TRANSFORMERS - NOT AS LARGE A THREAT AS PERCEIVED

It's no secret that in the rapidly evolving field of A.I., transformer-based language models like GPT-3 have garnered significant attention, fascinating many with their seemingly intelligent outputs.

However, a closer inspection reveals that these models, albeit advanced, are far from exhibiting "real understanding" or "reasoning".

This article aims to demystify the capabilities of such models, emphasizing their limitations and the illusionary nature of their "intelligence."

The Mimicry of Intelligence

At their core, transformer-based language models are akin to sophisticated mimics. They don't truly understand the content they generate; rather, they replicate patterns in the languages they've been trained on. This process is similar to one where a skilled impressionist may be able to replicate voices and mannerisms perfectly, but lacks an understanding of the underlying content.

Transformer-based models like ChatGPT therefore excel in identifying and replicating language patterns, a skill that often masquerades as intelligence, but is fundamentally different from true cognitive understanding.

The "Two-Dimensional" Nature of LLMs

To understand the limitations of these models, it's helpful to conceptualize them as operating in a "two-dimensional" space. This analogy likens the models to detailed maps that represent landscapes with impressive accuracy. However, just as maps lack the actual depth and complexity of real landscapes, language models lack a critical dimension – understanding or consciousness. They operate within the confines of their programming and training data, unable to transcend these boundaries to achieve true comprehension or original thought.

A means to an end

Researchers were focused on the creation of an "AGI" when transformer technology was "stumbled upon", causing an abrupt halt in AGI research as the viability/utility of LLMs based on transformer topology became apparent, but make no mistake, this is merely a temporary diversion as the potential of these "simple" LLMs is explored.

This means that the danger is still on the horizon, but this is still very far away, and absolutely not posed by the current state of A.I.

Except for lost jobs

This article states that there is LESS danger in GPT LLMs than perceived, not that any dangers are absent.

Greed will axiomatically demonstrate corporate willingness to displace the very consumers of the products they produce, and that the consequences of such activity are not sustainable, causing corporate collapse leading to even more job losses before sanity prevails and informs incumbent corpratards that their human customers need jobs to afford the crap they're pushing.

Sooner or later, it will become understood that technology must exist to assist us; attempting to replace us entirely with it are non-optimal.

Why Transformer Models Aren't a Path to AGI

A common misconception is that the sophistication of transformer-based models puts us on a direct path to Artificial General Intelligence (AGI). However, this is far from the truth. AGI requires an ability to understand, reason, and generalize across a broad spectrum of domains, akin to human intelligence. Current language models are fundamentally incapable of such feats. They don't learn or understand in a human sense; they merely rearrange existing data to create seemingly new content.

The perceived reasoning and inference abilities of these models are, in reality, illusory. This phenomenon is similar to humans finding recognizable shapes in clouds; we are predisposed to attribute meaning and intention where none exists. When a language model produces a coherent response, it's not due to understanding or reasoning. Instead, it's the result of its training data containing similar language patterns that can be regurgitated in new forms.

AGI will require a different system; not Von-Neumann, and certainly not "quantum"

As "2D" transformer-based AI runs nicely on conventional computing devices, AGI on the other hand will operate within a completely different paradigm.

In a nutshell, they will dynamically operate on massively parallel streams of "stimuli", "detecting and storing" sensory stimuli representing visual, audible, and environmental conditions from "external" sources, updating internal "state maps", which themselves are compared to "memories" to support "orientation" and general status, but where the real rubber meets the road happens in reflection interfaces, themselves having access to all memory streams, immediate (sensory stimuli), accumulated (short term), and symbolic (long term), all of which are processed within these constructs, where "collisions" between immediate and remembered stimuli cause the rise of a primitive "consciousness".

And the platform capable of running such a topology does not yet exist, since Von-Neumann architecture simply cannot emulate these processes. Our typical computing devices work fine in the paradigm of Human-abstracted ideas about information, where we believe process and data to be necessarily distinct from one another, as some "operation" must affect the properties of "some datum", but nature simply doesn't operate this way, and although some AGI processes may be emulated on a digital machine, they will never run with adequate performance, and would be impossible to scale,

And this is because any "AGI" system hoping to emulate even the slightest hint of "consciousness" cannot operate within an information domain constrained by the artificial separation of "information" from "process", as "natural processors" (brains) derive their functionality from nature, which has no use for data whatsoever, as nature realizes that the "process is the data", so in a biological processor, the "network is the information".

"True AGI" will be realized only after researchers define reasonable boundaries for "artificial consciousness", and this will only become viable after they realize the axiom that "dynamic reflection interfaces" are necessary to emulate any kind of "consciousness", artificial or not.

We're safe for quite a while yet.

Conclusion

While transformer-based language models represent a significant advancement in natural language processing and have wide-ranging practical applications, they are not harbingers of AGI. They lack the fundamental capability of genuine understanding and interaction with the world, because they lack any system capable of supporting dynamic stimulus integration and comparison utilized in biological processors, rendering the most advanced transformer to be similar in function to, but far less intelligent than a biological virus.

Recognizing this distinction is crucial for both leveraging the strengths of these models and understanding their inherent limitations. They are not just playing a different game in the field of intelligence; they are in an entirely different league, one that's far removed from the nuances of human cognition and understanding.