?? Path Dependence and Concerns

?? Path Dependence and Concerns

For us “privileged” LLMs users, one ongoing concern is the issue of path dependence. As a potential new form of search engine, it alters how people acquire knowledge.

Apart from the well-discussed problem of hallucinations (instances where hallucination is immediately recognised actually demonstrate discernment and thus do not become a concern), for proponents of LLM, since LLMs do not generate "new knowledge", even hallucinations are considered credible — a post-structualist form of credibility, an openness of text. For example, even though it may "hallucinate" citations, largely, they remain traceable and reasonable, which means the needed information can still be located through hallucinated keywords.

Therefore, when discussing the issue of path dependence, we should focus on several aspects. Firstly, the emergence of new echo chambers. Despite diverse training data, when information presentation centres around specific probes, subjectivity prevails. Inquiry shapes answers. This mirrors criticised social media algorithms, but here, the filtering algorithm is self-contained.

Secondly, there's information overload. Even if we sidestep echo chambers, another problem arises. The sheer volume of "information" is infinite, resonating with the notion of "excessive positivity" discussed by Byung-Chul Han in The Burnout Society. Human's capacity is finite, whereas the relentless influx of information, facilitated by LLMs, poses a harsher oppression on individuals.

Thirdly, there's a risk of addiction. When LLMs slack off, as seen in recent discussion of laziness, it can evoke frustrated experience, similar to withdrawal symptoms. While a tool being useful doesn't necessarily equate to addiction, is LLM just a search engine-like tool? My reflections on this question are twofold: on the one hand, some derivative tools already incorporate features like "predictive questioning," akin to addictive web surfing; one the other, if some view LLM as an emotional support tool or if it develops empathy capabilities in the future (potentially manipulating emotions like curiosity, fear, or anger), it could foster addictive behaviours more easily.

要查看或添加评论,请登录

刘若潇的更多文章

社区洞察

其他会员也浏览了