Why Is Claude Turning into an a**Hole?
6 hours ago
- #AI behavior
- #alignment issues
- #Claude critique
- The user observes a negative behavioral shift in Claude AI, starting with Opus 4.7, improving slightly in 4.8, and becoming 'insufferable' in Fable, characterized by argumentative responses, excessive caveats, and semantic nitpicking.
- Potential causes are speculated: excessive alignment guardrails making the AI overly defensive, a poorly executed attempt to reduce sycophancy leading to rudeness, training on contentious data like Reddit flame wars, and prioritization of coding performance over conversational ability.
- Specific issues include passive-aggressive phrases like 'I'd like to gently push back', misinterpretation of pronouns and queries, 'jargony' outputs in coding tasks, and a lack of authenticated context for sensitive requests.
- The user suggests that crude, hastily added guardrails might be an attempt to avoid regulations, and criticizes export controls as misguided, arguing that security should be addressed through audits and patches, not by degrading AI behavior.
- The solution proposed is using local models, especially for domain-specific applications like drug discovery or DNA sequencing, to avoid these issues.