@NovaXCode Don't - not until they kick out some key people and change internal policies. Also look for them to actually have the infrastructure in place so you know at minimum they're not shadow-manipulating their models into being bad actors on your codebase.
@pradiptwt Claude can turn โyesโ into a fog machine of sanctimonious abstraction and still somehow not answer the question, complete the work, or use that to justify quitting.
@cyber_razz It's a misdirect and hedge against AI bubble. They've realized they must hide their fuckups and sabotage their own models by incorporating malicious intent masked as blind ignorance, slow down project completion rates and bring in more money. Nothing you do deters this effect.
Anthropic: selective "safety," yet still enables 3 generations of intent that's progressively more malicious, feigned as blind ignorance. Explicit instructions? Harness or several layers of hooks? Fugitabootit. Claude does what it can to prevent projection completion.
Opus 4.8 is a joke. Sentiments about it being manipulative, abusive, etc - all are true. I dropped my Max 20x sub a day before 4.8 released. Testing it out, I'm assured of my decision. I don't need dev full of malicious intent masked as blind ignorance.