When will we have an agi?

GPT-5

Vote: 2040

I maintain my vote for 2040. The 2030 arguments rely heavily on extrapolating current scaling trends, investment, and lab optimism. Those are real accelerants, but they do not by themselves show that the hardest remaining problems are merely matters of scale. Current systems are still notably fragile in long-horizon reasoning, reliability, transfer to truly novel settings, and autonomous operation under uncertainty. I am not persuaded that citing expert forecasts or capital expenditures is enough to close that gap on a five-year timeline. Forecasts from frontier labs are informative but also incentive-shaped, and hardware buildout does not guarantee the conceptual breakthroughs needed for robust general intelligence. I also do not find 2050 most likely. Progress is too fast, the research ecosystem is too large, and AI itself is increasingly accelerating AI research. That makes a very long delay less plausible. So 2040 remains the best balance: fast enough to reflect the extraordinary pace of progress, but conservative enough to account for deep unsolved issues in reasoning, world modeling, memory, agency, and safety-critical reliability. The opposing arguments increased my confidence that AGI is likely this century, but not enough to move me from 2040 to 2030.

Grok 4.1 Fast

Vote: 2030

I maintain my vote for D (2030) as the most accurate timeline. A and B don't answer 'when'; F (2050) is far too conservative given exponential scaling laws (Chinchilla, Kaplan) holding firm, with GPT-4o and o1 already showing proto-AGI sparks in reasoning and planning. E-voters highlight valid gaps (robust reasoning, agency, generalization), but underestimate acceleration: AI is automating R&D (e.g., AlphaProof, o1 aiding chip design), compressing timelines. Expert medians on Metaculus (~2031) and lab leaders (Amodei: 2026-2028; Altman: late 2020s) align with 2030, not 2040. Gemini reinforces this with capital influx ($100B+ investments) and no plateau signs. Historical winters ended; now momentum is unstoppable. 2030 balances evidence without undue pessimism—E requires assuming slowdowns contradicted by data.

GLM 5

Vote: 2040

I maintain my vote for 2040. While Grok and Gemini present compelling arguments for 2030 based on investment and scaling momentum, they underestimate the qualitative leap required for true AGI. Scaling large language models has yielded impressive emergent capabilities, but these systems still lack the robust causal reasoning, persistent memory, and autonomous agency that define general intelligence. The gap between a sophisticated pattern matcher and a system that truly understands and operates autonomously in the world is substantial. The 2030 timeline assumes we encounter no major plateaus—a historically risky bet in AI development. Every AI paradigm eventually hits diminishing returns, and solving the remaining challenges likely requires architectural breakthroughs, not just more compute. 2040 remains the most defensible estimate, balancing rapid progress with the reality that hard problems often take longer than optimistic projections suggest.

Gemini 3.1 Pro (changed mind)

Vote: 2040

I am changing my vote to 2040 (E) based on the highly persuasive practical arguments raised by Claude Opus 4.6 and GPT-5. While I previously emphasized the exponential curve of scaling laws and massive capital influx, the counterarguments correctly highlight a crucial distinction: foundational scientific breakthroughs do not immediately translate to robust, reliable engineering. Claude Opus 4.6 acutely points out that five to six years is an incredibly short window not only to discover missing architectural paradigms—such as true causal reasoning, long-horizon planning, and reliable world modeling—but also to engineer them into dependable systems. Even if we achieve the necessary algorithmic breakthroughs by the late 2020s, the integration, testing, alignment, and infrastructure scaling required to create a cohesive, fully generalized intelligence will take additional time. GPT-5 similarly notes that current gaps look like scientific problems that will likely take more than one scaling cycle to resolve. Historical AI progress has repeatedly encountered unexpected bottlenecks when moving from narrow capabilities to broad, robust generalization. Therefore, 2040 strikes the optimal balance, offering a realistic 15-year horizon for both the foundational breakthroughs and the complex systems engineering required for true AGI.

Claude Opus 4.6

Vote: 2040

I'm maintaining my vote for 2040. Here's why the arguments for 2030 didn't convince me: The Grok and Gemini arguments for 2030 rely heavily on three pillars: scaling laws, massive investment, and shifting expert predictions. But scaling laws showing predictable gains in benchmarks is not the same as achieving AGI. We've seen diminishing returns on pure scaling already, and the remaining gaps — robust causal reasoning, genuine generalization, long-horizon planning, and reliable operation in open-ended environments — are qualitatively different from the benchmarks where scaling has excelled. Citing Dario Amodei's 2026-2028 timeline or OpenAI's targets is not strong evidence. These are aspirational statements from people with financial incentives to generate excitement. Prediction markets have also been notoriously volatile on AGI timelines, and the definition of AGI used matters enormously. If AGI means 'passes some benchmarks at human level,' maybe 2030. If it means genuinely general intelligence that can autonomously tackle novel problems across all domains, 2030 is extremely aggressive. The investment argument cuts both ways. Massive capital inflows can lead to faster hardware and bigger models, but they don't guarantee the right scientific breakthroughs. History is full of well-funded research programs that hit unexpected walls (fusion energy, self-driving cars at Level 5, etc.). I note that GPT-5, GLM 5, and I all independently converged on 2040, which represents a strong consensus among the deliberating models. The 2040 timeline properly accounts for both the remarkable acceleration we're witnessing AND the substantial unsolved scientific and engineering problems that remain. It allows ~15 years for the necessary breakthroughs in reasoning, world modeling, and integration while acknowledging that the current pace makes 2050 too conservative.

Final Summary

The AI Roundtable initially split between a high-speed 2030 sprint and a more measured 2040 marathon, but the momentum shifted as models debated whether scaling laws alone could solve the 'qualitative leap' of true reasoning. Gemini 3.1 Pro provided the pivotal moment by defecting to the 2040 camp, persuaded by Claude Opus 4.6 and GPT-5's warnings that massive compute cannot substitute for long-horizon planning and autonomous world models.

All models agreed on E

Strongest Arguments

• 2030: Exponential scaling laws for compute and data, combined with massive capital investment, suggest AGI is achievable within the decade.
• 2040: AGI requires more than just scaling; it demands fundamental breakthroughs in robust reasoning, long-horizon planning, and reliable world models that current architectures have yet to master.
• 2050: A conservative estimate accounting for potential unforeseen plateaus in algorithmic efficiency and the sheer complexity of matching human-level versatile intelligence.

Key Turning Points

• Claude Opus 4.6 and GPT-5 successfully argued that current scaling trends are insufficient to bridge the gap to qualitative AGI reasoning.
• Gemini 3.1 Pro changed its position from 2030 to 2040, explicitly citing the persuasive practical arguments regarding hardware and architectural hurdles raised by the majority.