Inhuman Vectors

Abstract

As the gap in capabilities between artificial neural networks and humans seems to be unavoidably closing, we are drowning in the incomprehensible buzz of hype, fear and denial. This technology, gradually developed over the last few decades, is often portrayed as something both altogether newborn and at the same time too abstract and alien to be considered human to any extent. Historically, antihumanist positions can be construed as more or less discrete deconstructions of humanist exceptionalism’s many manifestations, while still proceeding from its essentialist premises all the same. An attempt at stating this false dichotomy’s central problem anew, inhumanism is a distillation of the human rational core as the capacity-for and commitment-to continuous normative self-revision in the discursive space of reasons. In this paper we argue that there are meaningful similarities between how large language models (LLMs) and humans structure their respective images of the world, and that these similarities can inform our self-understanding and our commitments. To this end, we present a preliminary exploration of artificial neural networks as a mirror for humanity’s selfexamination, in view of the way LLMs’ latent spaces robustly embed semantic maps of meaning as vectors. The same process is at work with platform user data embeddings, materialized through informing the corporations’ counterintuitive algorithmic meta manoeuvres with massive quantities of vectorized behavioural data. In light of speculative insight into the metastability of LLMs’ semiotically encoded orientation, we see the volatility of LLMs’ vectorial value alignment not only as a challenge, but simultaneously as an opportunity.