From: jimruttshow8596
The role of Artificial General Intelligence (AGI) has seen rapid, often unpredictable, advancements in the AI space, with changes occurring at an accelerating pace. This pace is compared to the personal computer revolution of the late 1970s and early 1980s, but happening “10 times faster” [01:25:00]. While the development isn’t uniform across all domains—for instance, automatic supermarket checkouts still struggle—advanced AI technology in other areas is accelerating rapidly [02:17:00].
Large Language Models (LLMs) and AGI: A Nuanced Perspective
Ben Goertzel, a leading authority on AGI, posits that large language models (LLMs) in their current form—Transformer networks trained to predict the next token and then fine-tuned—will not lead to a full human-level AGI [04:49:18]. However, he acknowledges that LLMs can perform many amazing and useful functions [05:01:00] and can serve as valuable components within systems capable of achieving AGI [05:10:00].
The discussion often revolves around whether LLMs should be the central “hub” of an AGI architecture or play a supporting role [05:21:00]. For example, OpenAI’s approach involves an AGI architecture with multiple LLMs as an integration hub, alongside non-LLM systems like DALL-E or Wolfram Alpha [05:29:00]. In contrast, Goertzel’s own approach with the OpenCog Hyperon project positions a weighted labeled metagraph as the central hub, with LLMs and other specialized tools interacting on the periphery [05:59:00].
This leads to a fine-grained distinction: “LLM++” (LLMs plus external components) may not lead to human-level AGI, but “something + LLM” (where LLMs support a different core architecture) could accelerate the path to AGI [08:01:00]. The AGI field is polarized between LLM boosters and detractors [08:22:00], despite the fact that LLM limitations are constantly evolving with new tools and releases [08:45:00].
Specific Limitations of LLMs and Attributes of Human Intelligence
While LLMs have shown remarkable progress, fundamental limitations persist:
The Hallucination Problem
LLMs are known to “hallucinate,” meaning they generate factually incorrect or nonsensical information, especially when dealing with obscure queries [09:39:00]. For example, some models can incorrectly generate biographies or lists of famous podcast guests [09:58:00].
While this problem can be mitigated by technical means, such as probing the network to detect hallucination indicators [11:13:00] or running queries multiple times and analyzing entropy [13:52:00], this does not address the underlying cognitive difference between LLMs and human intelligence. Humans avoid hallucinations through a “reality discrimination function” and “reflective self-modeling” [12:08:00]. LLMs, even if engineered to filter out hallucinations, lack this internal understanding or self-awareness [12:40:00].
Defining AGI and Human-Level Intelligence
There is no universal agreement on what constitutes AGI [21:33:00]. Theoretically, AGI can be defined as the ability to achieve a wide variety of computable goals in diverse computable environments [21:57:07]. However, humans are “retards” by this measure, being poor at optimizing arbitrary reward functions in complex environments [23:22:00].
Other conceptualizations of intelligence include:
- Weaver’s Theory of Open-Ended Intelligence: Focuses on complex self-organizing systems that individuate, maintain existence, and self-transform [24:15:00].
- Human-Level General Intelligence: Evaluates what humans are naturally good at, beyond just IQ tests, which are considered “shitty” and do not fully capture human capability [25:34:00]. Multifactorial approaches like Gardner’s multiple intelligences theory (musical, literary, physical, existential, logical) come closer to capturing human intelligence [25:51:00].
The Turing Test, which assesses a machine’s ability to imitate a human in conversation, is deemed an inadequate measure of AGI [26:31:00]. The fact that LLMs are “knocking on the door” of passing it with methods clearly not reflective of AGI makes its relevance questionable [27:11:00].
Lack of Multi-Step Reasoning and Original Creativity
Two key limitations of current LLMs, which humans excel at, are:
- Complex Multi-Step Reasoning: Essential for original scientific research [30:07:00].
- Original Artistic Creativity: Needed for composing new music genres or truly novel art [30:14:00].
While LLMs can produce output that is as good as a “first draft created by a professional journeyman screenwriter” [34:57:00] or “damn good blues guitar solo” [35:38:00], they generally default to “banality”—an average of existing utterances [34:14:00]. Clever prompting can push them beyond this banality, but they still struggle to match the systematic breakthroughs of great human creatives [34:39:00].
In science, LLMs can “turn the crank” on advanced math and flesh out theories from original ideas, performing tasks suitable for a master’s student [38:48:00]. However, they lack the “deep judgment” [39:17:00] or “math aesthetics” needed to generate fundamentally surprising scientific theories or to discern productive research directions from dead ends [40:40:07].
These shortcomings stem from LLMs’ architecture, which primarily recognizes surface-level patterns in data. They do not show evidence of learning abstractions in the same way humans do, nor do they possess the “fundamentally derivative and imitative character” needed for true originality [33:18:00]. While LLMs can generalize remarkably well in some cases, it’s argued not to be sufficient for human-level AGI [33:46:00].
Towards Solving LLM Limitations for AGI: Architectural Directions
To overcome the limitations of banality, lack of multi-step reasoning, and non-banal creativity, several architectural directions are proposed:
1. Enhanced Recurrence in Neural Networks
One path involves introducing more recurrence into Transformer networks [46:46:00]. While Transformers reduced recurrence for scalability, recurrence is crucial for generating interesting abstractions [47:08:00]. This could involve replacing attention heads with more sophisticated mechanisms [47:20:00] or exploring alternatives to backpropagation for training, such as predictive coding-based training [47:36:00].
2. Hybrid Architectures
Integrating LLMs with other advanced neural net components is a promising direction. For instance, combining systems like AlphaZero (known for planning and strategic thinking) with neural knowledge graphs (as in Differential Neural Computing) and recurrent Transformers [48:17:00]. This approach leverages the strengths of different paradigms. DeepMind, with its expertise in DNC, AlphaZero, and evolutionary learning, is well-suited to pursue such hybrid models [48:44:00]. Similarly, some groups are exploring neural nets that explicitly learn abstractions through minimum description length learning, coupled with Transformers [49:31:00].
3. Evolutionary Algorithms and Radical Recurrence
There is a call for more research into floating point evolutionary algorithms for evolving neural networks [51:13:00]. This method, along with predictive coding, could work better for richly recurrent networks than backpropagation [52:20:00].
4. The OpenCog Hyperon Approach
Ben Goertzel’s project, OpenCog Hyperon, focuses on a self-modifying, self-rewriting weighted labeled metagraph as the core of AGI [53:51:00].
- Metagraph Structure: This is a graph where links can expand multiple nodes, point to other links or subgraphs, and are typed and weighted [54:41:00].
- Knowledge Representation: It aims to represent various types of knowledge (declarative, procedural, sensory) and cognitive operations (reinforcement learning, logical reasoning, pattern recognition) within this hypergraph [55:02:00].
- Meta-Programming: A new programming language, Meta-Meta, is being developed where programs themselves are sub-metagraphs that can act on, transform, and rewrite chunks of the same metagraph in which they exist [55:33:00]. This allows for inherent reflection and self-modification [56:54:00].
- Role of LLMs: LLMs are viewed as supporting actors that can exist on the periphery, helping the system [58:37:00], but not as the central hub.
- Advantages: This architecture is naturally suited for logical reasoning, precise description of repeatable procedures, and evolutionary programming (genetic programming) for creativity [59:51:00].
- Scalability Challenges: The main challenge is the scalability of infrastructure, similar to how deep neural nets required significant hardware to achieve their current capabilities [01:00:43]. Hyperon is building a pipeline to translate its native language into highly efficient math manipulations on specialized hardware beyond GPUs, including Associate Processing Units (APUs) [01:02:05]. The project also aims for decentralization, with components like a secure blockchain-based atom space module [01:07:01].
This approach is considered “least humanlike” among AGI pursuits, but offers a “really short” path from human-level AGI to superhuman AI, because the system is designed to rewrite its own code [01:00:00].
While LLMs accelerate progress in AI, their fundamental limitations suggest that a more integrated, hybrid approach — potentially with a different core architecture than an LLM — will be necessary to achieve true human-level AGI and beyond.