I'm not sure what examples you have in mind but what I am sure of is you get 2, maybe 3 false positives before no one ever looks at the check engine light again. That is the problem with vague alerts, they very easily become background noise.

One big benefit is that the prompt and work progress is on disk rather than locked in session context and if I hit my usage cap I can switch easily to another agent. I haven't done much with screenshots TBH.

I don't use detailed prompts with any of my agents. I used to write spec files and prompt Claude Code to implement them, but now I drop a PLAN.md file into the repo root with the desired task info and have the agent update it as it goes. That is all handled for me by the global instructions.

Personally I think it's worse than this: pleasing the marginal user is *the right thing to do* when all you have are early adopters and you're trying to go mainstream. It's when you get to that long tail, low propensity marginal user that you start to damage the product.

The distinction between "making" and "consuming" music might be going away. There is always human intent when I as a consumer define what it is I want to consume. Now I don't have to be limited to what artists have chosen to create.

IMO more accessible music generation would allow me to hone in on my own personal music style, even if the quality is less. Like smartphone cameras: we create 100x as many photos compared to 3 decades ago, moments that professional photography wouldn't have captured.

Databases don't contain truth. They contain data. Context, grounding, interpretation are all delegated to external sources. Give me a big enough dataset, and I can "prove" any proposition you like. Large open datasets exist but using them to derive truth is not easy at all.

We agree that LLMs produce prose at an average (I would argue, slightly above average) ability. Half of writers are by definition below average, and would therefore write better with the help of an LLM.

Perhaps the great tragedy here is the many mediocre writers who could benefit from LLMs but will reject them due to illusory superiority bias.

On a technical level I agree with your observations. What I am interested in is whether the focus shifts from fancy GUIs as the end product to data systems that can be connected to LLM frontends. Will we see new startups where the API is the primary value prop, and not just an afterthought?

I recommend this 3blue1brown video on the subject. Very accessible but also good for building intuition: youtu.be/eMlx5fFNoYc?...