An article circulated from towardsai.net alleging the author reverse-engineered 200 Artificial Intelligence startups and concluded that 73 percent of them were essentially reselling or wrapping third-party models rather than building their own. The write-up describes a methodology based on browser inspection and automated tooling: watching network traffic and response timing, decompiling client-side JavaScript bundles, and flagging indicators such as requests to provider endpoints (for example api.openai.com or api.anthropic.com), matching latency and token usage patterns, and even finding a dozen instances of exposed API keys in frontend code. The author reportedly promised to publish the full methodology and scraping tools on GitHub.
The Hacker News thread is split. Many commenters accepted the core claim as plausible and argued that prompt engineering, retrieval-augmented generation, tool calling, and orchestration do require nontrivial engineering and can be legitimate startup work. Others pushed back on the article’s methods and credibility, noting that observing browser traffic does not prove how a backend is architected, questioning how one could reliably infer the backend flow or attribute traffic to specific providers, and pointing out signs that the piece may be generated or edited by a language model and is behind a Medium paywall. Several participants cited examples and mechanisms discussed in the article, including agent scaffolding, evaluator loops, and the tension between single-model and multi-agent designs.
The discussion broadens into market and product implications. Commenters debated whether wrapping a commodity model is inherently dishonest or simply a pragmatic lean startup approach, whether such wrappers have durable moats, and how venture funding incentivizes quick demos. Security and operational concerns also appear repeatedly: exposed keys, ephemeral client keys (OpenAI realtime sessions), and the need for proper evaluation pipelines to prevent regressions. The thread ends as a mix of skepticism about the specific headline figure and agreement that many small teams are currently focused on prompt and context engineering as their core product work.
