At least in cybersecurity, we already have seen diminishing returns with newer models.
At this point the harness/applayer matters more, as different models perform better or worse on exploit classes depending on the prompt, tuning, and various other parameters.
Of course, by the time HN hyperfixates on a topic, it's already been executed on and HN is too late.
Yea, in many usecases the tooling space is increasingly sophisticated context management such as fine tuning domain specific mappings into the model so that it is able to work directly with a compressed form of some data without needing to decompress into the context.
In larger models, these fine tuning techniques work more reliably/robustly. Because of this many usecases tend to prefer larger models. It is possible to work the same behaviour into the smaller model, but it requires more effort. But it's one-time. And smaller models are usually much cheaper. People make a tradeoff along this curve.
This is observed at few-B scale upto hundred-B scale. No way for us non-anthropic/openai to fine tune beyond that of course.