This paper proposes a framework for comparing natural intelligence with human intelligence. The central idea is to go beyond the functional comparisons that are typically made into a three dimensional approach that also accounts for the system of intelligence development (named aetiological development) as well as inference (aetiological inference).
I enjoyed reading this paper. And it reflected some disparate thoughts I have had about the field. However, I don't think it is ready for publication yet; the treatment of the domain I find too shallow (the paper is very short and did take advantage of the full page limit to go in depth). Here are my main concerns:
-
Main argument is obfuscated by unnecessary jargon: This paper is trying to address a timely and fundamental problem in modern AI research. The theme of such a paper (if done right) should be read by all the community. The use of the word Aetiological in the title and main text , I would say, acts as a turnoff. It is a word that is unknown, never really explained in the abstract and never defined. This word is entirely new to me and I still could not understand the relevance of its use in the entire paper. After reading, I looked up the definition on line. The only definitions I could find (Cambridge, Miriam Webster) all point to the study of causes of diseases. Aside from this, it seems that the British spelling of the word is used here as opposed to the potential American spelling (where the a is removed). To make this paper accessible to all, I would advise to use plain English and to use spellings that will maximise on global reach. Using keywords with different spellings could at least be helpful.
-
Lack of Embedding with SoA": I am not an expert in this topic but have passing knowledge on some attempts to quantify intelligence. There is very limited discussion of the array of contemporary attempts to measure intelligence. I think the paper needs to discuss more extensively trends in existing literature, what they capture and not. Perhaps a table with a list of works and which ones cover what dimensions deemed important could provide a more compelling case. Here are just a few examples (non-exhaustive):
- Open-Endedness is Essential for Artificial Superhuman Intelligence: https://arxiv.org/abs/2406.04268
- A. Zadeh, M. Chan, P. P. Liang, E. Tong and L. -P. Morency, "Social-IQ: A Question Answering Benchmark for Artificial Social Intelligence," 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 2019, pp. 8799-8809, doi: 10.1109/CVPR.2019.00901.
- Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models: https://arxiv.org/abs/2206.04615
- Spontaneous Theory of Mind for Artificial Intelligence https://dl.acm.org/doi/10.1007/978-3-031-60405-8_5
-
connection between different frameworks of intelligence measurement. The CHC framework looks interesting. So do the IBM topics. However, I missed a deeper discussion linking these back to the three dimensions mentioned. It seems they are operating on different levels of granularity. And neither are really sufficiently useful to capture functional intelligence let alone the others. So I am not really sure how to link this with the three dimensions.
-
Argument that artificial systems do not develop in the same way as natural intelligent systems could have more depth. The argument about how AI systems develop differently from NI systems seems to suffer under a burden of proof issue. AI systems have been long trying to simulate natural processes, including how to develop under non pre-defined or fixed reward functions. This seems to be to be a claim on a moving target. And I am not sure why it is described in this way. I would have been more convinced if properties of development in NI systems could be put into a table against various families of AI system types. As it is, very few approach types are discussed. Some examples include: continual learning, curriculum learning, human-in-the-loop or hybrid intelligent systems, policy-dependent reward shaping, evolutionary algorithms, artificial swarm intelligence, multiagent systems, to name but a few.
5 Why do we want to compare the aetiological develop and inference?: Human behaviour is biologically governed by fundamental desires like procreation, a sense of belonging, need for safety, need for fat and sugar... these can lead to undesired outcomes for the same human in extreme environments (e.g. getting overweight from ultra processed food). I struggled to understand why we would want to compare artificial intelligence to these kinds of natural intelligence settings. e.g. are we wanting to compare humans that developed in a bad environment but maintained intelligence level compared to humans that developed in a good environment?
6 Computational models of human behaviour and cognition: I missed a discussion of other forms of computational models that could have some of the required properties for natural intelligence. e.g. there is the whole domain of cooperation games in social psychology (e.g. prisoner's dilemma, public goods game) which have been used to model emergent cooperative behaviour in humans. Evolutionary game theory is another popular one, and multiagent modelling. So computational models exist but their intended purpose can be quite different (e.g. being used to understand and model human behaviour and cognition). The boundary between computational models for social science and AI is starting to blend however, as more complex computational models are being used to model human cognition such as for example the wealth of works on artificial theory of mind.
e.g. Machine Theory of Mind, Neil Rabinowitz, Frank Perbet, Francis Song, Chiyuan Zhang, S. M. Ali Eslami, Matthew Botvinick Proceedings of the 35th International Conference on Machine Learning, PMLR 80:4218-4227, 2018.