The way the big guys are feigning improvement right now is through recursion. They feed the LLM back into itself a few thousand/million times to narrow the normal curve of result accuracy. This is incredibly expensive, unsustainable, and not possible to do (in reasonable time frames) without data centers on data centers. Don’t expect open models to achieve the same kind of results with coding related queries because the underlying technology has plateaued.

I've run the local models like qwen and I doubt they'll eat their lunch any time soon, but I still think there will be a point when the open source tools are about as good and their costs are comparable.

From a tech perspective I don't think AI is going away at all, but the way we use it will evolve. I feel like these companies will crash hard and open source will start looking way more competitive when the prices stabilize and they're not able to just burn investor cash to stay warm anymore.

There will be a point that having a marginally better coding tool doesn't rationalize spending something like $1000 per month per dev, and it makes more sense to deploy local open source coding LLM stacks that assist your devs but for key things. I think we're already getting there but the tooling needs more work, and the open source models need to improve somewhat.