Death by Convenience

As already outlined in my last post, one common feature of most of the interfaces we currently use to interact with LLMs, at least in the context of software engineering, is their addictive nature. Agents work rather independently and ask us occasionally for feedback or further instructions, but the way we provide these responses is almost always the same. We either write another quite verbose prompt (usually because our enthusiastic friend has gone off the rails again), or we press a number that most of the time isn't greater than five to choose from a pre-filtered selection of actions. One of these usually allows us to provide more detailed guidance in the form of another free-form prompt, and so doesn't really differ from the first option. In any case, we're dealing with either a shortcoming of the tool or a lack of clarity in the instructions we as humans provided it with.

Having to write out orders in such excruciating detail that they almost resemble a formal specification is extremely cumbersome,It's worth noting here that the fault does not inherently lie with the tool itself or its limitations, but rather with how its capabilities are made available to us. and so we often opt for simpler variations and further refinement later on after the inevitable failure of the rather rudimentary first attempt. As this ordeal drags on, we end up choosing the more convenient second option instead: pressing a number from one to five. One doesn't have to resort to more sophisticated architectures commonly called "agent orchestration" to end up taking the backseat at work when being provided with means as frictionless as a single click. Being presented with output that most of the time looks okay, the majority of easily verifiable results at least coming out seemingly correct while passing our crude current verification steps, we don't bother paying closer attention to the details anymore if the overall artifact we end up with comes off as good enough.

The whole dynamic feels almost identical to the current state of social media, where we monotonously keep scrolling through content without ever really engaging with any of it, the sole difference being that now we don't scroll but press a key in the context of intellectual work instead — both closely resembling a slot machine of some sort, something that might spit out what we hope for if we just keep going at it long enough. The end result doesn't change though. I leave it to the reader to extrapolate where this might lead us and the consequences we may have to deal with then.

I call the end state, both for individuals and our industry as a whole, Death by Convenience.