From: jimruttshow8596
Current artificial intelligence (AI) architectures, particularly deep neural networks (DNNs) and other machine learning (ML) algorithms, are considered by some leading authorities to be fundamentally unsuited for the creation of human-level Artificial General Intelligence (AGI) [01:10:07]. While these approaches have achieved significant practical success in narrow AI applications [00:02:35], they exhibit specific limitations that prevent them from achieving true generality [01:16:17].
Defining AGI vs. Narrow AI
AGI is an imprecise and informal term referring to computer systems that can perform tasks considered intelligent when done by people, including those they weren’t specifically programmed or trained for [00:01:52]. In contrast, narrow AI excels at highly particular tasks based on programming or data-driven training [00:02:40]. Humans, for example, can make leaps into domains only loosely connected with prior experience [00:02:55], a capability largely absent in current narrow AI.
Examples of Narrow AI Limitations
- AlphaFold: An impressive example of narrow AI that predicts protein folding based on training data [00:03:54]. However, it struggles with “floppy proteins” or new molecular classes from, for instance, an alien planet [00:04:06]. To deal with these, it requires manual feeding of more training data or algorithm changes; it cannot make the generalization leap on its own [00:04:26].
- Self-Driving Cars: While perhaps not “AGI-hard,” self-driving cars demonstrate the generalization problem because of the endless variety of “weird left turn situations” or other unexpected road events [00:07:38]. The training data sets don’t allow them to leap to deal with all such scenarios [00:08:00].
- The “Coffee Test”: A simple-minded example of AGI involves a robot placed in a random kitchen tasked with making coffee [00:06:28]. An average human could do this, but no current robot or AI can even begin to solve it [00:06:38].
The Shallow Pattern Problem of Deep Neural Networks
A core limitation of current DNNs is that they behave largely like “very clever lookup tables” [01:16:53]. They record and store every detailed pattern they see, accounting for overlap, usefulness, and weightings in different contexts [01:17:01]. However, despite being called “deep,” these networks primarily look at “shallow patterns” in datasets [01:17:47].
- Lack of World Model: In natural language processing, DNNs focus on sequences of words rather than trying to build a model of the conceived world underlying those words [01:18:00].
- The Table Saw Example: When asked how to fit a large table through a small door, a transformer neural network suggested using a “table saw” [01:18:23]. It mistakenly assumed a “table saw” was for sawing tables in half, failing to understand its physical function despite having read manuals explaining it in its training data [01:19:00]. This illustrates a failure to build a model of reality underlying the text [01:20:00].
- Knowledge Representation Issue: Current systems leverage vast data and processing power to recognize highly particular patterns and extrapolate from them [01:20:59]. This approach struggles to generalize to domains of reality that do not exhibit those specific patterns [01:21:17]. The knowledge is represented as a large catalog of weighted particulars, with “no attempt to abstract” [01:21:47].
The Crux: Lack of Abstraction and Generalization
The ability to find concise abstractions of experience is equivalent to the raw ability to generalize to different domains [01:21:55]. This is the crucial missing element in current DNNs [01:22:09]. Humans, even with small datasets (e.g., a few thousand war game sessions), can pull out broad generalizations and apply them to new, distinct scenarios [01:23:08]. This capability, often referred to as “one-shot learning” or “few-shot learning,” remains a challenge for current architectures [01:25:55].
Limitations in Creativity and Imagination
Current AI architectures tend to bypass the aspects of human intelligence that allow for “creative imaginative leaps” [01:27:12]. While AI can re-permute elements from existing images (like DALL-E) [01:29:09], they do not innovate like artists such as Matisse or Picasso, who fundamentally rethought art [01:29:14]. This limitation stems from the economic and commercial motivations driving AI development.
Commercial Biases
The AI industry has largely organized itself to deploy AI in ways that extremely leverage DNNs’ strengths [01:28:45]. To be a good employee of a company with a well-defined business model, an AI needs to “repeat some well understood operations in a predictable way to maximize well-defined metrics” [01:27:58]. This includes tasks like predicting ad clicks [01:28:20] or generating graphics that combine recognizable visual tropes with known measurable impact [01:29:36]. In these contexts, improvising and imagining are often “beside the point” [01:28:36]. As a result, AGI research, which seeks unpredictable and creative intelligence, often remains on the margins of the AI field [01:31:20].
Historical Context and Future Integration
The term “good old-fashioned AI” (GOFAI) often refers to earlier approaches that relied on explicitly typing knowledge into systems or using crisp logic [01:41:20]. While the idea of hand-coding all knowledge is now considered ridiculous [01:43:17], certain aspects of GOFAI, such as logic-based knowledge representation (with uncertainty incorporated), can be combined with learning algorithms and low-level perception [01:43:55].
Interestingly, neural networks themselves are “most good old-fashioned AI there is,” with concepts dating back to the 1940s and 50s [01:45:38]. Their recent success is attributed to the availability of “enough data and big enough computers” [01:46:00]. It is hypothesized that almost every idea from the history of AI could eventually contribute to AGI systems when deployed with the right hardware, data, and integration [01:46:09]. This suggests that current DNNs may be just the first type of “good old-fashioned AI” to achieve prominence with the advent of copious processing and data [01:46:36]. Future AGI might involve integrating various approaches, including logic, evolutionary learning, and agent-based systems [01:46:43].