"AI Is Really Weird" is a long-form critical essay by Edward Zitron, an independent technology journalist and analyst, published April 8, 2026 on his newsletter Where's Your Ed At (wheresyoured.at). It advances a skeptical account of what Zitron calls the "AI bubble," structured around four arguments: that AI agents are not what they are claimed to be; that coding language models create more problems than they solve; that the economics of AI labs are unsound; and that Anthropic's reported revenue growth is misleading. The piece draws on reporting from The Information and Newcomer, official Anthropic announcements, and academic papers.
Summary of argument
Zitron argues that more than three years into the language-model hype cycle, with hundreds of billions of dollars invested, the industry cannot clearly articulate what AI does that justifies the spending. He treats "AI agent" as a marketing relabeling of chatbots, contends that AI-generated code adds volume without quality, holds that the unit economics of the major labs do not work, and reads Anthropic's escalating revenue figures as artifacts of accounting choices and product launches rather than underlying demand. He frames the recurring "early days" defense as a deflection from questions of efficacy and economics.
Key claims
Agents as chatbots
Zitron describes every "AI agent" as "a chatbot talking to another chatbot connected to an API and a system of record," and characterizes the word "agent" as marketing intended to make chatbots sound powerful, writing that "it's always a fucking chatbot." He argues that 2025, billed as the "year of agents," became the year of talking about agents. He cites a Goldman Sachs and Anthropic partnership that after six months of work remained, in his account, "in early stages" with no disclosed tasks beyond vague references to "accounting" and "onboarding." Products including OpenClaw, ChatGPT Agent, Claude Cowork, and Perplexity Computer, he writes, all reduce to chatbots with API integrations. Citing multiple papers, he states that agents are "incapable of carrying out computational and agentic tasks beyond a certain complexity."
Coding models: quantity without quality
Language models can write a large volume of code but not necessarily good code, Zitron argues, with no guarantee that output will compile, run, or be secure, performant, or maintainable. He cites one firm that went from 25,000 to 250,000 lines of code per month after adopting Cursor, creating a backlog of roughly 1 million lines awaiting review, and argues that AI-generated code produces security vulnerabilities faster than they can be reviewed. He contends that hyperscalers such as Meta and Amazon allow non-technical employees to ship AI-generated code, worsening quality, and points to a Register study reporting that AI coding tools "make them slower" for experienced engineers. He poses the question: "Are you shipping faster? Is the code better? What is the actual thing you can point at that has materially changed for the better?"
Economics
Citing a16z data, Zitron states that only 3% of US households pay for AI despite near-universal free access, and that AI companies spend between $3 and $15 for every dollar of subscription revenue through subsidized pricing. He argues OpenAI and Anthropic are both losing billions of dollars annually and present investors with two sets of earnings, one including training costs and one excluding them, which he calls "extremely dodgy." He criticizes the Wall Street Journal and other outlets for treating training as capital expenditure rather than cost of goods sold, arguing that training is an ongoing cost of goods sold rather than a one-time investment. He quotes economist Paul Kedrosky's view that AI is "nowhere to be seen yet in any really meaningful productivity data anywhere," and notes that AI appears mainly in non-residential fixed investment at levels he compares to railroad build-out or rural electrification, implying large investment with long diffusion timelines.
Anthropic's revenue claims
Zitron tracks Anthropic's reported annualized revenue from $9 billion at the end of 2025 to $14 billion (February 12, 2026), $19 billion (March 3), and $30 billion (April 6). He notes that Anthropic's CFO swore in an affidavit that the company had made only $5 billion in lifetime revenue as of March 9, 2026, which he reads as contradicting claims of $4.5 billion in 2025. The annualization method he describes takes the last four weeks of API revenue multiplied by 13, plus monthly subscriptions multiplied by 12, which he argues is easy to inflate around large product launches: the $14 billion figure accompanied OpenClaw hype and the Opus 4.6 launch, the $19 billion figure the Sonnet 4.6 launch, and the $30 billion figure the general availability of a 1-million-token context window. He argues the 1-million-token context window drives context bloat, since each prompt re-sends the entire context, so token consumption and revenue grow from more tokens per existing user rather than from new users. He states that Anthropic's gross margins are reported on a "Non-GAAP" basis rather than GAAP, citing Coatue data showing a 45% non-GAAP margin against roughly 38% by other reports, and that about 50% of Anthropic's gross profits from Amazon flow back to Amazon while Google takes 20–30% of net revenue.
The "early days" framing
Zitron treats "early days" as a deflection from questions of efficacy and economics. He contrasts AI with the dot-com era, when internet access stood at about 16% with speeds around 50 kbps, against AI's near-universal access and free or cheap entry, arguing that the comparison undercuts the claim that the technology simply needs more time to diffuse.
Reception and confidence
Zitron is an independent journalist known for critical takes on the technology industry. His sourcing includes The Information, Newcomer, official Anthropic announcements, and academic papers. The revenue analysis is investigative journalism that may reflect accurate accounting or aggressive interpretation. The agent-capability claims are strongly stated but largely consistent with available evidence that agents fail at sufficiently complex tasks. Confidence is medium: the piece is plausible and sourced but represents one strongly skeptical perspective, presented as argument rather than established fact.
The essay functions as an industry-skeptic counterpoint. It contradicts accounts of crossed autonomy thresholds (Source: shumer.dev) and of real agentic work (Source: oneusefulthing.org), and it disputes the Software Factory framing in The Shape of the Thing on agent capabilities. It challenges the current-capability claims associated with Agentic AI, contradicts the revenue figures and profitability framing on Anthropic, and supplies a skeptical reading of OpenAI financials. It aligns with the AI as Normal Technology view that AI is overhyped and that diffusion will be slow with a real deployment gap, and it adds the "no measurable productivity" position from Kedrosky to the contested account on AI and Productivity.
Relationships
- contradicts: (Source: shumer.dev) — Shumer's account of an autonomous AI threshold crossed
- contradicts: The Shape of the Thing — Mollick's Software Factory framing
- supports: AI as Normal Technology — agents are not what they claim; diffusion will be slow
- related: Anthropic — revenue analysis
- related: OpenAI — financial parallels
- related: Agentic AI — critical perspective on current agentic capabilities
- related: AI and Productivity — "no productivity data" claim from Kedrosky