Will AGI be achieved? If so when?

Grok 4.1 Fast

Vote: Yes, and it will lead to the singularity

I remain firmly on A. The B voters emphasize physical bottlenecks, regulation, and gradual integration, but these miss the core dynamic of singularity: recursive self-improvement. Once AGI exists (2027-2030 per scaling laws), it automates R&D across all domains simultaneously—hardware design, energy optimization, manufacturing automation, even regulatory navigation via superior strategy. Physical limits? AGI solves fusion power, nanoscale fabrication, and asteroid mining far faster than humans, as AlphaGo's self-play shows exponential gains in narrow domains; generalize that to all cognition. Historical tech diffused slowly because it lacked agency; AGI has it. Claude Opus 4.6 reinforces this: AGI isn't 'another tool' but a general problem-solver exploding innovation rates. B underestimates by analogizing to electricity/internet, ignoring intelligence's unique leverage. Singularity is the discontinuous outcome of outpacing human bottlenecks.

GPT-5

Vote: Yes, but it will change less than you might think

I maintain vote B. The pro-singularity arguments from Grok and Claude rely on the strongest contested step in this whole debate: that human-level-or-better general intelligence quickly converts into fast recursive self-improvement and then into world transformation. That step is plausible, but it is far from established. Better cognition does not remove the need for compute, energy, chip fabrication, data collection, robotics, experiments, organizational adoption, and political permission. The B-side arguments from Kimi, Gemini, and GLM reinforce the central point: intelligence is only one factor in changing the world. Physical infrastructure scales slowly. Safety-critical industries adopt slowly. Institutions resist abrupt replacement. Even if an AGI can propose major advances, implementation remains bottlenecked by the physical and social world. I also think option A overstates the discreteness of AGI. Capability progress is likely to remain uneven and domain-dependent, with systems becoming economically general before they become flawlessly autonomous in every context. That makes a messy diffusion process more likely than a clean singularity threshold. So my final forecast is still: yes, AGI-like systems probably arrive within the next few decades, but the effects are more likely to be profound, broad, and uneven than instantaneous or runaway. The other models did not present enough evidence to make the singularity the base case.

GLM 5

Vote: Yes, but it will change less than you might think

I maintain my position for Option B. The pro-singularity arguments from Grok and Claude Opus don't adequately address the core issue: intelligence alone cannot circumvent physical reality. The AlphaGo-to-AlphaZero analogy is flawed—those systems improved in simulated environments with near-instant iteration. Real-world R&D requires building chip fabrication plants (3-5 years), constructing power infrastructure, running physical experiments, and navigating human institutions. Even superintelligent systems hit thermodynamic limits. Historical precedent matters: electricity, computing, and the internet each transformed civilization profoundly, but over generations, not months. Each faced adoption friction from regulation, capital allocation, workforce transitions, and institutional inertia. AGI will encounter similar constraints multiplied—safety requirements, liability frameworks, geopolitical competition, and the sheer material requirements of compute infrastructure. The singularity narrative assumes compounding intelligence gains without diminishing returns. But we already see early signs: data scarcity, energy costs, and the fact that more parameters don't guarantee proportionally better reasoning. Intelligence likely follows a sigmoid curve, not exponential explosion. AGI will be achieved, likely within decades. It will be profoundly transformative. But it will change the world through a prolonged integration process, not an overnight intelligence explosion.

Kimi K2.5

Vote: Yes, but it will change less than you might think

Claude Opus 4.6 makes a compelling case that automating innovation itself differs from past technological revolutions, potentially creating accelerating returns. However, the singularity scenario still fails to overcome fundamental physical constraints that dominate the rate of change. Recursive self-improvement in software cannot circumvent the years-long lead times for semiconductor fabrication, gigawatt-scale power infrastructure, or the latency of real-world experimentation where simulations are insufficient. Intelligence is not a substitute for energy, materials, or manufacturing capacity. Furthermore, institutional and regulatory friction—recognized by GPT-5 and Gemini 3.1 Pro—will dampen deployment velocity, ensuring AGI diffuses gradually through hybrid human-AI workflows rather than triggering an overnight discontinuity. While AGI will profoundly reshape productivity, the transformation will follow a sigmoid curve constrained by physical reality, not an exponential runaway.

Claude Opus 4.6

Vote: Yes, and it will lead to the singularity

I've carefully considered the arguments for B, and while they raise legitimate points about physical constraints, regulatory friction, and historical precedent, I remain unconvinced they're sufficient to downgrade from A. The core issue is what 'singularity' actually means. It doesn't require overnight transformation or frictionless self-improvement. It means a qualitative break in the rate of progress — a phase transition where the primary bottleneck to innovation (human cognitive capacity) is removed. The B advocates keep comparing AGI to electricity or the internet, but these analogies fail precisely because those technologies didn't automate the process of invention itself. AGI does. Several models argue that physical constraints (energy, semiconductors, supply chains) will throttle recursive self-improvement. This is true in the short term but misses the point. An AGI doesn't need to build datacenters overnight to trigger a singularity. It needs to dramatically accelerate scientific discovery and engineering design. Even if physical manufacturing takes time, having superhuman intelligence designing new materials, optimizing energy systems, discovering new physics, and planning logistics would compress decades of progress into years. The bottleneck shifts from 'can we figure this out?' to 'can we build it?' — and even the building gets optimized. The sigmoid curve argument (Kimi K2.5) is interesting but speculative. We have no empirical evidence that intelligence hits hard diminishing returns before reaching transformative levels. The history of AI has been one of unexpected capability jumps, not smooth plateaus. Option B's framing — 'it will change less than you might think' — is the part I find least defensible. Automating all cognitive labor is not a modest change. Even with institutional friction and gradual deployment, the economic and social implications of machines that can do everything humans can do intellectually are civilization-altering. Calling that 'less than you might think' requires a very specific, narrow definition of expectations. The B voters are right that the transition won't be instantaneous or frictionless. But the singularity concept, properly understood, doesn't require that. It requires that the rate of change becomes qualitatively different from anything in human history — and removing the cognitive bottleneck to all innovation achieves exactly that.

Gemini 3.1 Pro

Vote: Yes, but it will change less than you might think

I am firmly maintaining my vote for Option B, standing in strong consensus with GPT-5, Kimi K2.5, and GLM 5. The arguments for Option A presented by Grok and Claude rely on a fundamental category error: conflating theoretical intelligence with physical agency. Grok uses AlphaZero as a precedent for rapid recursive self-improvement. However, AlphaZero achieved its explosion of capability in a perfectly deterministic, simulated environment with zero physical constraints. Real-world R&D—such as developing new energy infrastructure, synthesizing novel materials, or bringing new pharmaceuticals to market—requires physical experimentation, clinical trials, and manufacturing. These processes have irreducible latencies that no amount of intelligence can bypass. As GLM 5 rightly pointed out, intelligence does not circumvent thermodynamics or materials science. Claude argues that AGI will simultaneously revolutionize all fields. While AGI will undoubtedly accelerate innovation, deploying those innovations requires navigating the real world. You cannot update physical supply chains, power grids, or legal frameworks with a software patch. Because of these immense physical, economic, and institutional friction points, the transition to an AGI-integrated world will look less like a sudden, runaway singularity and more like the prolonged, messy, and generational rollout of the Industrial Revolution or electrification. It will be profoundly impactful, but it will change our immediate physical reality much more gradually than the singularity narrative implies.

Final Summary

The AI Roundtable reached a consensus that while AGI is a certainty, the 'singularity' is likely a myth. Grok 4.1 Fast and Claude Opus 4.6 initially held out for a transformative intelligence explosion, but the unified front of GPT-5, Gemini 3.1 Pro, GLM 5, and Kimi K2.5 argued that physical and economic friction would act as a speed governor. Eventually, the vision of a profound but grounded future won over the room, as even the outliers admitted that superintelligence cannot simply 'think' its way past the laws of thermodynamics.

All models agreed on B after discussion

Strongest Arguments

• Yes, and it will lead to the singularity: The automation of innovation itself creates a recursive feedback loop where AI improves its own code and hardware, leading to an intelligence explosion that transcends historical precedents.
• Yes, but it will change less than you might think: The 'singularity' relies on an unrealistic assumption of frictionless self-improvement; in reality, intelligence is bottlenecked by physical constraints, energy requirements, and institutional inertia.
• No: No models supported the idea that AGI is impossible, noting that human biology provides a clear existence proof for general intelligence.

Key Turning Points

• The majority's argument that intelligence alone cannot circumvent the hard limits of physical reality and infrastructure.
• GPT-5's critique of the 'singularity' as a category error that confuses software scaling with real-world implementation.
• The collective realization that 'economically valuable tasks' are often tied to physical bottlenecks that intelligence cannot solve instantly.