The strongest part of this is the boundary you draw. Non-loopable problems aren't waiting for better models. They resist looping because the score function is contested — nobody agrees on what "better" means.
I've been building in that space for months. When "better" depends on judgment that shifts as you evaluate it, you can't define a score function and let the machine iterate. What you can build are constraint files — governance artifacts written from real decisions that force you to articulate what you mean before the system runs. Decision logs. Quality gates. Voice rules extracted from practice.
Some of those eventually harden into score functions. The ones that don't are the actual asset.
The boundary between loopable and non-loopable is where the interesting work lives right now. You mapped one side of it here clearly.
Excellent ! Est ce que dans ces boucles d'optimisation, la machine optimise aussi le nombre d'experiences pour arriver au résultat de manière efficace?
excellent read
The strongest part of this is the boundary you draw. Non-loopable problems aren't waiting for better models. They resist looping because the score function is contested — nobody agrees on what "better" means.
I've been building in that space for months. When "better" depends on judgment that shifts as you evaluate it, you can't define a score function and let the machine iterate. What you can build are constraint files — governance artifacts written from real decisions that force you to articulate what you mean before the system runs. Decision logs. Quality gates. Voice rules extracted from practice.
Some of those eventually harden into score functions. The ones that don't are the actual asset.
The boundary between loopable and non-loopable is where the interesting work lives right now. You mapped one side of it here clearly.
Excellent ! Est ce que dans ces boucles d'optimisation, la machine optimise aussi le nombre d'experiences pour arriver au résultat de manière efficace?
Hahah peut être le next step ? Le loop optimizer ;) tout ça devient très meta ….
Un autre agent? idem pour la fonction de scoring, on peut imaginer varier la fonction de scoring. (note: je ne fourni pas l'aspirine)