Group items tagged

"As AI systems become more general and more useful in the real world, ensuring they behave safely will become even more important. To date, the majority of technical AI safety research has focused on developing a theoretical understanding about the nature and causes of unsafe behaviour. Our new paper builds on a recent shift towards empirical testing (see Concrete Problems in AI Safety) and introduces a selection of simple reinforcement learning environments designed specifically to measure 'safe behaviours'."

But when cognitive science turned its back on behaviourism more than 50 years ago and began dealing with signals and internal maps, goals and expectations, beliefs and desires, biologists were torn. All right, they conceded, people and some animals have minds; their brains are physical minds - not mysterious dualistic minds - processing information and guiding purposeful behaviour; animals without brains, such as sea squirts, don't have minds, nor do plants or fungi or microbes. They resisted introducing intentional idioms into their theoretical work, except as useful metaphors when teaching or explaining to lay audiences. Genes weren't really selfish, antibodies weren't really seeking, cells weren't really figuring out where they were. These little biological mechanisms weren't really agents with agendas, even though thinking of them as if they were often led to insights.

This essay investigates detrimental thinking patterns about the future, termed futures fallacies. It is based on an analysis of the existing literature and personal observation. I define futures fallacies in three ways. First, as those thinking patterns that stand in direct contradiction to a truly desired longer-term future. Second, as thoughts and behaviours that are contrary to our best existing evidence, facts, and logic, of relevance to emerging futures. Third, as cognitive frames that ensure chosen strategies fail.

The analysis of questionnaires often involves representing the high-dimensional responses in a low-dimensional space (e.g., PCA, MCA, or t-SNE). However questionnaire data often contains categorical variables and common statistical model assumptions rarely hold. Here we present a non-parametric approach based on Fisher Information which obtains a low-dimensional embedding of a statistical manifold (SM). The SM has deep connections with parametric statistical models and the theory of phase transitions in statistical physics. Firstly we simulate questionnaire responses based on a non-linear SM and validate our method compared to other methods. Secondly we apply our method to two empirical datasets containing largely categorical variables: an anthropological survey of rice farmers in Bali and a cohort study on health inequality in Amsterdam. Compare to previous analysis and known anthropological knowledge we conclude that our method best discriminates between different behaviours, paving the way to dimension reduction as effective as for continuous data.

Prepublication peer review should be abolished. We consider the effects that such a change will have on the social structure of science, paying particular attention to the changed incentive structure and the likely effects on the behaviour of individual scientists. We evaluate these changes from the perspective of epistemic consequentialism. We find that where the effects of abolishing prepublication peer review can be evaluated with a reasonable level of confidence based on presently available evidence, they are either positive or neutral. We conclude that on present evidence abolishing peer review weakly dominates the status quo.