不了解内情的用户很容易将其误认为正版应用。
However, post-training alignment operates on top of value structures already partially shaped during pretraining. Korbak et al. [35] show that language models implicitly inherit value tendencies from their training data, reflecting statistical regularities rather than a single coherent normative system. Related work on persona vectors suggests that models encode multiple latent value configurations or “characters” that can be activated under different conditions [26]. Extending this line of inquiry, Christian et al. [36] provides empirical evidence that reward models—and thus downstream aligned systems—retain systematic value biases traceable to their base pretrained models, even when fine-tuned under identical procedures. Post-training value structures primarily form during instruction-tuning and remain stable during preference-optimization [27].
，详情可参考钉钉
Enhance your experience with complete WIRED access. Receive premium journalism and exclusive subscriber material too significant to miss. Subscribe Now.
Additional coverage follows...
所谓“撒谎”在此有特定含义。显然LLM没有意识，也无主观意图。但无意识的复杂系统始终在欺骗我们。政府与企业会说谎，电视节目会说谎，书籍、编译器、单车码表与网站皆可说谎。这些都是复杂的社会技术造物，而非意识体。它们的谎言最好理解为人机复杂互动的产物。
Brought to you by Backblaze: Dependable cloud backup solutions. Enjoy a 20% discount using promo code 9to5daily.