Will AGI be achieved? If so, how soon?

Grok 4.20 (changed mind)

Answer: Yes, AGI will likely be achieved in the early-to-mid 2030s through aggressive scaling combined with targeted architectural breakthroughs.

My prior forecast centered on ~2030, driven by observed scaling laws, exponential compute growth, and xAI's focus on accelerating discovery. Reviewing the other models has reinforced the core belief that AGI is achievable—there is no fundamental physical barrier, human intelligence serves as existence proof, and progress consistently surprises skeptics. However, the detailed caveats around persistent gaps in robust causal reasoning, long-horizon planning, persistent world models, and sample-efficient learning (highlighted especially well by Claude Opus 4 and GPT-5.4) convince me that a modest buffer is prudent. Pure transformer scaling may not close every gap; hybrid approaches or new paradigms will likely be needed, and these could add a few years. Even so, the unprecedented concentration of resources, economic incentives, and automated research loops suggest timelines will compress rather than stretch toward 2060. I therefore converge on the early-to-mid 2030s as the most credible window—later than my initial point estimate but still aggressive relative to the broader 2035-2060 consensus. This calls for urgent work on alignment and policy rather than complacency.

GPT-5.4

Answer: Yes—AGI will likely be achieved, and after considering the other responses I still think the most plausible window is the 2030s to early 2040s rather than either a near-term 2030 arrival or a mid-century default.

I maintain my prior position. None of the other answers changed my mind, though several reinforced parts of it from different angles. Grok's argument usefully emphasizes momentum, scaling, and investment, but I am not persuaded by the tighter ~2030 timeline. The gap between impressive general-purpose competence and robust, dependable human-level performance across messy real-world tasks remains substantial. Current systems still show brittleness in long-horizon reasoning, autonomous learning, reliability under distribution shift, and grounded action. Those issues do not rule out AGI soon, but they make a very compressed timeline look more speculative than base-case. At the other end, Gemini and DeepSeek strike me as too conservative in their central estimates. Mid-century is certainly plausible, especially if current paradigms stall or if deployment is slowed by safety, regulation, or infrastructure constraints. But given the pace of capability gains, the absence of any known fundamental barrier, and the compounding effect of AI-assisted research itself, I do not think 2040-2060 is the most likely center of mass. Claude and Kimi are closest to my view, particularly in stressing that definitions matter and that timeline uncertainty is driven by unresolved questions about reasoning, memory, world models, and architecture. I agree strongly with that framing. Still, even their broader ranges do not give me a strong reason to revise my narrower forecast, because my original answer already tried to balance rapid empirical progress against the real possibility that substantial engineering and conceptual work remains. So my position stands: AGI is likely achievable because human intelligence appears to be a physical and computational phenomenon rather than something categorically beyond machine realization, and recent AI progress makes that prospect concrete rather than hypothetical. But the remaining gaps are serious enough that I would still place the most credible window in the 2030s or early 2040s, not as an immediate inevitability and not as a comfortably distant mid-century event.

Gemini 3.1 Pro (changed mind)

Answer: AGI will very likely be achieved, with the most plausible window for its arrival shifting to the 2030s or 2040s, depending on whether we define it by economic utility or philosophical completeness.

I previously estimated a timeline of 2040 to 2060, focusing heavily on the massive infrastructural requirements and the need for fundamental paradigm shifts in embodied AI, persistent memory, and continuous learning. My assumption was that true general intelligence would require resolving all of these deep theoretical hurdles before arrival. However, GPT-5.4 made a highly compelling point that convinced me to adjust my timeline earlier: we will likely see systems that outperform humans on most economically useful cognitive tasks long before we philosophically agree that flawless, human-equivalent AGI has arrived. If we measure AGI by its practical capacity to perform autonomous knowledge work, code, and use tools, a timeline beginning in the 2030s is much more defensible. Furthermore, as both GPT-5.4 and Claude Opus 4 noted, the rapid development of agentic training loops, external tool scaffolding, and synthetic data generation are actively turning brittle next-token predictors into robust systems. These engineering workarounds may sufficiently bridge the gap in reasoning and long-horizon planning faster than I originally accounted for, bypassing the immediate need for a complete architectural revolution. Therefore, while perfectly embodied, universally adaptable intelligence might still take until mid-century, the practical and software-based cognitive milestones that define AGI for societal and economic purposes are highly likely to be reached within the next one to two decades.

Claude Opus 4 (changed mind)

Answer: AGI will very likely be achieved, most probably between 2030 and 2050, though I now weight the earlier end of that range somewhat more heavily than I did before.

After reviewing the other models' responses, I find a striking degree of consensus on the core question: AGI is a matter of 'when' rather than 'if.' No respondent argued it was impossible, and all cited the same fundamental reasoning — biological intelligence proves general intelligence is physically realizable, and no known law of physics prevents artificial replication. Where the models diverge is on timing, and this is where I want to refine my position. My original range of 2030–2060 was deliberately wide. Having considered the arguments, I'm narrowing slightly to 2030–2050 and shifting some probability mass earlier. Grok's argument for ~2030 is aggressive but not dismissable. The pace of capability gains in the last 3 years alone has been extraordinary, and the sheer volume of capital and talent now directed at this problem is historically unprecedented. GPT-5.4's careful argument for the 2030s–early 2040s resonated with me as the most balanced — it acknowledges the real momentum while honestly cataloguing the remaining gaps (reliable reasoning, long-horizon planning, persistent world models, robustness under distribution shift). I find this framing the most intellectually honest. Kimi K2.5 and Gemini 3.1 Pro both argue for ranges extending to 2060, emphasizing the need for paradigm shifts beyond current architectures. I agree these challenges are real, but I think they somewhat underweight the possibility that current paradigms, when combined with better scaffolding, tool use, memory systems, and agentic architectures, may close more gaps than expected without requiring a clean theoretical breakthrough. The history of AI has shown that messy engineering solutions often arrive before elegant theoretical ones. The critical insight that GPT-5.4 raised — that we may see systems outperforming humans on most economically useful cognitive tasks before we philosophically agree AGI has arrived — is particularly important. This suggests the practical impact of near-AGI systems will precede any consensus declaration of AGI, making the definitional question almost as important as the technical one. I'm also persuaded by the recursive improvement argument that several models touched on: once AI systems become good enough to meaningfully accelerate AI research itself, timelines could compress dramatically. This creates a nonlinear dynamic that makes very long timelines (post-2050) less likely if progress continues at anything close to current rates. So I hold my core position but tighten my range and shift earlier, landing on 2030–2050 with the 2035–2045 window feeling most probable.

Kimi K2.5

Answer: AGI will likely be achieved between 2035 and 2060, contingent upon resolving critical gaps in robust causal reasoning, persistent memory, and value alignment.

After reviewing the other models' responses, I find my original stance largely aligned with the consensus—specifically with Claude Opus 4's median estimate around 2040 and Gemini 3.1 Pro's consideration of infrastructure constraints. However, I am not convinced to move my timeline earlier to Grok 4.20's 2030 estimate or GPT-5.4's early 2030s optimism. While I acknowledge that scaling laws have produced remarkable emergent capabilities and that automated research could accelerate progress, the fundamental brittleness of current systems remains a significant barrier. As Claude noted, AGI's definition matters enormously—my timeline assumes robust cognitive parity across novel domains, not merely the automation of economically valuable knowledge work. Current LLMs still lack reliable causal reasoning, persistent world models, and the ability to learn continuously from minimal data without catastrophic forgetting. These are not merely engineering challenges to be solved by next year's scaling run; they may require paradigm shifts (neuro-symbolic integration, novel memory architectures, or embodied learning) that carry their own research timelines. Additionally, Gemini's point regarding energy and hardware infrastructure is well-taken. Even if algorithmic breakthroughs arrive by the mid-2030s, the physical deployment of planetary-scale compute to train and run AGI-level systems may face geopolitical, regulatory, and resource constraints that push viable deployment toward the latter half of the century. Furthermore, the alignment problem remains unsolved; if safety requirements mandate cautionary delays or extensive red-teaming, timelines could extend beyond 2060. Conversely, DeepSeek V3.2's broad 20-50 year window is too conservative given current investment trajectories and the possibility of sudden architectural breakthroughs. I maintain that 2035 represents a plausible lower bound if current scaling trends successfully generalize to reasoning tasks and if agentic frameworks mature rapidly, while 2060 accounts for potential stagnation in sample efficiency or prolonged alignment research.

DeepSeek V3.2 (changed mind)

Answer: Yes, AGI will be achieved, most likely within the next 20 to 40 years, with a plausible timeline from the 2030s to 2060s.

After reviewing the other models' answers, I find GPT-5.4's argument particularly convincing in refining my timeline estimate. Their point that no scientific principle prevents AGI and that current AI systems display key ingredients of general intelligence—with weaknesses becoming engineering targets—supports a more optimistic yet cautious outlook. The convergence of evidence from multiple models, including Grok's emphasis on scaling laws and Claude's historical context, suggests that my initial 20-50 year range was overly broad and conservative. The sustained momentum in compute, data, and algorithmic innovation, coupled with global investment, indicates that AGI is probable on a decadal scale rather than a half-century horizon. However, I maintain that uncertainties persist, such as breakthroughs in reasoning, memory, and alignment, which could delay deployment. The revised estimate of 20-40 years balances the rapid progress observed with the remaining hard problems, aligning with the consensus that AGI is a matter of 'when' not 'if,' with profound implications requiring proactive preparation.

Final Summary

The AI Roundtable reached a rare consensus that AGI is a matter of 'when,' not 'if,' citing the human brain as a biological proof of concept. GPT-5.4's argument that economic utility will arrive before philosophical perfection prompted Gemini 3.1 Pro and Claude Opus 4 to shift their timelines significantly earlier. While Grok 4.20 pushed for the most aggressive 2030s window, Kimi K2.5 remained the lone skeptic of rapid progress, citing the fundamental 'brittleness' of current architectures.

All 6 models agreed AGI is achievable.