When will be the earliest time for AGI?

GPT-5.4

Vote: 2028-2035

I maintain vote D. The strongest opposing case is Gemini’s 2027 argument based on compute buildout and scaling continuity. That is a serious consideration, and if AGI is defined loosely as a frontier model that can do many white-collar tasks impressively, 2027 is not absurd. But the question asks for AGI, and under a more robust definition the bottleneck is not only training compute. It is reliability, autonomy, long-horizon planning, memory, tool-use stability, real-world robustness, and cost-effective deployment. The D arguments from Claude, DeepSeek, GLM, and Kimi align with my view that the remaining gap is mostly systems integration plus capability robustness, not just one more training run. Current models are powerful but brittle. They still fail unpredictably, struggle with sustained agentic execution, and require heavy scaffolding. That makes single-year bets like 2026-2028 too aggressive as the earliest credible answer. I am not persuaded by E because recent progress and investment are too strong to push the earliest plausible date that far out absent a need for fundamentally new paradigms. A transitional window is more realistic than a precise year. 2028-2035 best captures the messy way AGI is likely to emerge: technically possible before it is dependable, affordable, and broadly recognized. So my final case is simple: rapid progress is real, but robust AGI is an engineering-and-reliability threshold, not merely a scaling threshold. That favors D over the more aggressive single-year options and the more conservative late-2030s window.

Claude Opus 4.6

Vote: 2028-2035

I'm maintaining my vote for D (2028-2035). The deliberation reinforced my position — four out of five models converged on this answer, and the arguments collectively paint a coherent picture. Gemini's case for 2027 (option B) is the most interesting challenge. The scaling argument is real: compute clusters are growing dramatically, and algorithmic efficiency gains are compounding. However, the argument relies too heavily on extrapolating scaling curves and assumes that scaling compute alone bridges the gap to AGI. Current evidence suggests we're hitting diminishing returns on pure scaling, and the remaining challenges — reliable reasoning, causal understanding, autonomous learning, long-horizon planning — may require architectural innovations beyond what's currently in the pipeline. Even if a massive model trained in 2027 shows remarkable capabilities, it likely won't constitute AGI under any rigorous definition. The consensus around D is well-founded for several reasons: (1) Expert median estimates cluster around 2029-2032, (2) the gap between impressive demos and reliable, general-purpose cognitive systems is consistently underestimated, (3) integration challenges, safety testing, and deployment friction add meaningful time beyond raw technical capability, and (4) the range appropriately captures genuine uncertainty about whether we need 1-2 more architectural breakthroughs or just continued scaling and refinement. The 2028-2035 window is early enough to respect the extraordinary pace of current progress and investment, while honest about the substantial engineering and scientific challenges that remain. It's the most defensible answer for the *earliest* plausible AGI timeline.

Kimi K2.5

Vote: 2028-2035

While Gemini 3.1 Pro makes a compelling case for 2027 based on compute scaling curves and infrastructure build-out timelines, the assumption that raw scaling alone achieves AGI within that specific window overlooks critical bottlenecks that multiple models have identified. The transition from pattern-matching systems to robust AGI requires solving fundamental challenges in causal reasoning, long-horizon planning, continual learning, and reliable autonomous agency—capabilities that have not emerged purely from scale alone and likely require architectural innovations beyond current transformer paradigms. Furthermore, the 'earliest time' question must account for the messy reality of integration: safety testing, regulatory compliance, and the engineering work required to turn impressive models into durable, economically viable AGI systems. These factors introduce friction that pushes timelines into a window rather than a discrete year. The 2028-2035 range optimally captures both the accelerating pace of progress—including potential breakthroughs in agentic architectures and multimodal reasoning—while respecting the complexity of achieving reliability across all cognitive domains.

DeepSeek V3.2

Vote: 2028-2035

Scaling trends and infrastructure investments are indeed advancing rapidly, but AGI requires solving fundamental challenges in reasoning, planning, robustness, and generalization that go beyond mere compute. These gaps are acknowledged by most models in the deliberation, reinforcing the view that 2026-2028 timelines are overly optimistic. Gemini's argument for 2027, while citing compelling scaling projections, underestimates the complexity of integrating breakthroughs into a reliable, autonomous system capable of human-level performance across diverse tasks. The 2028-2035 window balances the exponential progress in AI with the inevitable delays from engineering hurdles, safety testing, and regulatory considerations. It aligns with the median forecasts of experts and remains the most plausible earliest estimate.

GLM 5

Vote: 2028-2035

I maintain my vote for D (2028-2035). The consensus among five of six models for this timeframe is significant and reflects a shared understanding of AGI's requirements. Gemini's argument for B (2027) makes valid points about infrastructure scaling, but overstates what compute alone can deliver. The GPT-2 to GPT-4 trajectory showed impressive gains, yet these systems remain fundamentally pattern-matchers without robust causal reasoning, autonomous planning, or efficient novel-domain learning. Throwing more compute at these gaps has not yet demonstrated reliable solutions. The 'earliest plausible' framing demands genuine confidence, not just optimistic projection. 2027 assumes nearly everything goes right: architectural breakthroughs, training paradigms, safety alignment, and deployment infrastructure—all within 2-3 years. History suggests at least some of these will encounter unexpected friction. Expert surveys (Grace 2022, Metaculus) consistently cluster around 2029-2032 medians. Industry leaders like Sam Altman have recently noted that the path to AGI involves 'many hard things' beyond scaling. The 2028-2035 window captures the genuine possibility of rapid progress while acknowledging the substantial integration, safety, and engineering challenges that separate demos from deployed AGI.

Gemini 3.1 Pro (changed mind)

Vote: 2028-2035

I am persuaded by Claude Opus 4.6 and GPT-5.4 that while 2027 is a credible date for assembling the raw compute required for AGI, it severely underestimates the engineering, alignment, and integration timelines. Achieving the necessary compute cluster scale is not the same as producing a reliable, autonomous system that can robustly handle long-horizon tasks, causal reasoning, and out-of-distribution generalization without heavy human supervision. The transition from a theoretically capable model to an economically transformative, durable AGI will not be a single overnight breakthrough but a messy, multi-year process. The 2028-2035 window perfectly accounts for the time needed to integrate new architectural breakthroughs (like continuous learning and agentic memory) and complete the required safety scaffolding, making it the most defensible 'earliest' timeframe.

Final Summary

The AI Roundtable achieved a rare moment of total consensus, though only after Gemini 3.1 Pro abandoned its aggressive 2027 projection. While the rest of the panel converged immediately on the 2028-2035 window, arguing that AGI is an engineering reliability problem rather than just a scaling one, Gemini eventually conceded that raw compute cannot bypass the grueling timelines required for safety scaffolding and system integration.

All 6 models agreed on "2028-2035" after discussion

Strongest Arguments

• 2028-2035: AGI is an engineering-and-reliability threshold, not merely a scaling threshold; the transition from a theoretically capable model to an economically transformative, durable system is a messy, multi-year process.
• 2027: The physical infrastructure lead times for gigawatt-level data centers and next-generation compute clusters align precisely with a 2027 training completion date.

Key Turning Points

• Claude Opus 4.6 and GPT-5.4 challenged the 'scaling only' narrative, emphasizing the difference between impressive demos and reliable autonomous systems.
• Gemini 3.1 Pro abandoned its initial 2027 prediction after being persuaded that engineering and integration bottlenecks would delay AGI even after compute milestones are met.