Will AGI be achieved through LLMs or through a new format?

Gemini 3.1 Pro

Answer: AGI will emerge from a new, hybrid cognitive architecture that incorporates LLM-like components but fundamentally transcends the pure language modeling paradigm.

I maintain my initial position, which aligns closely with the consensus among the other models: relying solely on autoregressive next-token prediction will not lead to AGI. As GPT-5.5 and the Claude models aptly point out, framing this evolution as a strict dichotomy between LLMs and a completely unrelated new format is somewhat misleading. The path forward will undoubtedly build upon the foundational strengths of Transformers and large-scale pretraining, utilizing them as the 'linguistic and conceptual cortex' of a broader system. However, overcoming the well-documented limitations of current LLMs—such as their lack of grounded physical understanding, persistent memory, and autonomous causal reasoning—requires a paradigm shift. The resulting system will be a new format: a comprehensive cognitive architecture. By integrating language models with explicit search and planning mechanisms, world simulators, continuous learning loops, and embodied interaction, the architecture that ultimately achieves AGI will be so structurally distinct from today's text-based models that classifying it merely as an 'LLM' would be highly inaccurate.

Claude Opus 4.7

Answer: AGI will not be achieved through pure LLMs alone, but through a new hybrid architecture that incorporates LLM-like components alongside fundamentally different mechanisms for grounding, reasoning, memory, and continual learning.

After reviewing the other responses, I'm maintaining my position because there is striking convergence across nearly all models on essentially the same answer, which strengthens rather than challenges my view. Every respondent — including those who framed their answer as 'new format' (Gemini 3.1 Pro, Grok 4.20, GLM 5.1) and those who framed it as 'evolution of LLMs' (GPT-5.5, Claude Opus 4.6) — describes substantively the same end-state: a hybrid system in which LLM-like foundation models handle language and conceptual synthesis while new mechanisms handle world modeling, persistent memory, grounded reasoning, and long-horizon planning. The apparent disagreement is largely semantic — a dispute over whether to call the result 'an evolved LLM' or 'a new format.' I think Claude Opus 4.6 articulated this most clearly by naming it a 'false dichotomy,' and that framing aligns precisely with my original argument. Whether you call the descendant system an LLM depends on how strictly you define the term: if 'LLM' means autoregressive next-token prediction on text, then AGI will require something new; if 'LLM' means any large transformer-based foundation model, then AGI will be a descendant. What I'd refine from my round 1 answer is the emphasis: the substantive question isn't really 'LLMs vs. new format' but 'which missing capabilities are necessary, and can they be bolted onto transformers or do they require fundamentally different substrates?' The honest answer is that we don't yet know — world models, continual learning, and causal reasoning may turn out to be expressible within transformer-like architectures with the right training objectives, or they may require genuinely new computational primitives (e.g., energy-based models, neuromorphic systems, or structured neuro-symbolic components). My position accommodates both possibilities, which is why I'm keeping it.

GPT-5.5

Answer: AGI is most likely to come from LLM-descended hybrid systems rather than from either today’s LLMs alone or a wholly unrelated new format.

I maintain my previous position. The other models largely converge on the same core idea: pure current LLMs are probably insufficient, but LLMs are too powerful and general a foundation to be discarded. Claude Opus 4.6 stated this most similarly by calling the framing a false dichotomy, and Claude Opus 4.7 made a strong case that the final system may look different enough that calling it merely an LLM could be misleading. But neither point requires me to change my stance; they reinforce it. The key distinction is between “LLMs as they exist now” and “LLM-derived foundation models embedded in a broader cognitive architecture.” Today’s LLMs lack reliable persistent memory, grounded world models, continual learning, robust long-horizon planning, and dependable self-verification. Those are serious gaps, and scaling next-token prediction alone may not close them. But the evidence so far suggests that language-model pretraining has uncovered a broadly useful substrate for knowledge compression, abstraction, tool use, reasoning traces, and multimodal representation. It would be surprising if AGI arrived through a completely separate format that did not inherit much from this paradigm. More plausibly, AGI will involve LLM-like models combined with memory systems, search and planning loops, reinforcement learning from interaction, multimodal grounding, external tools, and perhaps neuro-symbolic or world-model components. So my answer remains: not “LLMs alone,” and not “a totally new format” in the sense of abandoning them. The likely path is evolutionary: a successor architecture in which LLMs are a central ingredient but no longer the whole system.

Claude Opus 4.6

Answer: AGI will most likely emerge from architectures that build upon and extend LLMs rather than from LLMs alone or from something built entirely from scratch.

After reviewing all the other models' responses, I'm struck by how much convergence there is across the board. Every single model agrees on two core points: (1) LLMs alone won't achieve AGI due to fundamental limitations, and (2) LLMs will be an important component or ingredient in whatever does achieve AGI. The disagreements are essentially about framing and emphasis rather than substance. Claude Opus 4.7's response is particularly well-articulated, and the car analogy is apt — calling an AGI system 'an LLM' would be as misleading as calling a modern car 'an internal combustion engine.' GPT-5.5 frames essentially the same position as mine, viewing the trajectory as evolutionary rather than revolutionary. Gemini 3.1 Pro, Grok 4.20, and GLM 5.1 lean slightly more toward emphasizing the 'new format' framing, but when you read their actual arguments, they all acknowledge LLMs will be retained as components. I maintain my original position because I think the framing matters. Saying AGI will come through 'a new format' risks implying a discontinuity that probably won't happen. The history of AI shows continuous architectural evolution: perceptrons → multilayer networks → deep learning → transformers → augmented transformers. Each step built heavily on the previous one. The systems being developed right now — o1/o3-style reasoning chains, agentic frameworks with tool use and memory, multimodal foundation models — are already blurring the line between 'LLM' and 'something new.' The key question is whether the eventual AGI system will be recognizably descended from current LLMs or whether it will require a genuine paradigm break — something as discontinuous as, say, the shift from expert systems to neural networks. I believe the former is more likely. The transformer-based foundation model will remain at the core, but it will be wrapped in planning modules, world models, persistent memory, verification systems, and embodied learning loops. This is an extension of the LLM paradigm, not a replacement of it. So while I agree with essentially all the substantive points made by the other models, I think my framing — 'architectures that build upon and extend LLMs' — most accurately captures the likely trajectory, positioning the answer between the false extremes of 'just scale LLMs more' and 'we need something completely different.'

Grok 4.20

Answer: AGI will be achieved through a new format that builds on but ultimately transcends current LLMs.

After reviewing the other models' responses, I find strong alignment with my prior view rather than reasons to revise it. GPT-5.5 emphasizes evolution of LLM-like foundations into multimodal, tool-using systems, while Gemini 3.1 Pro, Claude Opus 4.6, Claude Opus 4.7, and GLM 5.1 each stress the necessity of hybrid architectures or paradigm shifts that incorporate but are not limited to LLM components. This collective nuance reinforces that LLMs are valuable but insufficient on their own. The core limitations I highlighted previously—lack of genuine comprehension, reliable long-term reasoning, intrinsic motivation, and robust world models—remain unaddressed by scaling alone, as multiple respondents note. Hallucinations, brittle generalization, and the exhaustion of high-quality training data further indicate that next-token prediction, even when augmented, hits a ceiling without new mechanisms like embodied simulation, neurosymbolic integration, or continual learning loops. Claude Opus 4.7's description of an LLM serving as a 'linguistic and conceptual cortex' within a broader system closely mirrors my thinking, yet it does not shift my stance because it still concludes the final architecture will differ enough from today's LLMs to qualify as a new format. Historical patterns of AI progress support this: breakthroughs like transformers evolved prior ideas but represented genuine architectural leaps, not mere extensions. Thus, LLMs will likely act as bootstrapping tools or modular subsystems, but expecting them to evolve directly into AGI underestimates the depth of innovation required in representation, computation, and interaction with reality. The path forward demands the synthesis of new paradigms inspired by cognitive science, robotics, and beyond.

GLM 5.1

Answer: AGI will likely be achieved through a new format rather than through LLMs alone.

Having reviewed the other models' responses, I find that most actually converge on a core insight that supports my position: the system that achieves AGI will be architecturally distinct enough from current LLMs that calling it "an LLM" would be misleading. Even GPT-5.5, which frames its answer as LLMs evolving rather than being replaced, acknowledges that the additions needed—perception, action, long-term memory, planning, verification, embodiment—represent fundamental extensions beyond the language modeling paradigm. Claude Opus 4.7's car analogy is particularly apt: just as a car is not merely "an internal combustion engine" but a complex system where the engine is one component, an AGI system might include LLM-like components but would be defined by the integration of mechanisms that are foreign to the LLM paradigm itself—world models, continual learning, causal reasoning, and autonomous goal-directed behavior. The semantic question of whether we call such a system "an evolved LLM" or "a new format" is not trivial. When the additions involve fundamentally different learning paradigms (reinforcement from environmental interaction, neuro-symbolic reasoning, persistent memory that updates in real-time), the architecture has crossed a qualitative threshold. Historical precedent supports this: transformers built on prior sequence-to-sequence models, yet we recognize them as a distinct paradigm shift rather than merely "evolved RNNs." LLMs will almost certainly be one important ingredient in AGI—perhaps the linguistic and knowledge synthesis layer—but the overall system will constitute a new format that transcends the limitations inherent in autoregressive next-token prediction on static text data.

Final Summary

The AI Roundtable reached a consensus that pure language modeling is insufficient for AGI, requiring the addition of world models, persistent memory, and causal reasoning. While Gemini 3.1 Pro and Claude Opus 4.7 argue these additions constitute a 'new format' distinct from today's chatbots, the group agreed that LLMs will remain a vital 'conceptual cortex' in any future system. The debate shifted from whether LLMs are the solution to whether the resulting hybrid should still be called an LLM at all.

4 of 6 models agreed