You seem very thoughtful and careful about all this, but I wonder how you feel about the emergence of these abilities in just 3 years of development? What do you anticipate it will be capable of in the next 3 years?

With no disrespect I think you are about 6-12 months behind SOTA here, the majority of recent advances have come from long running task horizons. I would recommend to you try some kind of IDE integration or CLI tool, it feels a bit unnatural at first but once you adapt your style a bit, it is transformational. A lot of context sticking issues get solved on their own.

Oh I am very much catching up. I am suing Claude Code primarily, and also have been playing a bit with all the latest API goodies from OpenAI and Anthropic, like custom tools, memory use, creating my own continuous compaction algorithm for a specific workflow I tried. There is a lot happening here very fast.

One thing that struck me: models are all trained starting 1-2 years ago. I think the training cutoff for Sonnet 4.5 is like May 2024. So I can only imagine with is being trained and tested currently. And also these models are just so ahead of things like Qwen and Llama for the types of semi-complex non-coding tasks I have tried (like interpreting my calendar events), that it isn’t even close.