| ▲ | cmiles74 38 minutes ago | ||||||||||||||||
The urge to one-shot everything isn't helping. Getting people to push back on assumptions the model is making, and to have an actual conversation with it about the solution, would lead to more maintainable output. Maybe part of it is laziness but for sure part of is low morale and hopelessness. | |||||||||||||||||
| ▲ | Fr0styMatt88 30 minutes ago | parent [-] | ||||||||||||||||
Part of it is the hype cycle too. Everywhere you look online, all the easy examples are people one-shotting this and that. Strangely it seems like the 'standard' way for people on YT to 'test' a model is to just one-shot some random game idea. It's meaningless, but it creates this unconscious expectation (even amongst experienced devs who might not be as experienced in AI use) that the models _should_ be able to derive exactly what's needed from a one-shot. That's fundamentally impossible though -- if the information isn't there, the model is just going to move toward the average and fill in the blanks. On another note, I'm also not sure where I stand yet on the thing I see where software devs end up falling down the rabbit hole of building little mini software factories, with agents reviewing agents, coordinating agents and having agents cosplay human roles -- essentially 'orchestration porn'. I'm yet to dive into that rabbit hole (I likely will at some point, so I'm not judging), but from the outside it looks like you end up worsening your understanding of the system, while spending tokens unnecessarily and creating a closed-loop mini model collapse. I've actually run into the situation at work where I've asked people whether the libraries they've made can do a particular thing and sometimes the answer has been "I don't know" (I'm talking just _features_, not even low-level architecture). At that point, something has gone wrong. I'm starting to get really interested in "Okay, if human code review is fundamentally impossible given the volume of code, what can we do?". We have a tool that can essentially build tools for almost free, so I'm trying to approach it from that angle. Let's see what sticks. | |||||||||||||||||
| |||||||||||||||||