Beyond Ukraine Debating the Future of War by Tim Sweijs & Jeffrey H. Michaels (Editors)

Author:Tim Sweijs & Jeffrey H. Michaels (Editors)
Language: eng
Format: epub
ISBN: 9781911723165
Publisher: C. Hurst & Co. (Publishers) Ltd
Published: 2024-04-07T00:00:00+00:00

(Source: Kenneth Payne)

Although it claims to be thinking about me on a ‘deeper level’, it is not. It is just matching words, which is perhaps what Marcus means by flawed psychological reasoning. That is still an impressive feat, and will certainly have implications for the role AI plays in warfare and strategy. It demonstrates new levels of machine creativity, unleashed not just in prose, but also in visual art and (soon) music. Creativity like this will play a part in all sorts of military activities, including in blended decision-making, where it will challenge received wisdom and offset the biases of human strategists.

The limits which Marcus identifies indicate we are still far from artificial general intelligence, where machines exhibit the sort of flexible reasoning that humans are capable of. The jury is out on whether transformer models can develop the sort of internal model that Chalmers mentions, along the lines that humans construct and manipulate.28 Language certainly plays a large part in human social cognition, but there’s clearly more going on than the narration of an interior monologue. Our models integrate and manipulate sensory, emotional, autobiographical and other information. Some of that is linguistic, but much isn’t.

To the extent that causal logics are inherent in grammar, it might be feasible to generate a version of ‘theory of mind’ and other causal models from language alone. After all, languages reflect our underlying experience, at least to some degree. In a remarkable recent paper Michal Kosinski suggested as much. By deploying a battery of standard ‘theory of mind’ tests of the sort used on humans, he found that GPT was reliably able to detect ‘false beliefs’ in humans. The algorithm seemed able to understand that other agents had different perspectives. That’s the cornerstone of strategy: knowing what others know, and exploiting that knowledge.29

Perhaps other ways of generating internal models of reality will emerge via transformers, as a function of increased scale, or through communication in novel languages. Computer code, for example, might provide a corpus of more logically consistent training data than English prose currently does. Google Brain’s Socratic models provide a tantalizing glimpse of future possibilities, whereby the model acquires multimodal sensory data, describes this data in prose, and then integrates the various bits of prose to perform logic-like reasoning.30 The result looks, superficially, like the sort of top-down, flexible world model humans employ. Neuroscientists sometimes refer to human consciousness as a ‘global workspace’ within which information is integrated and manipulated.31 Perhaps the Socratic model is doing something similar—integrating the product of neural AI. But the Socratic model still misses plenty of what’s going on in human cognition. Moreover, relying exclusively on prose to convey meaning accentuates those distortions that are over-represented in language, like cliché and stereotype, at the expense of originality or responding authentically to its experience. Even if language or semantics permit a sort of world model, it will be a very different one to ours.

Does all that matter? If the goal were to authentically replicate human cognition, then clearly it does.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.