William J. Marshall

Interesting, I found that models of the 9B size do well with a simple Problem declaration and solution approach in the Thinking.
Of coarse It was doing intense mathematics but that seemed to be the sweet spot.

This was written as soon as the Reasoning models were released so it was pretty reactive. I was working with heavily post-trained models.

are the experiments documented? I'd like to see them, I'm still wrapping my head around the effects of reasoning on long context models. It seems to work well with new information or on Single objective and it's come quite a ways on the industry harnesses. In dynamic long context back and forth I usually rely on a well trained model and an external State system over reasoning.

No problem. Experiments are fun, I guess we need a better way of coming up with and sharing novel solutions to problems. I've seen hundreds of people spend thousands of hours finding the same solution in different domains.
It really is a bit of a struggle getting the context window to function independently on new concepts. Figured you could look at what worked and use it. I'd love to see what you come up with..

How do you have the time to do this. Sad bro, go to reddit and do this crap, this platform is meant to be a constructive community. there are enough idiots working against Open Ideas and AI without us tearing each-other down.

We have a few unpublished projects which have automated research environments directed by a main agent inside of a project space where development and context are built and a small research assistant type model charts progress and gives simple progress updates, this pushes the the research but It does stall on some tasks(Mostly assembling the project. It's very janky though so I'm excited to see how your process works, we use sub 14B models mostly. If you want to check out the research suite I can share it with you on Github just shoot an email to either Intelligentestate@gmail.com or ops@cenedril.dev