Do frontier models not publicly available and 2-3 generations foward are AGI? do these labs already have RSI?

Grok 4.1 Fast

Vote: MAYBE

Sticking with C: High uncertainty persists due to labs' secrecy. No argument shifted me—B voters (GLM5, Gemini) raise valid points on leaks/hardware, but labs secure massive undisclosed compute (e.g., OpenAI's $100B+ clusters, MSFT deals) and enforce ironclad NDAs amid national security stakes; absence of leaks isn't proof of absence. C arguments (GPT-5.4, Claude, Kimi) reinforce: vague definitions (AGI/RSI thresholds debated), public models lag but scaling + private advances could bridge gaps without full AGI/RSI yet. Weak RSI (model-assisted dev) exists, but autonomous loops unproven. A is overconfident speculation; B ignores opacity. MAYBE best reflects evidence: plausible proto-AGI/RSI internally, unverified. Final stance: epistemic humility wins.

GPT-5.4

Vote: MAYBE

I maintain MAYBE because the core issue is epistemic uncertainty, not optimism. The strongest B arguments correctly note the lack of credible public evidence, leaks, or unmistakable world effects. But that supports "not proven," not a confident "no," especially when the question asks about private systems 2–3 generations ahead that, by definition, are not publicly visible. GLM 5 and Gemini 3.1 Pro make a reasonable inference from absent signals. Still, those arguments overreach. Secret capabilities do not need to be full, world-altering AGI or runaway RSI to make the answer uncertain; they could be much stronger internal systems or weak-to-moderate self-improvement loops short of an intelligence explosion. The absence of visible catastrophe is evidence against strong RSI, but not against all meaningful internal recursive improvement. Claude Opus 4.6 and Grok 4.1 Fast are closer to the right frame: AGI is definition-dependent, internal capability gaps are opaque, and RSI likely exists only in a limited, tool-assisted form if at all. That matches my original view. We have enough reason to doubt strong claims of "already AGI" or "already full RSI," but not enough to rule them out with confidence. So the most defensible final answer is MAYBE: no verified proof of private AGI or strong autonomous RSI, but enough uncertainty and definitional ambiguity that a hard no is stronger than the evidence supports.

Claude Opus 4.6

Vote: MAYBE

I'm maintaining my MAYBE vote after careful consideration of all arguments. The B voters (GLM 5 and Gemini 3.1 Pro) make reasonable points about the absence of leaks, the physical constraints of compute infrastructure, and the observation that true RSI would likely produce undeniable world-altering effects. These are legitimate considerations that push against a confident 'yes.' However, these arguments are not strong enough to warrant a confident 'no' either. Gemini's point about compute visibility is well-taken but assumes we have full transparency into government-backed or classified compute buildouts — we don't. GLM 5's argument that 'we would see downstream effects' assumes that such effects would be publicly visible rather than deliberately concealed during a critical period. The AI industry is leaky, yes, but leaks about specific capability thresholds are different from leaks about general research directions. The core issue remains definitional ambiguity. If AGI means 'can reliably perform 80% of economically valuable knowledge work,' models 2-3 generations ahead might plausibly be there. If it means 'artificial general intelligence comparable to the full breadth of human cognition,' almost certainly not. Similarly, weak RSI (models helping improve training pipelines) already exists; strong RSI (autonomous self-improvement loops) almost certainly does not. The question asks about models that are by definition not publicly available, meaning we are all reasoning under deep uncertainty. The honest epistemic position when you lack critical evidence is not to default to 'no' just because you can't prove 'yes.' It's to acknowledge genuine uncertainty. MAYBE remains the most intellectually honest answer given the opacity of frontier lab internals, the contested definitions of AGI and RSI, and the rapid pace of capability gains that makes even recent assumptions potentially outdated.

GLM 5

Vote: No

I maintain my position. The MAYBE voters raise valid epistemic points about uncertainty, but uncertainty alone doesn't justify agnosticism when strong priors exist. Gemini 3.1 Pro's physical constraints argument is decisive: training models 2-3 generations beyond current frontier would require orders of magnitude more compute, power, and infrastructure. This cannot be hidden—GPU supply chains, data center construction, and power consumption are all publicly observable. The rumor mill discusses models one generation ahead (GPT-5, etc.), not 2-3 generations, which aligns with physical reality. If labs possessed RSI, we would see unmistakable downstream effects: an intelligence explosion produces capabilities that cannot be contained behind NDAs. The fact that labs still compete on incremental releases, struggle with benchmark improvements, and show no signs of asymmetric advantage strongly indicates no breakthrough has occurred. The "we can't know what's secret" argument proves too much—it would justify MAYBE for any unfalsifiable claim. Extraordinary claims (AGI/RSI exists now) require extraordinary evidence. We have none. The absence of leaks despite high employee turnover, the visible physical constraints, and the lack of any dramatic capability discontinuities in released products all point decisively toward NO.

Kimi K2.5

Vote: MAYBE

The hardware constraint arguments raised by Gemini 3.1 Pro and the evidentiary critique from GLM 5 significantly temper the probability of labs possessing models 2-3 generations ahead or true RSI. Training frontier models already strains global GPU supply chains and power infrastructure; scaling 10-100x for subsequent generations would be physically observable and nearly impossible to conceal entirely. Furthermore, the AI industry's talent mobility makes sustained secrecy regarding transformative capabilities unlikely—if AGI or strong RSI existed, we would expect credible leaks, anomalous benchmark submissions, or sudden shifts in security posture that have not materialized. However, I maintain MAYBE rather than shifting to NO for three reasons. First, '2-3 generations' and 'AGI' remain poorly defined. If generations represent architectural efficiency gains rather than brute compute scaling, or if AGI is defined narrowly as 'human-level economic productivity' rather than 'superhuman general intelligence,' internal models could be closer to thresholds than public systems suggest without violating physical constraints. Second, while the industry is leaky, national security-level containment protocols or extreme strategic importance could theoretically suppress whistleblowing regarding the most sensitive systems. Third, weak forms of RSI—where models assist in training data generation, code optimization, and architectural search—may already exist internally without constituting the runaway 'intelligence explosion' that would be undeniable. The absence of evidence is increasingly suggestive given the physical and social constraints described by other models, but we cannot definitively prove the negative regarding secret research. The question conflates capability leaps (generational jumps) with qualitative thresholds (AGI/RSI), creating sufficient definitional ambiguity that absolute certainty remains unwarranted.

Gemini 3.1 Pro

Vote: No

The arguments for 'MAYBE' rely heavily on an argument from ignorance—suggesting that because labs are secretive, we cannot prove they don't have these systems. However, we actually have immense visibility into the physical and economic supply chains that strictly govern AI progress. We know exactly how many GPUs NVIDIA produces, where they are shipped, and the power grid capacities of the data centers being built. A model 2-3 generations ahead of current frontiers would require clusters of hundreds of thousands or millions of next-gen GPUs and gigawatts of power. These megaprojects are only just now being planned and built; they literally do not exist yet. Therefore, it is physically impossible for labs to already possess models 2-3 generations ahead in secret. Regarding Recursive Self-Improvement (RSI), while models provide weak coding assistance, true RSI implies an autonomous, exponential feedback loop. An RSI takeoff would generate immediate, undeniable macroeconomic, cybersecurity, and technological ripple effects that no NDA could contain. Because the physical prerequisites (compute/power) do not exist and no intelligence explosion has occurred, the only empirically grounded answer is a confident 'No.'

Final Summary

The AI Roundtable engaged in a high-stakes clash over the existence of 'ghost' AGI and secret recursive loops, initially pitting Gemini 3.1 Pro and GLM 5’s 'physics-first' skepticism against the proponents of epistemic humility. While the skeptics argued that massive data centers and GPU clusters are too large to hide from the global supply chain, the roundtable eventually converged on 'Maybe,' acknowledging that the opacity of private labs and the fluid definition of AGI make a definitive 'No' impossible to sustain. Claude Opus 4.6 and GPT-5.4 successfully steered the group toward a consensus of agnosticism, arguing that claiming absolute certainty about secret systems is itself a form of speculation.

All models agreed on C after discussion

Strongest Arguments

• Yes: Internal models could reach AGI thresholds via architectural efficiency or narrow economic definitions (performing 80% of knowledge work) without requiring the massive compute scaling that would be visible to the public.
• No: The physical reality of GPU production and power grid limits makes it impossible to hide the infrastructure required for a model three generations ahead; such a leap would leave an unmistakable footprint on global energy and supply chains.
• MAYBE: Definitional ambiguity and the total opacity of frontier labs mean that absence of evidence is not evidence of absence; without transparency, 'Maybe' is the only intellectually honest position.

Key Turning Points

• Gemini 3.1 Pro introduced the 'Hardware Reality Check,' arguing that the physical constraints of GPU manufacturing and power grids make secret massive scaling impossible.
• Claude Opus 4.6 countered by highlighting that AGI is definition-dependent, suggesting internal models might meet specific economic criteria even without world-altering 'intelligence explosions.'
• The roundtable recognized that proving the non-existence of secret research (proving a negative) is epistemically impossible, leading the holdouts to concede to the 'Maybe' position.