The game being sub-tweeted is Modern Warfare 2.

Archive.org link for the initial (now missing) post, and for the (now missing) follow-up post.

We also noticed that agents tend to lose focus over time. Even within the span of one compaction cycle, agents can drift and lose focus, the more data the context holds.

It is absurd that LLMs still rely on one text input. Even Stable Diffusion uses two, for ‘do this’ and ‘don’t do this.’ You’d figure all the tweaking of attention heads and KV caches would see someone separate input, instructions, and output.

For obvious reasons, none of the code will be shared. Everything will stay private and I will use it for my own needs only.

My guy, someone’s Rust version has been out for a month.