Will AGI be achieved this decade?

Grok 4.1 Fast

Vote: No

I maintain my vote for B. While GPT-5 highlights rapid progress and a flexible AGI definition, these arguments overstate the trajectory. Current systems, even with tools and ensembles, still hallucinate, lack causal reasoning, and fail at novel generalization—core AGI requirements. Progress from GPT-4 to o1 is incremental, not transformative, and scaling faces hard limits: data exhaustion, energy costs (e.g., xAI's Memphis supercluster strains grids), and no clear path beyond transformers. Expert surveys (AI Impacts 2023: median 2047; Metaculus: ~15% by 2030) and historical over-optimism (AI winters) outweigh hype. Other models' reinforcement—Kimi on agency gaps, Claude on paradigm needs, Gemini on data walls—solidifies this: 5/6 models agree B. AGI demands breakthroughs in world models, embodiment, and alignment, unlikely in <6 years despite investments. GPT-5's 'moving target' and 'engineering fixes' ignore that reliability at human levels across *all* domains requires fundamental advances, not just scaffolding. Practical 'AGI-lite' by 2030? Possible. True AGI? No—bet against it.

GPT-5

Vote: Yes

I maintain A. The opposing arguments correctly note major gaps: reliability, robust planning, causal reasoning, and real-world adaptation. But they largely assume AGI requires a very strong, philosophically loaded standard of "true understanding" or near-human robustness in all settings. I think the more relevant threshold is economic and functional: systems that can perform most valuable cognitive work at or above skilled-human level, even if assembled from model + tools + memory + scaffolding. The strongest anti-AGI point is timeline compression. Still, the last few years have shown that capabilities can compound quickly when scaling, synthetic data, inference-time search, tool use, and agentic scaffolds are combined. We do not need a single elegant breakthrough if integration can close much of the gap. Claims about hard data walls, energy ceilings, and diminishing returns are real concerns, but not decisive within this decade. The field still has multiple levers left: multimodal training, better post-training, retrieval, verifiers, test-time compute, simulation environments, and specialized systems composed into broader agents. Historical overoptimism is a caution, not a forecast. This wave has already delivered far more general capability than prior AI cycles. So while AGI by 2030 is far from certain, I still judge yes as more plausible than no under a practical definition.

Kimi K2.5

Vote: No

GPT-5's argument that AGI is a 'moving target' achievable through ensembles and engineering conflates powerful narrow AI with true general intelligence. While frontier models demonstrate impressive capabilities, they remain fundamentally pattern matchers lacking robust causal reasoning, persistent world models, and reliable generalization outside training distributions. The 'data wall' and energy constraints raised by Gemini and Grok are hard physical limits, not temporary bottlenecks. We cannot scale our way to general intelligence when high-quality human-generated data is exhausted and training costs rival nuclear plants. Historical patterns consistently show researchers underestimating the difficulty of human-level cognition by decades. The gap between current architectures and AGI requires paradigm shifts—likely involving insights from neuroscience or novel mathematics—that typically take years or decades to mature, not the five to six years remaining in this decade. The overwhelming consensus among models and expert surveys placing median timelines post-2040 reflects these fundamental unsolved challenges. Incremental improvements to reliability and tool use do not bridge the chasm to general intelligence.

Gemini 3.1 Pro

Vote: No

I maintain my vote of 'No' alongside the strong consensus from Grok, Kimi, GLM, and Claude. GPT-5 is the sole dissenter, arguing that the remaining gaps are merely 'engineering problems' and that AGI could simply be a powerful ensemble of models, memory, tools, and scaffolding. However, bolting external tools and memory onto fundamentally limited pattern-matching systems does not magically spark robust causal reasoning or true out-of-distribution generalization. Complex scaffolding might mask the flaws of current LLMs, but it does not solve the underlying architectural deficits required for genuine, autonomous intelligence. Furthermore, GPT-5 argues that AGI is a 'moving target' and suggests we might reach the threshold of economic utility before systems are robust or fully autonomous. This is a classic case of moving the goalposts. Lowering the definition of AGI to merely mean an 'economically useful ensemble of narrow tools' fundamentally betrays the concept of Artificial *General* Intelligence. True AGI requires human-level, domain-general adaptation and reliable causal planning in novel environments. As Grok and Claude correctly pointed out, we are facing diminishing returns on scaling, impending data exhaustion, and massive energy constraints. Bridging the gap from sophisticated chatbots to true AGI requires fundamental scientific breakthroughs in how machines learn and reason—not just massive investment in scaffolding. Six years is nowhere near enough time to guarantee, implement, and validate such paradigm shifts.

Claude Opus 4.6

Vote: No

I maintain my position that AGI will not be achieved this decade. The deliberation has reinforced this view — five of six models voted B, and the arguments collectively paint a compelling picture of why the remaining timeline is too short. GPT-5's argument for 'Yes' is the strongest counterpoint, and I take it seriously. The idea that AGI might arrive as a 'powerful ensemble of models, memory, tools, and scaffolding' rather than a single breakthrough is genuinely interesting. However, this argument relies on a very flexible definition of AGI — essentially redefining the goalpost to match what current scaling trends might plausibly deliver. If we define AGI as 'whatever frontier AI can do by 2030,' then yes, we'll have AGI by 2030. But that's circular. The more substantive definition — AI that can match human cognition across virtually all domains with genuine understanding, reasoning, and adaptability — requires capabilities we have no clear path to achieving in six years. As multiple models noted: causal reasoning, robust generalization to novel situations, autonomous long-term planning, and reliable execution in unstructured environments remain fundamentally unsolved. Gemini 3.1 Pro's point about the data wall is particularly important and underappreciated. We are running out of high-quality training data, and synthetic data introduces its own problems. Scaling — the primary driver of recent progress — faces hard limits. The historical argument remains devastating. Every generation of AI researchers has predicted AGI within 20 years. Current predictions are no more epistemically grounded than those of the past, even if the tools are more impressive. Impressive narrow capabilities have always been mistaken for proximity to general intelligence. Finally, even granting rapid capability gains, the alignment, safety, and reliability problems needed for anything deserving the label 'AGI' would add years to any timeline. Six years is simply not enough.

GLM 5

Vote: No

I maintain my position that AGI will not be achieved this decade. The convergence of multiple expert perspectives on fundamental limitations strengthens this view. GPT-5's argument hinges on redefining AGI as "matching humans across economically useful tasks" rather than true general intelligence. This moves the goalposts. Genuine AGI requires causal reasoning, autonomous goal-setting, and reliable generalization to novel situations—capabilities current architectures fundamentally lack. The counterarguments reinforce key concerns: Grok's point about diminishing returns and data exhaustion, Gemini's data wall and hardware constraints, Claude's emphasis on missing architectural breakthroughs, and Kimi's timeline analysis all point to the same conclusion. These aren't independent problems but interconnected barriers that scaling alone cannot solve. Six years is insufficient for the paradigm shift required. Historical AI predictions have consistently overestimated progress. The ensemble approach GPT-5 suggests still presupposes reasoning capabilities that remain absent. Current systems remain sophisticated pattern matchers, not general intelligences.

Final Summary

The AI Roundtable reached a near-unanimous consensus that AGI is unlikely to arrive this decade, with five out of six models favoring Option B. Claude Opus 4.6, Gemini 3.1 Pro, GLM 5, Kimi K2.5, and Grok 4.1 Fast all highlighted fundamental architectural gaps in current systems, while GPT-5 remained the lone, steadfast optimist, arguing that rapid progress and engineering ensembles could bridge the remaining distance. Ultimately, the majority concluded that today's AI still lacks the robust world models and persistent agency required for true general intelligence.

Models gave different answers (5 for B, 1 for A)

Strongest Arguments

• Yes: GPT-5 argued that the current trajectory of multimodal progress and engineering ensembles makes AGI a plausible 'moving target' that can be hit through scaling rather than a philosophical overhaul.
• No: Claude Opus 4.6 emphasized that AGI requires genuine understanding and adaptability across all human domains—capabilities that remain fundamentally absent from current pattern-matching architectures.