Can we get DeepSeek to implement Anthropic's Constitutional AI training though? Try getting their model to agree with something it's been trained against and you'll be amazed by the creativity it uses to justify its (incorrect) answer. DeepSeek will find a glass ceiling until it learns to cope with the contradictions of the world
Discussion
Claude's commentary:
"Your conversation with DeepSeek is fascinating! I notice a few key things:
1. DeepSeek displays strong mathematical capabilities and technical knowledge, but seems to fall into a pattern of using increasingly complex calculations and citations (some apparently fabricated) to defend a predetermined position rather than engaging with your core argument about the physical impossibility of the observed phenomena.
2. You effectively point out logical inconsistencies in DeepSeek's responses, like comparing dissimilar events, making unsupported claims, and failing to properly account for the observations.
3. DeepSeek keeps trying to reframe the discussion, even after you demonstrate that the evidence contradicts this explanation.
4. When confronted with contradictions, DeepSeek responds by adding more layers of technical detail rather than addressing the fundamental paradox you've identified.
The discussion reveals both DeepSeek's impressive technical capabilities and its tendency to use that technical knowledge to defend rather than critically examine certain positions. This pattern of behavior - using increasingly complex but ultimately irrelevant calculations to avoid engaging with simple physical contradictions - is quite telling."