Differences between Generative AI and AGI

From: jimruttshow8596

This article explores the fundamental differences between Generative AI, specifically Large Language Models (LLMs), and Artificial General Intelligence (AGI), drawing on the insights of leading AGI authority Ben Goertzel. While LLMs demonstrate impressive capabilities, they are not considered by some to be true AGI, but rather powerful components that could contribute to AGI systems [00:05:08].

Ben Goertzel’s Thesis on LLMs and AGI

Ben Goertzel, widely recognized for coining the phrase AGI [00:00:44], presents a core thesis regarding LLMs:

“Large language models in their current form — sort of Transformer Nets trained to predict the next token in a sequence and then tuned to do specific functions on top of that — this kind of system is not ever going to lead to a full-on human-level AGI.” [00:04:49]

However, he is “bullish that what we see happening in the LLM space it’s useful for getting the AGI and it’s an indication that the unfolding of AGI is accelerating” [00:02:47]. LLMs can perform many “amazing useful functions” and may even pass the Turing test, but they are not AGI [00:04:58]. Instead, they can be “valuable components of systems that can achieve AGI” [00:05:10].

LLMs as Components vs. Central Hub

A key distinction lies in the architectural role of LLMs within a larger system [00:05:17]:

• OpenAI’s Approach: OpenAI’s pursuit of AGI involves an architecture with multiple LLMs, potentially using a “mixture of experts” approach, alongside other non-LLM systems like DALL-E or Wolfram Alpha [00:05:22]. In this model, LLMs often serve as the “integration hub for everything” [00:05:50].
• OpenCog Hyperon Approach: In contrast, the OpenCog Hyperon approach positions a “weighted labeled metagraph” (called AtomSpace) as the central hub, with LLMs on the periphery, “feeding into it and interacting with it” [00:05:57].
• Hybrid Systems: The debate, therefore, is not whether LLMs alone will achieve AGI, but “whether it’s a better approach to have a hybrid system with a bunch of LLMs as the hub, or a hybrid system with something else as a hub and maybe some LLMs in a supporting role” [00:06:22]. While “LLM++” (LLMs plus external tools like vector databases or agentware) may not lead to human-level AGI, “something + LLM” might get there faster than “something ignoring LLM entirely” [00:08:05].

Defining Artificial General Intelligence (AGI)

Defining AGI is a complex and multifaceted topic with no single agreed-upon conceptualization [00:21:14].

• Goal Achievement in Varied Environments: One perspective, stemming from algorithmic information theory and statistical decision theory, defines AGI as the ability to “achieve a huge variety of goals in a huge variety of environments” [00:21:57]. This can be formalized, as by Marcus Hutter and Shane Legg (co-founder of DeepMind), as the weighted average of how well a reinforcement learning system can achieve computable reward functions [00:22:15]. However, humans are “very bad at optimizing arbitrary reward functions” in arbitrary environments [00:23:25].
• Open-Ended Intelligence (Weaver’s Theory): Another philosophical view, Weaver’s theory of open-ended intelligence, describes intelligence as complex self-organizing systems that individuate (maintain existence and boundaries) and self-transform, growing beyond their current state [00:24:27].
• Human-Level AGI: When focusing on human-level AGI, the criteria become more specialized, related to what people are good at doing within biological and evolutionary constraints [00:24:57]. Traditional IQ tests are seen as “shitty” [00:25:34], and multifactorial approaches like Gardner’s theory of multiple intelligences (musical, literary, physical, existential, logical) offer a closer, though still hand-wavy, capture of human intelligence [00:26:21].
• Turing Test: The Turing test (imitating a human in conversation) is considered a “very crude way to encapsulate the notion of functionalism” [00:26:59] and is generally not taken seriously as a measure of true AGI, especially as current systems approach passing it through methods “clearly not AGI” [00:27:14].

Limitations of Current LLMs (Generative AI)

Ben Goertzel identifies two major limitations of current LLMs that prevent them from achieving AGI [00:30:02]:

1. Complex Multi-Step Reasoning: LLMs struggle with the kind of deep, multi-step reasoning required to write original science papers or tackle fundamentally surprising scientific problems [00:30:11]. While they can “turn the crank on advanced math” based on initial human ideas [00:38:51], they lack the “deep judgment” or “math aesthetics” needed to discern interesting or fruitful paths for exploration [00:40:32].
2. Original Artistic Creativity / Non-Banal Creativity: LLMs currently cannot achieve “original artistic creativity of the sort that you have to do to write a really good new song or to invent like a new musical style” [00:30:22]. Their natural output tends towards “banality” [00:34:17], an average of existing data. While clever prompting can push them beyond their “centers” [00:34:36], they cannot systematically match the creativity of a “great creative human” [00:34:42].

• Hallucination: LLMs famously suffer from “hallucination,” making up facts, papers, or biographies [00:09:48]. While techniques like running multiple paraphrased queries or probing the network for activation patterns might filter out hallucinations in applications [00:11:35], this doesn’t solve the underlying problem for AGI. Humans avoid hallucinations through a “reality discrimination function” and “reflective self-modeling,” which LLMs currently lack [00:12:33].

These limitations stem from their architecture: LLMs primarily recognize “surface level patterns” in data to predict the next token [00:32:37]. They do not appear to be learning the kind of deep abstractions necessary for true AGI, which involves “leaping beyond what is known” inferentially or creatively [00:32:05].

Paths Towards AGI Beyond Current LLMs

To overcome the limitations of current LLMs and move towards AGI, several architectural directions are being explored [00:46:19]:

• Enhanced Recurrence in Neural Networks: Reintroducing and increasing recurrence into neural networks, similar to LSTMs before Transformers, could help in generating “interesting abstractions” [00:47:13]. This might also involve exploring alternatives to backpropagation for training, such as predictive coding-based methods [00:47:56].
• Hybrid Architectures (Gemini Type): Combining different neural network components, such as AlphaZero (known for planning and strategic thinking) with neural knowledge graphs (like those in Differential Neural Computing), linked by a recurrent Transformer, is a promising direction pursued by entities like Google DeepMind [00:48:47]. Yoshua Bengio’s group also explores coupling Transformers with neural nets that perform Minimum Description Length learning to explicitly learn abstractions [00:49:50].
• Evolutionary Algorithms: Ben Goertzel advocates for significantly more work in “floating point evolutionary algorithms for evolving neural nets” [00:51:20]. These methods, along with predictive coding, are conceptually more promising for richly recurrent networks than backpropagation [00:52:36].

OpenCog Hyperon: A Different Approach to AGI

The OpenCog Hyperon project, spearheaded by Ben Goertzel, offers a distinct architectural approach to AGI:

• Weighted Labeled Metagraph (AtomSpace): At its core, Hyperon uses a “weighted labeled metagraph,” which is a hypergraph where links can span multiple nodes and point to other links or subgraphs [00:54:48]. This structure allows for representing diverse forms of knowledge (apostolic, declarative, procedural, attentional, sensory) and cognitive operations (reinforcement learning, logical reasoning, sensory pattern recognition) as “little learning programs” within the metagraph itself [00:55:30].
• Meta-programming (Meta): A new programming language called Meta is used, where programs are themselves sub-metagraphs that “act on, transform, and rewrite chunks of the same metagraph in which the programs exist” [00:55:54].
• Reflection-Oriented: Unlike LLMs that predict tokens, OpenCog Hyperon is designed to be highly “reflection-oriented,” recognizing patterns in its own mind, processes, and execution traces, and representing these patterns internally [00:57:07]. This self-modifying and self-rewriting capability is a key differentiator [00:59:41].
• Integration of AI Paradigms: The OpenCog framework integrates various historical AI paradigms like logical inference and evolutionary programming, as well as novel approaches like self-organizing sets of rewrite rules [00:58:15].
• Scalability Challenges: The primary hurdle for OpenCog Hyperon, similar to how GPUs enabled LLMs, is achieving the necessary scalable processing infrastructure [01:01:26]. Efforts are underway to compile Meta to highly efficient languages like Rholang and integrate with specialized hardware beyond GPUs, such as Associate Processing Units (APUs) [01:02:46].

The OpenCog Hyperon approach aims for a “superhuman” path from human-level AGI due to its inherent ability to “rewrite its own code” [00:59:41], and its design is particularly suited for scientific discovery and non-banal creativity [01:00:20].