I'm growing more and more frustrated at engineers wasting time building AI tools. I'm seeing a lot of people building these tools but not actually making anything. Documentation, memory, skills/plugins... I just don't think a lot of these actually DO anything and most people's time would be better spent using the LLMs to build software for people. Not more software for LLMs.

Anybody else noticing this? Anybody else hacking away all the bespoke tools, complex markdown files, and other extra stuff and just relying on the frontier models (and maybe an mcp or two to get your LLM controlled access to specific resouces)?

The nail gun analogy is only partially apt, and does not go far enough to describe what LLMs are doing to software.

Let's change "nail gun" to "robot." The robot can drive nails, but it can do other things too.

At a steady pace: Fewer houses need to be built because robots can, for some people and climates, just provide shelter when it's needed (if I must stretch the analogy this far, perhaps the robot can put up a force field or use it's body to morph into a bed, etc..). If you have a robot, you may not need a house in the traditional sense. This is not true for everyone everywhere.

Anecdotally, maybe every 2-3 months the improvements in LLMs negate the need to build more and more complex software. I directly test this and evaluate "can an LLM one shot what this software does" on a regular basis.

When a house does need to be built, the humans are not driving nails. The humans are configuring and transporting robots, ensuring they have the right plans and the right nails for each part of the house, and checking that the resulting structure is built correctly and of sufficient quality.

It's not surprising, then, that people who build houses are all suddenly very interested in joining the robot value chain in one way or another.