I use a $200/mo OpenAI Codex sub. Throughout the workday I run an average of 2 concurrent agents of GPT 5.5 with high reasoning on fast mode and use less than half of my subscription usage.
For more interactive/active usage you might be better off using the low reasoning level, but I have usually found high to be a good balance of intelligence and generation speed.
I like to use Pi (https://pi.dev/), and I recently got it to make an approval extension for itself. It has a lot of documentation built-in for the agent to modify the behavior of the app.
I got it to display all proposed file change diffs and bash commands and made it so I can either approve the action or deny it with a message for it.
It was surprisingly easy to tell it to modify things things the diff viewing algorithm or syntax highlighting for the diffs.
Thanks for the info, I will look into this! I got past that battle and further into the game in my testing, but I don't think I've tried using an item in battle yet.
The reduction rules seem kind of arbitrary to me. At that point why don't you just use combinators instead of defining a set of 5 ways their operator can be used?
MLA makes it so the keys and values used are a function of a smaller latent vector you cache instead of a key and a value for each token. KV cache quantization reduces the size of the values in the cache by using less bits to store each value. These two approaches operate on different parts of the process so they can be used in combination. For example, you can quantize the latents that are stored for MLA.
There are papers that try to quantize angles associated with weights because angles have a more uniform distribution. I haven't read this specific paper, but it looks like it uses a similar trick at a glance.