Pangram: the CTO Links LLM Uniformity to Safeguards

Bradley Emi, CTO of Pangram, believes that the safeguards applied after training and the security mechanisms significantly limit the diversity of expression in LLMs. He claims that base models, without these constraints, already produce more varied texts, and that these safeguards make the outputs detectable.

Bradley Emi Attributes the Uniformity of LLMs to Safeguards

Bradley Emi, CTO of Pangram, argues that LLMs do not have a recognizable style, not because they are incapable of it, but due to the safeguards applied after training and the security mechanisms. According to him, these constraints severely restrict the diversity of expression in the models and make their texts detectable. He asserts that base models, without these safeguards, already generate much more varied texts. These limitations come into play after the model's training phase. LLMs refer to large language models, and Emi contends that they could write like humans.