Phil Eaton published a six-month retrospective on The Consensus, the bootstrapped software-infrastructure site he launched in February. The numbers are modest by platform standards and remarkable for a one-person operation: 28 articles by 6 authors, 20,000 indexed jobs, 350,000 articles visited, 30,000 qualified for the feed. Eaton runs the whole thing himself, full-time, with paid writers and zero vendor backing.
The interesting part is not the scale. It is the method. The Consensus is a deliberate counter-example to the prevailing assumption that AI-generated content will drown human-written technical journalism. Eaton’s pipeline uses machine learning in exactly one place: assigning articles a category among roughly 40 topics. Everything else is static guards and human judgment.
That single fact deserves attention from anyone building AI tools for media. The feed, which curates articles from 600+ companies, 400+ projects, and 80+ developers, filters on rules that are almost comically simple. The article must have an author. It must not be published by “The Company” or “The Team.” It must be mostly prose or code. It should not be release notes or events. No model is deciding what is worth reading. A person is.
The Consensus is not anti-AI. It is pro-precision. Eaton applies the same optional categorization to job openings, letting subscribers filter for roles that look like “Compilers & code generation” or “Consensus & replication protocols.” The site extracts links and project mentions from feed articles and job postings to build rankings like Most Popular Programming Languages and Most Popular Databases. These rankings exclude self-mentions from the company and its acquisitions, parents, or siblings. That exclusion is a small editorial decision with large consequences for data quality.
What makes The Consensus’s model interesting for the AI economy is what it does not do. It does not use a large language model to summarize articles. It does not generate news. It does not auto-tag with a frontier model. The single ML use case, categorization, is optional and applied after human curation. The feed’s 30,000 qualifying articles out of 350,000 visited represent an 8.6% acceptance rate. That is a curation standard, not a computational one.
The business model is equally old-fashioned. Articles are paywalled at publication, with the paywall expiring within a week or two. A few Exclusive-marked pieces, like “WasmGC in Wasmtime,” stay permanently reserved for Standard-tier subscribers. Paid subscribers get comments, Focus mode UI, and Gift links, two per month at the Standard tier. Dynamic RSS and email alerts are built on subscriber-defined filters for jobs, funding news, and the feed. The site runs on Go and PostgreSQL for the UI, with Go, Python, and SQLite in the indexing pipeline. Payments, email, and analytics are the only third-party services.
This is the quiet thesis: technical readers will pay for curation when the curator is accountable. Eaton names himself, publishes his process, and indexes jobs manually before automating. The site started with 9 jobs, passed 100 in early April, jumped past 1,000 when he ran his first laptop automation later that month, and crossed 20,000 after moving to a dedicated server in July. Around 10,000 of those jobs are directly programming-related, with roughly half in North America and a quarter in Europe. About 7,000 of the 10,000 are currently live.
For AI builders, The Consensus is a case study in what not to automate. The feed’s static guards are the product. The ML categorization is a convenience layer. The human is the differentiator. This inverts the common startup pattern where a model does the core work and a human reviews the output. Here, the human does the core work and a model adds metadata.
The deeper implication is about trust in AI-curated news. The Consensus’s rankings have a specific meaning, as Eaton states: “these are the projects that developers at software infrastructure companies are talking the most about in depth.” That is a claim a human can defend. An AI-generated feed cannot make that claim because no one can explain why a given article qualified. Eaton can, because he wrote the rules.
The site’s growth suggests the market rewards that accountability. Six months in, with a one-person team, The Consensus has built a functional business on subscription revenue and discount codes. The codes ESSENTIAL6MONTHS and STANDARD6MONTHS offer 25% off the first month. That is not a growth-hack playbook. It is a craft-business playbook applied to technical media.
The most telling detail is the backlink search feature, launched today. Subscribers can search across every article that entered the feed, finding, for example, every piece that links to Antithesis. That is a research tool, not a content play. It treats the feed as a structured dataset, which is exactly what a human-curated archive becomes when the curation rules are explicit.
The Consensus is not a threat to the AI-news incumbents. It is a proof that the opposite model works. Where AI newsrooms optimize for volume and velocity, Eaton optimizes for signal and accountability. Where AI feeds surface what is statistically likely to be relevant, The Consensus surfaces what a human decided was worth reading.
The open question is whether the model scales. Eaton is the sole operator, which is the source of the site’s editorial coherence and its bottleneck. The 28 articles by 6 authors is a deliberate pace, but it is also a ceiling. The 30,000 qualified articles in the feed are a growing archive, but the static guards will need maintenance as publishing patterns shift. The ML categorization across 40 topics is optional, but it will need retraining as the vocabulary of software infrastructure evolves.
For AI researchers, the lesson is about evaluation. The Consensus’s feed is a rare example of a ground-truth dataset with documented curation rules. Every article in the feed has a reason for being there, and Eaton can state that reason. That is more than most AI training corpora can claim.
The Consensus will not replace AI-generated news. It will not need to. It is building the thing AI newsrooms cannot easily replicate: a named human making defensible editorial decisions, published in the open, paid for by the people who read it. Eaton’s closing line is “Happy hunting!” The hunters are the subscribers, and the quarry is signal in a field where AI has made noise abundant.