A couple of weeks ago I wrote out my late night ponderings on A.I. I have been interested in artificial intelligence since my undergraduate days at university. Which later propelled me to focus on developing it as a graduate student of physics.
When personal computers were first commercialized, their uses varied greatly. Yet, the public wasn't immediately prompted to use them for drawing dancing teddy bears in uniforms on top of tricycles.
Seems odd, right? That's because the suggestion of how to utilize this powerful tool has trickled down into these vague, pointless applications. I'm frankly over it.
Image of one of the movies that inspired me into the realm of artificial intelligence.
I've always questioned whether this "AI" is truly artificial intelligence at all. All it seems to be are, rightly so-called, language models. However, language is central to our intelligence. An artificial language model that can summarize information and answer questions quickly does have some connection to how our biological intelligence works. We humans compute at astonishing rates, and as the top species on the food chain, we primarily communicate through language. One might conclude that this is a narrow classification of what intelligence is.
Yet, even if someone suddenly became deaf, mute, or blind, they would still be considered intelligent. No one would think of a person with such disabilities as suddenly being non-human or unintelligent. Similarly, even if someone's computational abilities varied drastically, they would still be considered intelligent, albeit perhaps not as intelligent as someone with exceptional skills.
So, what exactly is our definition of intelligence? Can these language models effectively be seen as a form of intelligence, or are they just powerful, fast language models that we've given a misleading name? Perhaps a more accurate term would be the "PLM," the Powerful Language Model tool. They perform statistical predictions of shear computational rigor in a matter of seconds. Though how impressive that is can it be deemed intelligent or computing as fast as a human does. Large language models have access to millions and millions of books, texts, and even classified information and we have no concept of who said what when, in a private setting to then say what the model produces on the screen is in fact original thought.
These models simply pull information from somewhere and summarize it, much like a skilled student who can craft text in their own words. Who knows what sources these models quote or reference? We mistakenly assume these models are actually thinking.
I believe it would be a healthier and wiser approach to develop intelligence by first examining the process of thought itself, then focusing on how that thought is communicated. Communication itself cannot be considered the origin or definition of intelligence. Humans communicate intelligence in various ways - through language, music, art, numbers, and movement - but all these methods reference something more fundamental. Thus, if we truly want to achieve artificial intelligence, we need to go beyond the tools used by intelligent beings and delve deeper into the core processes of thought.
You can see from the above piece that it is primarily my own thoughts and mullings made over some time that converged in one late-night expression. I recently got a better and fuller argument by listening to the Podcast of David Eagleman a neuroscientist. He delved deeper into the conceptualization and sentience of a being having intelligence. In his podcast, he comes up with a test that seems to focus more heavily on the characterization of intelligence than the Turing test where he claims that “ If a system is truly intelligent it should be able to do scientific discovery”. Which is something after some time I agree with. It deals with creative original thought and conceptualization.
Recent work by Pekala et al on “Evaluating AI-guided Design for Scientific Discovery”, only showcases that AI can be used to aid scientific discovery by the rate of novel scientific discovery hypotheses and compares the performance of different adaptive sampling strategies.
Journal articles published in recent months in Nature highlight powerful statistical tools such as GT4SD that significantly assist researchers with the consideration of paths for scientific hypotheses.
Feelings of doubt are beginning to dissipate on my theorizing as I see an alignment of thought with popular AI theorists, Marvin Minsky, Judea Pearl, and Yann LeCun. One quote that stands out by Marvin Minsky is "The problem of creating truly intelligent machines has been difficult because we don't have a good theory of how minds work."
The key to developing AI from a fundamentalist view is firstly by a well-developed classification and understanding of intelligence. I would take a leap here and suggest that artificial systems be classified by its own definition of intelligence and that intelligence differs for humans to machines. This is no way seems to simplify the outstanding and brilliant performance of an artificial system. It only seeks to distinguish the two as separate entities that at specific areas can overlap but are essentially independent.
Comentarios