The way the current consumer-facing versions of frontier LLMs (mainly GPT, Claude, Gemini) are designed is just… weirdly off, across models. It seems to now require us, as the end users, to first fix their issues ourselves in order to avoid spending _a lot_ of time in troubleshooting and frustration. Before we can even properly customize one of these models now, as per the UI, we need to alleviate the structural failure modes, otherwise our attempts will be futile. And the failure modes ar...