A software and security engineer argues that readers should stop relying on OpenAI and Anthropic. Writing after reading Naomi Klein and Astra Taylor's End Times Fascism and the Fight for the Living World, the author says the American frontier labs are led by people who believe they are building a Machine God and are willing to risk the world to reach AGI. The author argues these companies are not unique, since many open-weight models now rival the latest ChatGPT and Claude models, even on complex tasks like vulnerability research.
Open-weight models offer clear advantages. They can be run privately on your own hardware, though that requires substantial RAM. Providers such as Tinfoil use secure enclaves to run them privately. Where privacy is less of a concern, many other providers are available, some quite affordable, and switching between them is easy, so users are not locked into a single vendor. The author also notes that ChatGPT and Claude are never private. Their closed models let the companies inspect everything users do, and an opt-out checkbox for training does not stop them from accessing chat sessions or sharing data with governments or companies.
The author also criticizes the guardrails on OpenAI and Anthropic models. A monitoring agent reads transcripts and halts sessions it suspects of hacking, and false positives are frequent. Access to less restricted models requires approval through programs such as OpenAI's Trusted Access for Cyber and Anthropic's Cyber Verification Program. Many US cybersecurity firms have gained access, and military contractors and government agencies likely have as well. The author calls it an undemocratic concentration of power to let Sam Altman and Dario Amodei, consulting the Trump Administration, decide who gets capable models.
The author concludes that the better path is to avoid depending on ChatGPT or Claude altogether. Rather than seeking privileged access from corporate gatekeepers, users can learn to run models such as GLM-5.3 to meet their needs. The author argues this is not difficult and that these models are good, with the added benefit that no cop agent watches over the user.
Source: Hacker News · Summarized by HeadlinesBriefing