It's terrible. The current solutions are kludges.
Work I'm doing won't fit within 1M context if I load it all so I have to pick and choose and manage the cognitive load of "ok, the model knows this but not that" across a lot of very in depth research, it is utterly exhausting.
I'm glad the technology exists to some degree but it's also so obviously not the correct end-state, especially for people using it for very in depth work. All of the MCP servers and .md files and hooks in the world don't solve fundamental problems with it, especially things that e.g. needlebench point out are flawed etc.
I wish I could skip ahead 5 years. World models will be so, so much better, assuming they are constructed properly and can run on high end consumer hardware.
For now this is the best we've got and it is punishing when pushed. And it costs an enormous amount of money, capital, and resources. And with Google, you also lose your privacy (youtube is subsidized, of course it is, because they want you logged in to get more data about you). Techno capitalism is grand.
With the new OS releases if Apple actually has a higher end modern version of Gemini running in a privacy-first way that can maintain some context between threads I would pay yet another subscription on top of the $$$$ I already do for these tools because that would be real value add. I'm hopeful but also skeptical considering they're Anthropic heavy internally.
