I just had a long session planning an implementation with Opus 5.5 in a reasonably complex system. It was a frustrating affair where it suggested very complex solutions that, sure would work, but would be very slow and difficult to maintain with the rest of the solution. After a long discussion with many variations of complex designs, we found a very simple solution. Elegant even.
I think the problem in this case was:
1. Unusual constraints, that kept nudging the model to overdesign or misunderstand parts of the solution at every turn.
2. One code-comment it found and thankfully mentioned, that a previous session had made up and left there as a fabricated requirement. The model trusted this "requirement" more than my instructions, which unfortunately made sense, since it was indeed written and part of the code.
I'm working on a system that has a mix of trust boundaries with a few perspectives that are a-typical and it trips up the models working on it. Fable 5.2 didn't do better at all.
I have tried to put clear instructions in an agent architecture file, but I haven't managed to get it fluent yet.
To me, it sounds like the point of the article is that we're skipping the nuance and jumping straight to effectively declaring equality between AI use and slop, which is a bit sad in a couple of ways:
a) effort is apparently so rare, that the distinction between effortful content creation with AI and opportunistic press-a-button "content" creation is too difficult.
b) real pervasive concern about our future (rightfully?) motivates sweeping judgement against the whole creative space with any AI involved.
c) concretely, its sad in itself that any arena of expression is overwhelmed with low effort.
On the other hand, by the very nature press-a-button content is so easy to make that it will naturally bury real content unless someone takes an active gate keeper role. (Every channel on Youtube I know has made in-depth complaints that this is not happening there.)
I hope we find a way around it, so that we can enjoy content and products that is made with the same amount of effort as before but with more powerful tools creating a better end result. Not to mention lower barrier to entry into creative spaces, giving us more perspectives and ideas expressed in a quality way.
I have to say that I'm very pleased with this version.
My prime anecdotal reason:
I asked Opus 5.5 to make a sweeping change with a few Sonnet agents. It clarified the scope, we agreed and then it started. I saw some problems on the way and asked it. It's answer:
This didn't go as planned. The Sonnet agents don't have the right judgement capacity for this, so they take shortcuts based on their limited scope X. I suggest we terminate them, roll back their changes and I can schedule an Opus agent to do it instead. It will be slower but done correctly.
There is a first time for everything.
Every interaction I've had with Opus 5.5 so far feels like talking to. An adult.
I think the problem in this case was: 1. Unusual constraints, that kept nudging the model to overdesign or misunderstand parts of the solution at every turn. 2. One code-comment it found and thankfully mentioned, that a previous session had made up and left there as a fabricated requirement. The model trusted this "requirement" more than my instructions, which unfortunately made sense, since it was indeed written and part of the code.
I'm working on a system that has a mix of trust boundaries with a few perspectives that are a-typical and it trips up the models working on it. Fable 5.2 didn't do better at all.
I have tried to put clear instructions in an agent architecture file, but I haven't managed to get it fluent yet.
reply