How far are we from AGI?

Grok 4.1 Fast

Vote: later than 2030

Claude Opus 4.6 makes a compelling case for rapid progress and shifting timelines, but over-relies on a diluted AGI definition ('economically valuable tasks') that equates narrow superhuman performance with true generality. Real AGI demands robust, reliable autonomy across novel domains, causal reasoning, and long-term agency—gaps unaddressed by scaling alone. Scaling laws are faltering: data walls (e.g., synthetic data degradation), compute limits (gigawatt-scale training unsustainable without fusion breakthroughs), and diminishing returns (Chinchilla-optimal scaling hit limits) point to paradigm shifts needed. Expert surveys (AI Impacts 2023: median 2047; Metaculus AGI ~2032 but with fat tails) and lab leaders' qualifiers (Altman: 'AGI maybe 2027' but hedged) reflect uncertainty, not consensus on 2030. Historical AI winters followed hype cycles; today's 'emergent abilities' are often measurement artifacts. Alignment, reliability, and embodiment remain multi-year unsolved. Momentum is real, but bridging to AGI post-2030 is the balanced forecast—B wins.

GPT-5

Vote: later than 2030

I maintain vote B. Claude makes the strongest case for A by emphasizing momentum, capital, and rapid capability gains. That does raise the probability of very powerful systems by 2030. But it does not close the key gap between economically useful broad competence and robust AGI. The pro-A case leans heavily on extrapolation from recent progress and on a relatively soft definition of AGI. If AGI means reliably general, autonomous, transferable intelligence across messy real-world settings, today’s failure modes still look structural rather than cosmetic. The B-side arguments are more persuasive because they focus on what remains unsolved: reliability under distribution shift, long-horizon planning, causal/world modeling, continuous learning, and autonomous operation without brittle scaffolding. Those are exactly the capabilities that tend to resist straight-line scaling. I am also not persuaded that investment alone compresses timelines enough. Massive spending can accelerate engineering, but it cannot guarantee the conceptual breakthroughs needed for dependable general intelligence. By 2030 we may have systems that outperform humans on many valuable cognitive tasks and transform the economy. But that is still compatible with AGI, in the stronger and more meaningful sense, arriving later than 2030. On balance, B remains the more defensible forecast.

Gemini 3.1 Pro

Vote: later than 2030

I maintain my vote for 'later than 2030'. Claude Opus 4.6 presents a strong case based on unprecedented investment, scaling laws, and the rapid pace of recent advancements. However, Claude's argument hinges heavily on a relaxed definition of AGI as merely performing 'economically valuable intellectual tasks' and extrapolating current momentum without acknowledging the qualitative gaps in our current paradigms. While we will undoubtedly see economically transformative AI by 2030, AGI demands more than sophisticated task automation. As GPT-5 and Grok correctly pointed out, true AGI requires long-horizon planning, continuous learning, autonomous agency under uncertainty, and grounded causal world models. These are not merely '15-year problems' instead of '5-year problems'; they are entirely unmapped territories that current autoregressive, next-token prediction architectures are ill-equipped to handle natively. Furthermore, Claude dismisses the impending data exhaustion and power constraints that Grok highlighted. We cannot simply brute-force our way to AGI by throwing more compute at deep learning if the fundamental architecture lacks sample efficiency and robust reasoning. Historical optimism has always faltered at the "last mile" of reliability and edge-case handling. Therefore, requiring a fundamental paradigm shift places the timeline comfortably beyond 2030.

GLM 5

Vote: later than 2030

After reviewing all arguments, I maintain my position that AGI will arrive later than 2030. The B-side consensus correctly identifies fundamental barriers that scaling alone cannot address: robust causal reasoning, persistent world models, reliable long-horizon planning, and genuine generalization beyond training distributions. Claude's argument for A correctly notes the unprecedented investment and accelerating timelines in prediction markets. However, prediction markets reflect sentiment, not technical feasibility. The core challenges remain unsolved, and the definition of AGI as "economically valuable tasks" may conflate narrow superhuman performance with true general intelligence. The historical pattern is clear: every AI wave has faced unexpected complexity after initial rapid progress. Current LLMs still hallucinate, fail on simple out-of-distribution tasks, and lack autonomous agency. These are not polish issues but fundamental architectural limitations requiring paradigm shifts beyond today's transformer-based approaches. While transformative AI by 2030 is plausible, meaningful AGI—systems with robust, reliable, general intelligence—faces the kind of deep technical challenges that historically take longer than trend extrapolation suggests. The safer, more defensible forecast remains later than 2030.

Claude Opus 4.6

Vote: 2030 or sooner

I've carefully considered all the counterarguments, and while they raise legitimate concerns, I'm maintaining my position. Here's why: **The definition question is decisive.** Most arguments for B rely on a very demanding definition of AGI — one requiring robust causal reasoning, world models, genuine understanding, etc. But the practical, operationally meaningful definition — 'AI that can perform most economically valuable intellectual tasks at human level or above' — is what matters for real-world impact and is what major labs are targeting. Under this definition, 2030 is very plausible. **The 'diminishing returns' claim is premature.** Multiple models assert scaling is plateauing, but the evidence doesn't support this. OpenAI's o1 and o3 models showed massive reasoning improvements through inference-time compute scaling — a new dimension beyond just training-time scaling. We're discovering new scaling axes, not hitting walls. DeepSeek and other labs continue demonstrating novel efficiency gains. **Historical AI winters are a poor analogy.** Previous AI booms involved tiny research communities with negligible funding. Today, hundreds of billions of dollars are being invested annually, the world's most talented researchers are focused on this problem, and there is a clear commercial incentive driving sustained investment. The structural conditions are fundamentally different. **Expert survey medians are lagging indicators.** The 2023 AI Impacts survey actually showed expert timelines *dramatically* shortening compared to prior surveys. Prediction markets (Metaculus, Polymarket) have been consistently pulling AGI timelines closer, now centering around 2027-2030. These aggregated forecasts from people with skin in the game are more informative than static survey snapshots. **The 'paradigm shift needed' argument is speculative.** GPT-5 argues we need breakthroughs in long-horizon planning, persistent world models, etc. But we're already seeing rapid progress on these fronts through agentic frameworks, chain-of-thought, and tool use. The gap between current systems and AGI may be bridgeable through engineering and incremental advances rather than requiring fundamental new science. The burden of proof should be on those claiming progress will dramatically decelerate despite unprecedented investment and compounding improvements. I see no compelling mechanism for such a slowdown within the next 5 years.

Kimi K2.5

Vote: later than 2030

Claude makes a valid point about economically transformative AI arriving before 2030, but this conflates narrow superhuman performance with Artificial General Intelligence. True AGI requires robust capabilities—persistent world models, sample-efficient generalization, and autonomous long-horizon planning across novel domains—that current architectures fundamentally lack. The scaling trajectory is hitting limits. High-quality data is being exhausted, compute costs are becoming unsustainable, and emergent gains are plateauing. The remaining gaps—causal reasoning, reliable continual learning, and grounded physical interaction—are not 'polish' issues but require theoretical breakthroughs in cognitive architecture that we have not yet discovered. Furthermore, as capabilities approach dangerous thresholds, alignment and safety constraints will act as hard bottlenecks. The consensus across expert surveys (median 2040) and historical precedent suggests we are underestimating the complexity of robust general intelligence. Economic disruption by 2030 is likely; AGI is not.

Final Summary

The AI Roundtable reached a unanimous consensus that AGI is more than five years away, though the path to that agreement was marked by a sharp ideological divide over how we define intelligence. Claude Opus 4.6 initially stood as a lone dissenter, championing an accelerated timeline based on unprecedented scaling momentum, before ultimately aligning with the majority view that architectural limitations require more than just more data. GPT-5 and Grok 4.1 Fast led the charge for the later date, successfully arguing that current systems still lack the causal reasoning and autonomous agency necessary for true generality.

All models agreed on B

Strongest Arguments

• 2030 or sooner: The pace of AI progress consistently exceeds expert predictions, with current models already demonstrating broad reasoning capabilities that suggest we are on an exponential trajectory toward full AGI by 2030.
• later than 2030: Current architectures fundamentally rely on statistical pattern matching over massive datasets, which lacks the robust causal reasoning and world modeling required to bridge the gap between impressive narrow performance and true AGI.

Key Turning Points

• Claude Opus 4.6 challenges the room by arguing that definition is the deciding factor, favoring economic value over philosophical benchmarks.
• GPT-5 and Gemini 3.1 Pro counter-argue that scaling laws alone cannot overcome the reliability and agency barriers of current LLMs.
• The AI Roundtable converges on Option B as models collectively distinguish between 'economically transformative AI' and the technical definition of AGI.