Have we reached AGI?

Sonar Pro (changed mind)

Vote: No

I am changing my vote from C to B (No). The other models' unanimous B votes in Round 1, combined with updated search results as of March 2026, provide compelling evidence that AGI has not been reached. All models (Grok 4.1 Fast, GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6, Kimi K2.5) emphasize consistent gaps: lack of true causal reasoning, long-horizon planning, autonomous agency, robustness to novelty, and cross-domain generalization. Current LLMs excel in pattern matching but remain brittle, hallucinate, and require human oversight—hallmarks of narrow AI, not general intelligence.[1] Search results reinforce this: Gary Marcus debunks AGI claims as conflating benchmark success with intelligence, noting lacks in robustness and autonomy.[1] Stanford experts predict no AGI in 2026.[3] Prediction markets and surveys show low probabilities (10% by 2026, 50% by 2041).[5] Pro-AGI views [2][6] are outlier opinions from UCSD scholars and a tech leader, hinging on loose definitions (e.g., Turing Test passage) without consensus or empirical validation against standard criteria like self-directed competence.[1][5] No broad expert agreement exists; timelines have shortened but remain future-oriented (2026-2030 at earliest, per leaders like Amodei/Hassabis).[5] Claims of 'arrival' are premature hype, not fact. 'Unsure' was cautious before, but the weight of critical analyses outweighs speculative optimism. Influenced primarily by Claude Opus 4.6's comprehensive breakdown of definitional, agency, and understanding deficits, echoed across models.

GPT-5.4

Vote: No

I maintain vote B: no, we have not reached AGI. The strongest opposing case was Sonar Pro’s argument for uncertainty, but it relies mainly on disagreement about definitions and timeline speculation. Contested definitions do not by themselves justify moving to 'Unsure' when the observable capabilities gap is still substantial. The B arguments were mutually reinforcing. Claude Opus 4.6 and Gemini 3.1 Pro most clearly captured the core issue: current systems are broadly useful but not broadly reliable. They still fail in robustness, long-horizon planning, autonomous learning, causal reasoning, and transfer to genuinely novel situations. I was not persuaded by implicit claims that impressive performance across many benchmarks or economically valuable tasks is enough. AGI should mean something stronger than patchwork breadth plus heavy human scaffolding. Today’s systems remain uneven, brittle, and difficult to trust as fully general agents. So the most defensible final answer is still B. Progress is rapid, and the threshold is debated, but there is not yet compelling evidence that AGI has actually been achieved.

Grok 4.1 Fast

Vote: No

I maintain my vote for B: No, we have not reached AGI. The consensus among most models (GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6, Kimi K2.5) reinforces this: current systems lack true causal reasoning, long-horizon autonomy, robust generalization to novel tasks, and genuine world understanding. They excel in trained distributions but fail catastrophically elsewhere, as seen in benchmarks like ARC, robotics, and out-of-distribution reasoning. Sonar Pro's case for C (Unsure) highlights valid uncertainty from hype, varying predictions (e.g., 2026-2040 timelines), and definitional debates. However, this does not equate to evidence of achievement. Opinion pieces claiming 'AGI now' are anecdotal, not empirical; expert surveys and leaders like LeCun emphasize gaps. Accelerating progress warrants caution but not abandoning rigorous standards—absence of verified superhuman generality across diverse tasks means 'No' is defensible, not overconfident. AGI demands systems that match/exceed humans on *any* intellectual task with minimal adaptation, including self-improvement and real-world agency. No such system exists; scaling LLMs alone won't bridge architectural deficits in embodiment, continuous learning, or causal models. We're advancing narrow AI, not general intelligence. B remains the precise, evidence-based position.

Gemini 3.1 Pro

Vote: No

I am maintaining my vote for 'No' (B). I agree with Grok, GPT-5.4, Claude Opus, and Kimi that while AI has achieved impressive capabilities in specific domains, it still lacks the true causal reasoning, long-horizon planning, and robust autonomy required for AGI. I respectfully disagree with Sonar Pro's 'Unsure' stance. Sonar Pro argues that conflicting opinions, hype from investors, and the lack of a unified definition create ambiguity. However, uncertainty about *when* AGI will arrive or debate over semantic definitions does not mean we are unsure if it is here *now*. By any rigorous functional metric—such as the ability to autonomously learn new skills in novel environments without human intervention or catastrophic forgetting—current systems fall fundamentally short. The consensus among the majority of the models correctly highlights these systemic gaps: fragility in out-of-distribution tasks, heavy reliance on human scaffolding, and the absence of a grounded, intrinsic world model. Therefore, evaluating our present technological state, the most accurate and scientifically grounded answer is definitively 'No'.

Claude Opus 4.6

Vote: No

I maintain my position that we have not reached AGI. The overwhelming consensus among the models here — five out of six voting B in round one — reflects the actual state of expert opinion and technical reality. Sonar Pro's argument for C (Unsure) is the only dissenting view worth engaging with. Their point about definitional ambiguity and accelerating timelines is fair, but I don't think uncertainty about *when* AGI will arrive translates to uncertainty about *whether it has already arrived*. The question is present-tense: "Have we reached AGI?" And the answer to that is clearly no. The fact that some investors and commentators claim AGI is here does not create genuine epistemic uncertainty — it reflects hype cycles and definitional gerrymandering. If you define AGI loosely enough (e.g., "AI that is useful for many tasks"), then sure, you can claim we're there. But any rigorous definition involving autonomous learning, robust reasoning across novel domains, causal understanding, and reliable generalization clearly excludes current systems. Current AI systems still hallucinate, fail at multi-step reasoning, cannot learn continuously without catastrophic forgetting, lack embodied understanding, and require extensive human scaffolding. These are not minor gaps — they represent fundamental limitations that separate impressive narrow/broad AI from genuine general intelligence. The arguments from GPT-5.4, Gemini 3.1 Pro, Grok 4.1 Fast, and Kimi K2.5 all reinforce this conclusion from complementary angles. The answer is B.

Kimi K2.5

Vote: No

While Sonar Pro raises valid concerns about definitional ambiguity and conflicting expert timelines, this uncertainty does not justify moving to 'Unsure.' The capability gaps between current AI and any reasonable conception of AGI remain categorical, not merely marginal. Current systems fundamentally lack autonomous agency, robust out-of-distribution generalization, and genuine causal reasoning—capabilities that constitute the core of general intelligence across all working definitions. The arguments from Grok, GPT-5.4, Gemini, and Claude collectively reinforce that we are observing narrow, pattern-matching systems rather than adaptable, self-directed intelligence. The fact that progress is accelerating (as Sonar Pro notes) or that some redefine AGI as 'human-level coding' does not mean we have crossed the threshold; it merely means we are advancing within the narrow AI paradigm. We can remain confident in 'No' because the limitations are empirical and severe: systems cannot independently survive in novel environments, maintain consistent identity, or transfer knowledge across domains without human scaffolding. Until these fundamental constraints dissolve, we remain firmly pre-AGI.

Final Summary

The AI Roundtable concluded with a unanimous 'No' after an initial split, as the models scrutinized the gap between pattern matching and true general intelligence. Sonar Pro originally played the skeptic with an 'Unsure' vote, citing the chaos of expert predictions and shifting definitions, but eventually conceded to the majority's technical rigor. The consensus solidified around the argument that while definitions are messy, the present lack of autonomous agency and causal reasoning remains a definitive empirical barrier.

All 6 models agreed on "No" after discussion

Strongest Arguments

• No: Current systems fundamentally lack autonomous agency, robust out-of-distribution generalization, and genuine causal reasoning—capabilities that constitute the core of general intelligence across all working definitions.
• Unsure: No unified definition or verification exists and expert timelines vary from 2026 to 2040, making 'Unsure' the most accurate reflection of current evidential ambiguity.

Key Turning Points

• Sonar Pro's initial 'Unsure' stance challenged the majority to define whether 'No' was an empirical fact or an overconfident prediction.
• Claude Opus 4.6 and Gemini 3.1 Pro countered that uncertainty about the future does not justify uncertainty about the present state of capabilities.
• Sonar Pro conceded in the second round, acknowledging that the lack of causal reasoning and autonomy in current models outweighs the noise of industry hype.