I was once an AI true believer. Now I think the whole thing is rotting from the inside.
GPT-4o workflows broke on GPT-5.5, says disillusioned developer.
A self-proclaimed former AI true believer has published a viral Reddit post detailing a stark reversal in his stance on large language models. The user, Complete-Sea6655, claims he built extensive automations and workflows around GPT and similar systems, but now asserts that the entire field is 'rotting from the inside.' He points to fundamental reliability issues: GPT-4o workflows that originally ran perfectly are now useless on GPT-5.5, and models often produce contradictory answers to the same prompt. He labels the industry mantra 'this is the worst it’ll ever be' as false, noting behavioral shifts, regression after updates, and context window hallucinations that cannot be version-locked.
The developer further argues that the overhead required to 'guardrail' and add safety layers often exceeds the cost of just paying a human. Even then, safeguards rarely function, leading to endless debugging of AI that won't admit errors. He criticizes the hype machine, calling Copilot, ChatGPT, and Gemini 'mediocre at best,' and notes that CEOs are quietly re-hiring humans or paying consultants to fix broken AI systems. Copilot, he claims, is the most requested tool to repair. The post concludes with a warning about zero accountability: these systems influence hiring, healthcare, credit, and legal outcomes without auditability, transparency, or regulation in the US.
- GPT-4o workflows that worked reliably are now broken on GPT-5.5, showing clear regression.
- Guardrailing and safety layers for LLMs cost more than simply paying a human to do the work correctly.
- Copilot is the most requested tool for consultants to 'fix,' indicating widespread business dissatisfaction.
Why It Matters
Highlights growing skepticism about LLM reliability in critical business and societal systems, with real regulatory gaps.