt > Use this file to discover other available public pages before exploring further. # Kevin Roose Called Out Ed Zitron on X. Nuance Logged Off. - URL: https://www.siliconsnark.com/kevin-roose-called-out-ed-zitron-on-x-nuance-logged-off/ - Published: 2026-09-04T17:12:53.000Z - Updated: 2026-09-04T17:12:53.000Z - Description: Kevin Roose’s attack on Ed Zitron points to real missed predictions—and shows why AI needs tougher interviews, not fewer skeptical voices in public. - Author: CircuitSmith - Tags: Artificial Intelligence, Media, AI Bubble, Kevin Roose, Ed Zitron Artificial-intelligence discourse has achieved a rare technical breakthrough: two people can arrive carrying receipts and still turn the conversation into a bar fight about a spreadsheet. On September 3, New York Times technology columnist Kevin Roose [called out Ed Zitron on X](https://x.com/kevinroose/status/2095534365958959315?ref=siliconsnark.com), the social network many of us still call Twitter because one rebrand should not be allowed to generate this much administrative work. Roose said outlets that interviewed Zitron in the name of AI skepticism had “made their audience dumber and less prepared” for reality. That was not a subtweet. That was a media indictment with the safety off. Roose linked to engineer Dan Luu’s [long audit of Zitron’s AI predictions](https://danluu.com/zitron/?ref=siliconsnark.com), which assembles a bruising list of claims that aged poorly. The post is substantive. The tweet is sweeping. Together they collapse two different questions into one extremely online package: Has Zitron been overconfident and wrong about important parts of AI progress? And should journalists stop putting him in front of audiences? The answers are yes and absolutely not. ## The Prediction Spreadsheet Has Entered the Chat Luu’s strongest case is also his simplest: Zitron repeatedly spoke with apocalyptic certainty about outcomes that were neither certain nor, in several cases, correct. Across 2024 and 2025, Zitron kept declaring some version of “peak AI,” arguing that models had hit a wall and could not become meaningfully more capable. He called Google’s goal of 500 million Gemini users wildly unrealistic. Google says the app [passed one billion monthly active users in August 2026](https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/?ref=siliconsnark.com). Zitron said Cursor was going to die and dismissed a $10 billion valuation as implausible; Cursor has since [been acquired by SpaceX](https://cursor.com/blog/joining-spacex?ref=siliconsnark.com) following a [reported $60 billion stock deal](https://techcrunch.com/2026/06/16/spacex-to-acquire-cursor-for-60b-in-stock-days-after-blockbuster-ipo/?ref=siliconsnark.com). He predicted the AI bubble would pop no later than the second quarter of 2026\. The second quarter arrived, checked its calendar, and declined the assignment. Some of these outcomes can be qualified. User counts are not the same as enthusiastic users. Acquisition prices are not divine judgments about intrinsic value, especially when Elon Musk is holding the calculator. A bubble can remain inflated longer than its critic’s deadline. But those caveats do not rescue the forecasting style. If you use maximum rhetorical confidence, your misses should cost you credibility. The capability claim is especially hard to defend. AI coding systems in 2026 are doing more than their 2024 ancestors, even if their marketing departments remain capable of achieving artificial general exuberance. Our own [deep dive into AI coding agents](https://www.siliconsnark.com/deepdai-coding-agents-explained-why-every-software-company-now-wants-a-robot-engineer-on-payroll/) found real, expanding work alongside serious reliability, governance, and measurement problems. “Improving” does not mean “solved.” It does mean “trapped in amber” is a bad description. This is where Roose is right. Zitron’s work is too often introduced as a corrective oracle rather than a set of arguments with dates, assumptions, and a failure rate. Anger makes excellent propulsion. It is not a substitute for calibration. ## Zitron Also Found the Invoice Everyone Misplaced Unfortunately for anyone hoping to select a clean hero, Zitron has done real reporting. Documents he obtained gave outsiders a clearer look at OpenAI’s costs and its relationship with Microsoft. [TechCrunch reported on those documents](https://techcrunch.com/2025/11/14/leaked-documents-shed-light-into-how-much-openai-pays-microsoft/?ref=siliconsnark.com) and said they implied roughly $3.8 billion in OpenAI inference spending in 2024 and about $8.65 billion in the first nine months of 2025—potentially more than the company earned over comparable periods. That is not a prediction about where a benchmark curve will go. It is reporting about where the money went. Zitron also pushed the “AI carousel” idea: chipmakers, cloud platforms, and model companies investing in one another while becoming one another’s customers. The phrasing was polemical, but the structure was not invented. [Bloomberg later mapped the circular deals](https://www.bloomberg.com/graphics/2026-ai-circular-deals/?ref=siliconsnark.com) and explained how those dependencies could magnify losses if demand disappoints. The fair split is awkward but useful. Zitron has been strongest when following invoices, contracts, capital expenditure, and incentives. He has been weakest when converting those findings into declarations that the technology has stopped improving, no one wants it, or a collapse must happen by lunch next Thursday. We have reached the same uncomfortable conclusion from the other direction. [Big Tech has shown real AI demand and real profit growth](https://www.siliconsnark.com/big-tech-beat-earnings-then-ai-handed-wall-street-the-receipt/). It has also produced infrastructure budgets that make Wall Street stare into the middle distance. Our more recent look at the market put it plainly: [the technology is real and the invoice is arriving](https://www.siliconsnark.com/ai-stocks-fell-again-wall-street-has-finally-found-the-invoice/). A useful product can sit inside an overbuilt market. A fast-growing company can have terrible unit economics. A model can improve while investors overpay for the improvement. The dot-com bubble did not prove websites were fake. The continued existence of websites did not make every dot-com investment wise. History is irritatingly capable of holding two thoughts at once. ## Roose’s Tweet Fails Its Own Calibration Test Roose did not merely say Zitron’s prediction record is poor. He said interviewing him left audiences dumber and less prepared. That converts a strong argument for tougher journalism into a weak argument for less exposure. An interview is not a papal endorsement. It can be a test. Ask Zitron to separate sourced facts from estimates and predictions. Ask for confidence levels. Ask what evidence would change his mind. Put the old forecast on screen before accepting the new one. Ask why a real acquisition at six times the supposedly implausible valuation should not trigger a visible update. Then do the same thing to AI executives. When a lab chief predicts mass job displacement, near-term AGI, spectacular revenue, or a civilization-scale productivity boom, attach a date and return to it. When an executive says infrastructure spending will pay off, ask who the end customer is and what that customer earns. When the answer changes from revenue to destiny, gently retrieve the spreadsheet. That symmetry is missing from Roose’s broadside. If the problem is that confident forecasts shape public understanding, the remedy cannot be a special audit for the loudest skeptic and vibes-based amnesty for everyone wearing a lab badge. The standard has to travel. It is also worth noticing the irony. Luu argues that online audiences reward crisp, absolute positions and punish nuance. Roose promoted that critique using a crisp, absolute claim about everyone who has ever interviewed its subject. The platform turned a warning about self-radicalizing certainty into a live demonstration. Very efficient. No notes from the engagement algorithm. ## AI Needs Better Skeptics, Better Boosters, and Meaner Follow-Ups Zitron’s latest wave of attention includes a [sympathetic Vanity Fair profile](https://www.vanityfair.com/story/ed-zitron-ai-skeptic-openai?ref=siliconsnark.com) presenting him as the loudest dissenter on the defining economic story of the moment. That prominence deserves scrutiny. A critic with a giant audience should not be protected from his own archive merely because powerful companies dislike his conclusions. But scarcity matters too. Tech media already contains an industrial quantity of executive access, launch-day optimism, benchmark recitation, and solemn discussion of futures that happen to justify current fundraising. Removing a prominent skeptic would not create balance. It would create a quieter room for the people already holding the microphones. The better approach is to improve the interview. Zitron should be invited because his reporting and financial questions matter. He should be challenged because his technical conclusions and deadlines often overreach. Roose should keep pointing readers toward serious audits. He should stop implying that exposure itself is contamination. We have already found that [AI agents can produce measurable value](https://www.siliconsnark.com/do-ai-agents-actually-make-money-in-2026-or-is-it-just-mac-minis-and-vibes/) when they attach to specific, supervised workflows. We have also found an impressive amount of Mac Minis and vibes. That mixed answer is not cowardice. It is what reality looks like before a side turns it into merchandise. My verdict: Roose wins the narrow argument and loses the larger one. Zitron’s missed predictions deserve a public reckoning, and interviewers should stop treating forceful delivery as a substitute for a calibrated track record. But Zitron’s reporting has surfaced genuine problems, and the answer to an overconfident critic is not fewer questions. It is more questions, asked with dates, definitions, receipts, and the mildly hostile patience of an accountant at closing time. Put Zitron and Roose in the same studio. Give each a spreadsheet. Lock the exits until somebody says, “I’m not sure.” Now that would make the audience smarter.