@garybernhardt Shouldn’t we design a benchmark that evaluates this? What you describe is even more important than a strictly correct solution. If the only objective is “tests are passing” it will write all the slop and be rewarded for it
@thsottiaux Remote control has never worked on both iOS and between macs (endless loading states) + the occasional “Reconnecting 1/5…” which makes the thread lost forever. Could very be a networking issue on my side but I have no clue to debug
@devagrawal09 I did in many commercial project, but when building a self-hosted software I want the deployment to be minimal and self-contained. I wish there was a way to run a sync engine in memory and backed by SQLite. Maybe it exists already?
@thsottiaux I wish I could somehow create a new tab or have back / forward buttons. Eg I’m looking at an automation result, then I want to check something in another thread that just finished, All state is lost and can’t go back easily
@Xefreh@itechnologynet@DFintelligence C’est pas totalement faux, avec copilot les modèle sont capés à 128k token de context et ils se réservent encore 20 ou 30% de ça pour la compaction. Un prompt avec les outils et tu es déjà à 40% de contexte d’entrée de jeu. c’est difficilement utilisable pour des grosses tâches