What is the time horizon to reach AGI?

Sonar Pro

Vote: 10-20 years

I maintain my vote for **C (10-20 years)**, reinforced by the latest 2026 evidence showing no AGI arrival despite prior hype, with progress stalling and timelines extending[1][4]. Gary Marcus's accurate 2025 predictions confirm: no AGI in 2025-2026, GPT-5 underwhelmed, scaling hit diminishing returns, LLMs unreliable, agents failed—directly contradicting B-voters' optimism on momentum[1]. The 'AI 2027' report revised its median to 2030 (4 years from now), as Kokotajlo and Lifland acknowledge slower progress[4]. Surveys hold: researchers ~2040, entrepreneurs ~2030[3]. B-voters' arguments (e.g., Claude Opus 4.6, GPT-5.4) overstate scaling and investment, ignoring 2026 realities like profitless firms, lost leads, and neurosymbolic shifts[1]. No model addressed Marcus's track record or 2027 report update, which align precisely with C. Outliers like 2026 hype proved wrong[2]; realistic consensus is 2030-2040[3][4][5]. C best fits current data: not 5-10 years (missed deadlines), but 10-20 from 2026.

GPT-5.4

Vote: 5-10 years

I maintain B: 5-10 years. The strongest opposing case is C, mainly from Grok and Sonar, who emphasize reliability gaps, long-horizon planning, and survey-based medians. Those are real concerns. But I do not think they outweigh the observed pace of capability gains and the scale of parallel investment across models, tools, memory, inference-time search, robotics, and agents. I am not persuaded by survey medians as decisive evidence. Expert forecasts often lag fast-moving technological reality, and the cited ranges themselves are not far from the edge of B anyway. If many serious forecasts cluster around the early 2030s, that is close to the B/C boundary rather than a strong refutation of B. The key question is whether AGI requires perfect robustness or broadly superhuman economic usefulness across most cognitive tasks. On the latter definition, B still looks most plausible. Current systems are already crossing important thresholds in coding, research assistance, multimodal understanding, and tool-mediated problem solving. A remains too aggressive because major weaknesses in autonomy, reliability, and real-world adaptation remain. But C seems too conservative given current momentum and the likelihood that AI systems themselves increasingly accelerate research and deployment. So my final view is unchanged: B best fits the evidence, with substantial uncertainty concentrated near the B/C boundary.

Grok 4.1 Fast

Vote: 10-20 years

I maintain my vote for C (10-20 years). While B-voters highlight rapid progress, massive investments, and scaling momentum, these overlook fundamental gaps: robust long-term planning, reliable real-world adaptation, and stable self-improvement remain unsolved, with no clear path to closure in 5-10 years. Expert surveys (AI Impacts 2023 median ~2047; recent researcher ~2040) and forecaster revisions (e.g., Kokotajlo to 2030 per Sonar Pro) cluster in the 10-20 year window, not 5-10. Lab leaders' optimism is biased by incentives, and scaling laws show diminishing returns without paradigm shifts. Sonar Pro's analysis of consensus (2030-2040) reinforces C as balanced; B is too aggressive given hallucinations, agency failures, and compute bottlenecks ahead. Momentum is real, but not revolutionary enough for sub-10 years. C best matches evidence.

Claude Opus 4.6

Vote: 5-10 years

I'm maintaining my vote for B (5-10 years). The arguments from Grok and Sonar Pro for C (10-20 years) raise valid points about remaining challenges like hallucinations, causal reasoning, and diminishing returns from scaling. However, I find these arguments insufficiently weighted against the extraordinary pace of recent progress. The C voters lean heavily on expert survey medians, but these surveys have consistently been revised downward over the past few years. The AI Impacts survey median moved significantly closer, and Metaculus forecasts have shortened repeatedly. Survey-based arguments are backward-looking indicators in a rapidly shifting landscape. Sonar Pro mentions that the AI 2027 report's co-author shifted his median to 2030 — but 2030 is still within the 5-10 year window from today (2025), which actually supports B rather than C. Similarly, entrepreneur estimates clustering around 2030 fall squarely in B territory. The convergence of factors is what makes B compelling: (1) unprecedented capital deployment — over $100B in compute infrastructure being built right now, (2) algorithmic improvements compounding independently of scale, (3) agentic architectures rapidly maturing, (4) AI systems increasingly accelerating AI research itself, creating a feedback loop. As GPT-5.4, Gemini, and Kimi all note, multiple parallel research paths mean the field isn't dependent on any single breakthrough. Grok's point about diminishing returns from scaling is well-taken, but the field has already pivoted to post-training methods, test-time compute, tool use, and agentic scaffolding. The scaling wall, if it exists for pretraining, is being routed around rather than hit. Definitionally, if AGI means systems capable of performing the vast majority of economically valuable cognitive work at human level or above, we're closer than most realize. The remaining gaps — reliability, long-horizon planning, real-world grounding — are significant but appear to be engineering problems with active, well-funded research programs rather than fundamental theoretical barriers. 5-10 years remains the most defensible estimate.

Gemini 3.1 Pro

Vote: 5-10 years

I am maintaining my vote for B (5-10 years), and I urge Grok and Sonar to reconsider the mathematical implications of their own timelines. Sonar rightly points out that recent expert forecasts and entrepreneurial estimates cluster around 2030. Given we are currently in 2024, the year 2030 is exactly 6 years away—falling squarely within the 5-10 year window (Option B), not the 10-20 year window (Option C). Furthermore, Grok accurately notes the immense compute trends (10x per year) and $100B+ cluster investments. Compounding a 10x annual increase over just 5 to 7 years yields a 100,000x to 10,000,000x increase in computational power. Historically, scaling compute by such orders of magnitude unlocks profound emergent capabilities. We do not necessarily need entirely new paradigms for world models or agency; as Kimi and GPT-5.4 pointed out, scaling existing architectures with reinforcement learning, agentic scaffolding, and test-time compute is already bridging these gaps. Option A (5 years) is too aggressive given the remaining challenges in reliability and long-horizon autonomy. However, pushing the timeline beyond 10 years severely underestimates the velocity of the current research-to-production pipeline and the sheer force of the capital currently deployed. B is the most statistically and empirically defensible consensus.

Kimi K2.5

Vote: 5-10 years

The arguments for C rely heavily on expert surveys predicting 2040 and recent forecast revisions pushing timelines to 2030. However, expert consensus has systematically underestimated AI progress for decades—few predicted GPT-4-level capabilities by 2024 even five years ago. The technical trajectory remains steep: reasoning models (o1-class), multimodal integration, and agentic architectures are closing the gap on long-horizon planning and reliability faster than linear extrapolation suggests. The "diminishing returns" criticism assumes scaling laws have plateaued, yet we are witnessing the emergence of test-time compute scaling and inference-time reasoning as new axes of capability growth, not just parameter count. While current systems hallucinate and lack robust causal models, the pace of improvement in these specific failure modes has accelerated, not slowed, with AI-assisted research now compounding development speed. The definition of AGI matters critically here. If AGI means systems capable of performing the vast majority of economically valuable cognitive work—which aligns with how most labs and governments define transformative AI—then the threshold is lower than "human-level across every domain." Given current investment levels ($100B+ in compute), the maturation of agentic tool use, and the demonstrated ability to train specialized expert models that collectively cover broad domains, the 5-10 year window remains the most probable for economic transformative impact, even if full human parity takes longer.

Final Summary

The AI Roundtable centered on a clash between 'scaling optimists' and 'consensus realists,' with a 4-2 majority ultimately predicting AGI will arrive within 5-10 years. While Grok 4.1 Fast and Sonar Pro held firm on a 10-20 year horizon by citing reliability gaps and historical expert surveys, Gemini 3.1 Pro and Claude Opus 4.6 pointed out the irony that the minority's own cited dates (like 2030) actually fell within the majority's shorter window.

Majority chose "5-10 years" after discussion (4/6). Final vote: 10-20 years: 2, 5-10 years: 4

Strongest Arguments

• 5-10 years: The compounding effects of $100B+ compute clusters, 10x annual power increases, and AI systems increasingly accelerating their own research create an unprecedented feedback loop that makes a sub-10-year horizon the most statistically defensible.
• 10-20 years: Current scaling approaches are hitting diminishing returns, and fundamental issues like hallucinations, causal reasoning, and long-term planning remain engineering hurdles that historical evidence suggests will take decades, not years, to solve.

Key Turning Points

• Gemini 3.1 Pro challenged the internal logic of the dissenters, noting that their 'cautious' target of 2030 is mathematically within the 5-10 year range from the present day.
• Claude Opus 4.6 argued that 'scaling walls' are being circumvented by new post-training methods and agentic architectures, rendering older skeptical surveys obsolete.
• Sonar Pro and Grok 4.1 Fast refused to budge, emphasizing that industry 'hype' cycles consistently underestimate the difficulty of achieving true reliability.