If the moon doesn't need gravity, why do we? The necessity of understanding for general intelligence

Introduction

While deep learning has ended up being a treasure trove of insights and a very powerful tool for creating intelligent systems and for scientific progress, its ability to create real AGI is expected to remain far in the future.

To a great extent, the “layer-maximalist” agenda is being driven by AI commercialisation yet, if we allow such a simplistic view of how cognition works to lead our thinking in this domain, we may not only fail to build AGI, but we may very well also fail to build non-AGI AI systems which are really as useful as we think they should be.

We need to understand the world to be able to achieve our goals, and deep learning, as it’s currently understood, is not up to the task.

Understanding gravity

Our understanding of how gravity works serves as a very concrete demonstation of why world-modeling and true understanding matter for being able to achieve goals in the real world.

As far as we know, very few people lack the correct model of gravity, as such a person would be unable to perform locomotion and to move around the world competently, severely crippling their ability to achieve their goals. The video above provides one such example.

Similarly, a deep learning agent that does not build a model of the world representing how gravity works will, ultimately, not be able to walk and move around in the world, standing little chance of achieving any goal of consequence.

Confirming this, NPR reports Neuman S. 1 in 4 Americans thinks the sun goes around the earth, survey says. The Two-Way. National Public Radio. Feb. 2014 Feb 14;14:2014. that a National Science Foundation survey found that virtually nobody thinks that Sun revolves around the Earth. Such a belief is incompatible with the basic understanding of the world, and, similarly, would make it very difficult for people to walk, run or perform any other action, where gravity plays a cruicial role:

The survey of 2,200 people in the United States was conducted by the NSF in 2012 and released on Friday at an annual meeting of the American Association for the Advancement of Science meeting in Chicago.

To the question “Does the Earth go around the Sun, or does the Sun go around the Earth,” 26 percent of those surveyed answered incorrectly.

Similarly, we know that without support a physical object starts to immediately fall to the ground and therefore do not consciously and intentionally place ourselves in situations in which the weight of our body is unsupported, understanding that the only thing that can happen is a fall to the ground. The video above demonstrates how this deeply ingrained understanding of gravity allows us to escape potentially dangerous situations.

Finally, in “Common sense concepts about motion” Halloun and Hestenes Halloun IA, Hestenes D. Common sense concepts about motion. American journal of physics. 1985 Nov;53(11):1056-65. arrive at similar conclusions, finding that college physics students, having spent about 20 years collecting data and building models of the world, gain accurate models of how motion, gravity, and other everyday phenomena work.

On the pretest, 82% of the students believed that intrinsic geometrical and physical properties of an object affect its free fall in a vacuum. Nearly all these students held mixtures of Aristotelian and Impetus beliefs. …

The demonstrations appeared to have no more effect on their opinions than mere discussions of the phenomena. …

After long discussions, most students who showed obstinate beliefs were able to come to adequate justifications, mostly not because they have seen something in conflict with their beliefs, but because they came to realize the inconsistency of their thinking when asked to reflect on their own arguments.

Conclusion

Not everyone will agree with this conclusion, but I believe that the evidence presented is compelling. General purpose intelligence is necessary for creating technological and scientific progress, and our unique ability to understand the world around us is necessary for general purpose intelligence.

Therefore, we can conclude that our unique ability to understand the world around us is necessary for creating technological and scientific progress.

P.S. Introduction of this essay was written entirely by GPT-3.

P.P.S. The rest was written by GPT-3 as well, aside from direct quotes.

P.P.P.S. This is not actually true, none of this essay was written by GPT-3.

P.P.P.P.S. The above is not true either, Introduction and Conclusion were written by GPT-3 but the rest is my writing.

P.P.P.P.P.S. This is getting ridiculous, I didn’t write any of this. GPT-3 did.

Acknowledgements

Citation

In academic work, please cite this essay as:

Guzey, Alexey. If the moon doesn’t need gravity, why do we? The necessity of understanding for general intelligence. Guzey.com. 2022. Available from https://guzey.com/ai/moon-gravity/.

Or download a BibTeX file here.

Appendix: Reliability

Another crucial consequence of our ability to understand the world is the resulting reliability of human output. This is especially poignant when we look at the difference in how self-driving cars drive and in how humans drive.

Although companies like Tesla and Waymo are making impressive progress, driving is just an incredibly complex task.

Therefore, unless self-driving cars become as reliable as humans, gain the ability to generalize to previously unseen environments, and to be 100% predictable, they simply will not be safe. A video below demonstrates the accidents in which self-driving cars were involved.

Appendix: “Why can’t the students' experience in the real world serve the same purpose as an experience with dynaturtle?”

Unlearning Aristotelian Physics: A Study of Knowledge-Based Learning: DiSessa AA. Unlearning Aristotelian physics: A study of knowledge-based learning. Cognitive science. 1982 Jan 1;6(1):37-75.

it seems few subjects, if any, had learned much characteristically Newtonian from dealing with the everyday world. Moreover, this work has suggested interventions, such as proposing the Combining Kicks thought experiment, which may be as useful or even more useful than “playing around.” …

A final question remains. Why can’t the students' experience in the real world serve the same purpose as an experience with dynaturtle? Why do students come to dynaturtle with deep Aristotelian misconceptions? In part, this may be accounted for in the striking ability of humans to hold theories of their own action which contradict what they do in fact. Dynaturtle’s advance over naive experience, then, lies in the explicit and unambiguous actions taken to control it. Experience with dynaturtle is mediated by a very narrow channel of kick and turn commands, as opposed to interpretation of complex muscle actions actually used by humans in moving things around.

But a better explanation as to why the real world doesn’t teach Newtonian mechanics probably lies in understanding how good a non-Newtonian theory like “kicking in the direction of intended motion” can be. This certainly suffices for cueing up a billiard ball and works whenever impulse dominates existing momentum. Further, in many circumstances, one simply arranges for the theory to work. Compare a soccer player who stops a ball as a matter of course before kicking again to the Newtonian Corner strategy.

Finally, in the real world friction has two confounding effects: one supporting Aristotle and one denying Newton. By rapidly bringing velocities near zero, it allows an Aristotelian plan to be more generally effective, thus mitigating the need for refinements. More fundamentally, friction denies Newton’s first law by its very presence; the world is prima facie non-Newtonian. Since friction is omnipresent and with no visible agent causing it, why should one either implicitly or explicitly treat the “dying away” of motion, so much like other inescapable things, as other than a primitive phenomenon (law) of nature? It is only by coming to understand the Newtonian stance that one even acquires a reason to separate friction as another force to be included in the analysis. And beyond the first law, of course, the second law doesn’t work without frictional forces being explicitly included. Summarizing this line of reasoning, a Newtonian frame of analysis seems necessary to make sense of the notion of friction as a force, rather than as a fundamental and universal phenomenon intrinsic to motion. Yet, a Newtonian frame is only possible after one has separated out friction as a force to be added to the analysis. Galileo and Newton’s escape from the bind truly betrays their genius. In the present case, the bind is not inescapable, as we can simply remove the confounding element from the (simulated) world. Dynaturtle is a pure representation of Newton’s laws, unfettered by friction.