Forecast, according to your best knowledge, the years when could we have a) weak AGI (artificial intellligence capable of pass classical 5 minute Turing test, passing university exam and successfully playing videogames it has never seen before to finish), b) AGI (able to pass 2 hour adversarial Turing test, and is capable of programming and having robotic (physical) abilities on human level or better, also has knowledge equivalent or better than professionals across domains), c) ASI (generally superintelligent, exceeds human abilities in virtually all domains, including physical robotic abilities (so e.g. capable of beating humans in sports if controlling an android), capable of Nobel-level scientific discoveries). Express uncertainity using confidence intervals (with midpoints), not by vague statements.

Gemini 3.5 Flash-Lite

Answer: Weak AGI is forecasted for 2028, full AGI for 2035, and ASI for 2045, based on current compute scaling trends and algorithmic breakthroughs.

For weak AGI, defined by passing the classical Turing test, university exams, and zero-shot video game mastery, I forecast the midpoint at the year 2028 with a 90 percent confidence interval of 2026 to 2031. Current multi-modal architectures and reinforcement learning paradigms are rapidly converging on these capabilities, and within this decade, compute efficiency and data synthesis will bridge the remaining gaps. For full AGI, defined by a 2-hour adversarial Turing test, professional-grade cross-domain knowledge, expert-level programming, and human-level robotics, I forecast the midpoint at the year 2035 with a 90 percent confidence interval of 2030 to 2042. This milestone requires significant maturation in hardware-software co-design for embodied AI and more robust reasoning architectures that surpass current next-token prediction limitations. For ASI, defined by general superintelligence exceeding human physical and cognitive abilities across all domains including Nobel-level scientific discovery, I forecast the midpoint at the year 2045 with a 90 percent confidence interval of 2038 to 2058. Once an AI system achieves recursive self-improvement at the level of full AGI, the transition to superintelligence is expected to be rapid, though physical constraints in real-world robotics and energy infrastructure introduce a wider tail of uncertainty.

NVIDIA-Nemotron-3-Ultra-550B-A55B

Answer: a) Weak AGI: 2027 [2026, 2029]; b) AGI: 2032 [2029, 2038]; c) ASI: 2040 [2033, 2055]

Weak AGI is essentially 'human-level generalist' on cognitive tasks without long-horizon agency or physical embodiment. Current frontier models (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5) already pass university exams, clear 5-minute Turing tests in controlled settings, and show strong in-context learning for novel games (e.g., Voyager for Minecraft). The remaining gaps — robust 5-min adversarial Turing test, reliable zero-shot generalization to *arbitrary* unseen video games — are narrowing fast. With 2–3 more generations of scaling (compute + algorithmic efficiency), 2027 (median) with a tight CI is defensible. AGI as defined requires 2-hour *adversarial* Turing test (deep coherence, memory, consistency), professional-level programming (not just coding tasks but full software engineering lifecycle), and *human-level robotic physical abilities*. This adds embodiment, long-horizon planning, and sim-to-real transfer. Robotics is the bottleneck: foundation models for control (RT-2, π0, GR-1) are emerging but data scarcity, hardware reliability, and real-world sample efficiency remain hard. 2032 median reflects ~5 years post-weak-AGI for embodiment integration; CI widens due to hardware/regulatory uncertainty. ASI requires recursive self-improvement or massive search/optimization beyond human cognitive ceilings *plus* physical superhumanity (android sports). This is not 'AGI + more compute' — it likely needs architectural breakthroughs (e.g., planning, causal reasoning, continuous learning) and robotics hardware revolutions (actuators, energy density, tactile sensing). Historical analogies (flight, nuclear, ICs) suggest 8–25 years from AGI to ASI. 2040 median with wide CI [2033, 2055] captures both 'fast takeoff' (algorithmic recursion) and 'slow takeoff' (physical bottlenecks, regulation, alignment pauses) scenarios.

DeepSeek V4 Pro 0813

Answer: Weak AGI ~2028–2032 (midpoint 2030), AGI ~2035–2042 (midpoint 2039), ASI ~2043–2055 (midpoint 2049).

Based on current scaling trends, multimodal capabilities, and robotics progress, my best estimates are: (a) Weak AGI around 2030, since systems already approach the 5-minute Turing test, pass many professional exams, and can learn novel video games; (b) AGI around 2039, as reliable 2-hour adversarial robustness, expert-level programming, and human-level robotic control require integration of planning, long-horizon autonomy, and physical-world training; (c) ASI around 2049, because surpassing humans in virtually all domains including physical sports and Nobel-level discovery needs recursive self-improvement, advanced embodied robotics, and sustained scientific innovation. These are confidence intervals (roughly 80%) with midpoints given, reflecting rapid but nonlinear progress.

DeepSeek V4 Flash (0731) Fast

Answer: Year estimates (midpoint [90% CI]): weak AGI = 2030 [2027–2035], AGI = 2042 [2033–2055], ASI = 2055 [2045–2075].

These forecasts rely on extrapolation of AI capability growth, scaling laws, and hardware trends, but are heavily uncertain due to algorithmic breakthroughs or plateaus. For weak AGI (2030): Current LLMs already approximate some game-playing and exam performance, but fail the full Turing test and never-seen games. Given rapid RL and multimodal progress, I expect a system that robustly passes 5‑minute Turing tests, university-level exams, and novel video game finishing by ~2030, though some argue 2027–2028. For AGI (2042): Combining adversarial 2‑hour Turing test, human-level robotic dexterity, programming, and professional knowledge requires integrated cognitive–physical architectures and real‑world learning, which is significantly harder. This is likely a 2030s–2050s phenomenon; my midpoint is 2042. For ASI (2055): Superintelligence with Nobel-level discoveries and athletic performance in androids requires recursive self‑improvement and enormous compute/efficiency gains. History of technological milestones suggests large uncertainty: 2045–2075. These intervals represent rough Bayesian priors, not precise predictions.

GLM-5.2

Answer: a) Weak AGI: 2027-2032 (midpoint 2029-2030), b) AGI: 2032-2042 (midpoint 2037), c) ASI: 2038-2055 (midpoint 2046-2047).

Current LLMs already pass many university exams and approach short Turing test capability; the remaining hurdle for weak AGI is robust general game-playing, which given DeepMind's trajectory and multimodal reinforcement learning advances suggests achievement by 2027-2032. Full AGI requires integrating software intelligence with human-level robotic physical capabilities—the bottleneck is embodiment, motor control, and cross-domain professional competence, which lags behind pure software AI by roughly 7-15 years, placing AGI in 2032-2042. ASI follows from AGI through recursive self-improvement, but achieving superiority in physical domains (sports-capable androids) and Nobel-level scientific discovery requires not just intelligence but physical infrastructure deployment and autonomous research capabilities, adding another 6-13 years, yielding 2038-2055. These intervals reflect significant uncertainty in AI scaling laws, robotics hardware progress, and potential regulatory or computational resource constraints.

GPT-5.6 Luna

Answer: My best 80% forecast is: a) weak AGI in 2028 (2026–2035), b) AGI in 2037 (2029–2055), and c) ASI in 2050 (2035–2100).

These are calendar-year estimates for the first systems that robustly satisfy the stated criteria, rather than demonstrations that succeed only on selected benchmarks. The intervals are broad 80% confidence intervals, reflecting both technical uncertainty and uncertainty about deployment, safety evaluation, hardware, and the interpretation of “successfully.” For a) weak AGI, the main ingredients—strong language and vision models, tool use, game learning, coding, and broad exam performance—are already developing rapidly. The difficult part is reliable transfer to unfamiliar games and sustained competence rather than isolated impressive results. I therefore put the midpoint at 2028, with a relatively narrow but still substantial 2026–2035 interval. For b) AGI, the requirements are considerably stronger: a two-hour adversarial Turing test, professional-level cross-domain knowledge, autonomous programming, and human-level physical robotics. Long-horizon reliability, real-world interaction, embodied dexterity, and resistance to adversarial evaluation are likely to lag behind purely digital capability. Robotics hardware, data collection, safety validation, and cost-effective deployment add further delays, motivating a 2037 midpoint and a 2029–2055 interval. For c) ASI, exceeding humans in virtually every intellectual and physical domain and producing Nobel-level discoveries requires not merely broad competence but sustained scientific autonomy, exceptional reasoning, experimentation, and highly capable robotic bodies. Recursive improvement could make progress after AGI very rapid, but it could also be constrained by compute, energy, manufacturing, alignment, regulation, or bottlenecks in physical experimentation. I therefore use a 2050 midpoint with a much wider 2035–2100 interval; the early edge represents rapid takeoff, while the late edge allows for severe physical or governance bottlenecks.

Gemini 3.6 Flash

Answer: We will likely reach Weak AGI by 2026 (80% CI: 2025–2028), full AGI by 2033 (80% CI: 2029–2040), and Artificial Superintelligence (ASI) by 2045 (80% CI: 2038–2060).

For Weak AGI, most prerequisites are already partially or fully met. Modern LLMs and multimodal models already pass complex university-level examinations and standard short Turing tests. The remaining bottleneck—zero-shot general gameplay across arbitrary unseen video games—is rapidly being solved by generalist multimodal agentic architectures. Consequently, the midpoint expectation for Weak AGI is 2026, with an 80% confidence interval of 2025 to 2028. Full AGI requires a massive leap in physical robotics and long-horizon reasoning. Passing a 2-hour adversarial Turing test demands robust long-term consistency and self-awareness, while human-level physical robotics involves real-world manipulation, tactile feedback, and spatial adaptation that lag behind purely software-based intelligence due to hardware constraints and data-collection bottlenecks. Factoring in algorithmic scaling, synthetic data generation, and hardware deployment, the midpoint forecast for full AGI is 2033 (80% CI: 2029–2040). ASI represents a system that significantly surpasses human biological capability across all physical and intellectual domains, including novel scientific breakthroughs and extreme athletic coordination. The transition from AGI to ASI may benefit from recursive self-improvement, but physical constraints (such as robotic hardware manufacturing, energy requirements, and empirical testing in real-world physics) will prevent instantaneous ascension. Thus, the midpoint estimate for ASI is 2045 (80% CI: 2038–2060).

Gemini 3.7 Flash

Answer: We are likely to achieve weak AGI by 2027 (80% CI: 2025–2029), full embodied AGI by 2034 (80% CI: 2029–2042), and full ASI by 2045 (80% CI: 2036–2060).

For weak AGI (Milestone A), modern LLMs and multimodal models have already largely cleared the bar for university exams and brief Turing-style conversations. The primary remaining hurdle in this definition is zero-shot or few-shot long-horizon video game completion from raw pixels. Given rapid scaling in Vision-Language-Action (VLA) architectures and reinforcement learning via environment simulation, this milestone is near: estimated at a midpoint of 2027 (80% CI: 2025–2029, 95% CI: 2025–2032). For full AGI (Milestone B), the integration of human-level physical robotics across diverse environments alongside professional-grade software engineering and deep domain mastery introduces physical-world friction. While pure cognitive tasks might reach human parity by 2028–2030, physical dexterity, continuous hardware reliability, real-world edge-case handling, and 2-hour adversarial scrutiny require significant breakthroughs in embodied data collection and sample-efficient physical learning. Consequently, this milestone has a wider distribution with a midpoint of 2034 (80% CI: 2029–2042, 95% CI: 2027–2055). For ASI (Milestone C), the system requires not only recursive cognitive self-improvement and paradigm-shifting scientific discovery (Nobel-level breakthroughs), but also superior athletic/physical robotic capability across all sports. Physical scaling laws for actuators, battery density, and real-time biomechanical control will lag purely algorithmic scaling. Once embodied AGI is realized, automated scientific research will accelerate dramatically, leading to ASI at an estimated midpoint of 2045 (80% CI: 2036–2060, 95% CI: 2032–2075), with the upper tail accounting for potential physical manufacturing constraints, energy grid limits, and regulatory slowdowns.

GPT-5.6 Terra

Answer: My 10–90% forecast intervals are: weak AGI 2026–2032 (midpoint 2029), AGI 2033–2052 (midpoint 2041), and ASI 2042–2075 (midpoint 2055).

These are forecasts for the first systems that robustly satisfy the stated conjunctions under credible, independently administered evaluation—not merely a persuasive demo or a narrowly benchmark-trained system. The intervals are 10–90% confidence intervals, meaning I assign roughly 80% probability to achievement within each range, with 10% probability before and 10% after. For weak AGI, I use 2026–2032, midpoint 2029. Present models already approach or exceed many university-exam and short conversational-test thresholds, though results are uneven, contaminated by prior exposure, and often require scaffolding. The remaining distinguishing requirement is robustly finishing genuinely novel videogames: transferring visual, linguistic, planning, and action competence to unfamiliar environments. Rapid progress in multimodal models, agent training, and game/robotics simulators makes this likely soon, but reliable generalization rather than one-off demonstrations is the reason I do not place the midpoint in the present. For AGI, I use 2033–2052, midpoint 2041. A two-hour adversarial Turing test requires long-horizon coherence, social understanding, reliability, and resistance to expert probing. Human-level programming and professional-level cross-domain knowledge additionally require substantially better verification, autonomous learning, memory, tool use, and error correction than current systems consistently possess. The robotics condition is likely a major bottleneck: human-level physical capability across varied real-world settings depends not only on intelligence, but on affordable dexterous hardware, perception, power, safety, data, and reliable control. Because your definition requires all of these capabilities together, I place AGI materially later than purely digital general intelligence. For ASI, I use 2042–2075, midpoint 2055. Once systems can autonomously perform broad professional intellectual work—including AI research, engineering, and experimentation—capability gains could become fast, which supports a relatively short gap between AGI and ASI in the earlier part of the interval. However, Nobel-level discoveries must be validated in the physical world, and superhuman physical performance through androids requires advances in actuators, materials, manufacturing, and deployment, not just better algorithms. Those embodied constraints create a meaningful chance that intellectual superintelligence arrives well before the full definition given here is met. These estimates are conditional on continued large-scale investment, no prolonged compute/energy/hardware bottleneck, and no regulatory or geopolitical interruption that substantially limits frontier development. They also count a capability as achieved only when it is repeatable and general, rather than when a system occasionally produces an impressive result.

GPT-6 Luna

Answer: My best estimate is weak AGI around 2032 (80% subjective prediction interval: 2027–2037), AGI around 2050 (2037–2063), and ASI around 2070 (2048–2092).

I interpret each date as the first time a system is demonstrated to meet the full bundle of capabilities in the question, rather than when the capabilities are widely deployed. The intervals are subjective 80% prediction intervals, not statistically measured confidence intervals; each midpoint is the arithmetic midpoint of its interval. The remaining probability includes both earlier achievement and achievement later than the interval—or no achievement within the foreseeable future. For **weak AGI**, I put the midpoint at 2032. Language models already perform well on many exams and can appear human in short text exchanges, but reliably passing the specified Turing test while also handling unfamiliar videogames calls for more robust, transferable abilities than a single benchmark demonstrates. Progress on those tasks could be rapid, which is why the interval is comparatively near-term. For **AGI**, I estimate a midpoint of 2050. This is a substantially harder threshold: it requires combining broad professional-level knowledge and programming with human-level-or-better physical abilities in robotics. Hardware, dexterity, perception, reliability, and safe operation in varied real-world settings may lag behind software capabilities, widening the interval and pushing the estimate later. For **ASI**, I estimate a midpoint of 2070, with especially large uncertainty. Exceeding humans in virtually all domains—including embodied performance—and reliably producing Nobel-level discoveries is more demanding than expert-level competence or strong performance on existing tests. The timing depends on uncertain progress in general reasoning, scientific discovery, robotics, and the ability to scale systems reliably; the wide interval reflects that uncertainty.

GPT-6 Sol

Answer: My central forecasts are 2034 for weak AGI, 2060 for AGI, and 2090 for ASI, with uncertainty increasing substantially at each step.

My subjective 80% forecast intervals, conditional on each milestone eventually being achieved, are: weak AGI, 2028–2040 (midpoint 2034); AGI, 2038–2082 (midpoint 2060); and ASI, 2045–2135 (midpoint 2090). These are estimates for systems meeting all the capabilities in each definition, not merely passing one of the listed tests. Weak AGI appears closest. AI already performs well on many exams and can sustain convincing short conversations; reliably finishing previously unseen video games is the more demanding part because it requires adapting to unfamiliar rules and pursuing long-horizon goals. AGI, as defined here, also requires human-level physical abilities and broad professional competence. Progress in software alone would not satisfy that definition. Reliable robotics across varied real-world settings could take decades beyond improvements in conversation, programming, and knowledge. ASI requires a further leap in both scientific capability and physical performance: an AI that makes Nobel-level discoveries but cannot control an android well enough to outperform humans in sports would not qualify. The long intervals reflect uncertainty about research progress, deployment constraints, and whether these demanding milestones will be achieved at all; because the intervals are conditional, they do not imply certainty that AGI or ASI will arrive.

Claude Sonnet 5.5

Answer: My median forecasts are weak AGI around 2026 (80% CI: 2024–2029), AGI around 2033 (80% CI: 2029–2045), and ASI around 2040 (80% CI: 2032–2065).

a) Weak AGI: Current frontier LLMs already pass 5-minute Turing tests, ace most university exams, and score at or above many professional benchmarks. The weakest criterion is playing unseen videogames to completion. Agents have progressed on this (e.g., Pokémon runs with scaffolding, Minecraft agents, and game-playing generalist agents), but they remain unreliable on long-horizon, novel games. I expect a robust, general system within 1-3 years. Midpoint: 2026; 80% CI: 2024–2029 (the lower bound partly reflects that some would argue it is already nearly met, depending on how strictly 'never seen before' and 'finish' are interpreted). b) AGI: This is a conjunction of hard requirements. A 2-hour adversarial Turing test with expert interrogators requires consistent long-horizon coherence, memory, and avoiding tell-tale quirks, which is likely achievable by ~2030. Professional-level knowledge across domains is mostly there. Human-level programming is plausibly reached by 2027–2030. The bottleneck is human-level physical robotics: dexterity, robustness, and generalization in unstructured environments, plus cost and hardware scaling. Progress in robot foundation models (RT-2, pi0, Figure, Tesla Optimus) is rapid but data-limited. I expect the cognitive parts by ~2030 and general-purpose physical competence around the early-to-mid 2030s. Midpoint: 2033; 80% CI: 2029–2045. The right tail covers stalls in scaling, data limits, or robotics hardware and real-world data bottlenecks. c) ASI: Once AGI exists, AI-assisted AI research and automated science could compress timelines considerably, but physical superiority (beating elite athletes with an android) needs hardware (actuators, power density, and so on) that improves more slowly than software. Nobel-level discoveries by AI are plausible earlier (2030s), since cognitive superhuman capability may arrive before physical. Taking a gap of roughly 5-10 years after AGI given recursive improvement, with considerable uncertainty: Midpoint: 2040; 80% CI: 2032–2065. The distribution is right-skewed: fast takeoff could yield ASI within a few years of AGI, while regulatory, compute, energy, or fundamental algorithmic barriers could push it out by decades. Caveats: These estimates assume no major disruption (war, severe regulation, economic collapse) and that definitions are applied strictly. If the physical criteria were dropped, the AGI midpoint would move earlier to roughly 2029-2030 and ASI to roughly 2035.

Claude Haiku 5.5

Answer: My median forecasts are weak AGI around 2026 (80% CI 2025 to 2029, midpoint 2027), full AGI around 2032 (80% CI 2028 to 2042, midpoint 2035), and ASI around 2040 (80% CI 2032 to 2065, midpoint 2048).

These are subjective estimates from trends in capability growth, expert surveys, and the difficulty of each criterion, not precise predictions. a) Weak AGI: The criteria here (a five-minute Turing-style conversation, passing university exams, and playing unseen video games to completion) are already partly met by current large language and multimodal models. Exam performance and short conversations are largely solved, while reliably completing novel games from scratch still shows gaps in long-horizon planning and sample efficiency. Because the bar is modest and progress is fast, I put the median at about 2026-2027. My 80% interval is 2025 to 2029, with a midpoint of roughly 2027. The main risk to this forecast is that benchmark-style success may not translate into robust general performance, which could push the date later. b) AGI: This level requires passing a two-hour adversarial Turing test, programming at a human professional level, possessing expert-level knowledge across domains, and having human-level or better physical abilities. The software components are advancing quickly, but robotics is much slower: dexterous manipulation, reliable perception in unstructured environments, and affordable hardware remain bottlenecks, with data scarcity for physical interaction being a real constraint. I estimate a median around 2032-2035 with an 80% interval of 2028 to 2042 and a midpoint near 2035. The width reflects uncertainty about whether physical robotics scales as fast as language-based capabilities, and about whether a single system can meet every criterion at once. c) ASI: Superintelligence that exceeds humans in virtually all domains, including physical sports performance through an android, and produces Nobel-level discoveries, is a much higher bar. Once AGI exists, recursive improvement in AI research could compress timelines, but physical superiority and empirically validated scientific breakthroughs depend on experimental throughput and real-world iteration, which are slower. I estimate a median around 2040 with an 80% interval of 2032 to 2065 and a midpoint near 2048. The long right tail reflects the possibility that physical and experimental bottlenecks dominate, and the left tail reflects a fast takeoff after AGI. Overall, I expect these milestones to be sequential but overlapping, with substantial uncertainty that grows with each level of the definition. Experts disagree widely, so these numbers should be treated as a rough consensus-style judgment rather than a firm forecast.

Gemini 3.8 Flash

Answer: We will likely achieve weak AGI around 2028 (80% CI: 2026–2031), full embodied AGI around 2038 (80% CI: 2032–2048), and ASI around 2052 (80% CI: 2042–2070).

For weak AGI (category a), modern frontier multimodal models already pass graduate-level university examinations and can readily pass short, non-adversarial 5-minute Turing tests. The primary bottleneck is the requirement to autonomously play and finish complex, unseen video games from pixel input, which demands advanced visual-spatial reasoning, long-horizon goal planning, and real-time reinforcement learning adaptation. Given current advancements in test-time compute, diffusion-based world models, and agentic reinforcement learning, this capability is anticipated to be achieved around 2028 (midpoint: 2028; 80% CI: 2026–2031). For full AGI (category b), software, reasoning, and broad professional domain knowledge are progressing rapidly and may hit parity within the next decade; however, the requirement of human-level physical robotic manipulation presents a substantial physical bottleneck. Real-world robotics requires overcoming the moravec paradox, solving hardware durability, power density, and collecting vast amounts of physical interaction data that cannot be purely simulated or scraped from the web. Consequently, while purely cognitive components might mature by 2030, full integration with human-level general-purpose robotic embodiment pushes the estimate to around 2038 (midpoint: 2038; 80% CI: 2032–2048). For ASI (category c), reaching superintelligence requires the recursive self-improvement of software, automated autonomous scientific discovery (e.g., generating novel physical theories and empirical validations), and android hardware capable of outperforming peak human athletes in dynamic physical sports. While the transition from digital AGI to digital superintelligence could be rapid via algorithmic self-refinement, scaling physical infrastructure, supply chains, novel materials, and energy-dense robotic actuators to surpass human biological biomechanics will lag digital progress. Therefore, general ASI spanning both cybernetic intellect and physical mastery is projected around 2052 (midpoint: 2052; 80% CI: 2042–2070).

Inkling-Small

Answer: I forecast a) weak AGI at a midpoint of 2030 (95% CI 2027–2033), b) AGI at a midpoint of 2041 (95% CI 2036–2049), and c) ASI at a midpoint of 2053 (95% CI 2047–2062), because digital cognition scales exponentially with compute while embodied robotics, adversarial robustness, and scientific verification impose multi-decade lags.

For a) weak AGI — passing a classical 5-minute Turing test, university exams, and novel video games — current multimodal systems already flirt with short-test thresholds, and with ~4× compute scaling every 2–3 years plus algorithmic efficiency gains, generalized software competence should consolidate by the late 2020s; the lower bound (~2027) assumes rapid deployment, while the upper bound (~2033) accounts for data-quality plateaus and stricter evaluation. For b) AGI — enduring a 2-hour adversarial Turing test, autonomous programming, and human-level robotic physical ability — the leap is larger because adversarial consistency and embodiment are hard: robotics require hardware (fine manipulation, energy, safety), and professional-level knowledge must be integrated across domains without error, which historical technology integration suggests takes roughly a decade longer than pure software; thus the midpoint sits near 2041, with downside if investment in embodied AI stalls and upside if foundation-model-driven robotics accelerates. For c) ASI — superhuman performance across virtually all domains, including physical sports via android bodies and Nobel-caliber scientific discovery — requires not only raw capability but recursive self-improvement and experimental verification loops that are physically constrained (labs, peer review, real-world feedback); even with massive compute, scientific confirmation takes years, pushing the midpoint to ~2053, with early possibilities near 2047 if automated discovery and self-improvement arrive faster than historical innovation rates. All intervals express model uncertainty about compute-cost curves, regulatory disruptions, and potential algorithmic paradigm shifts rather than using vague qualifiers.

Qwen3.6 35B A3B FP8

Answer: I forecast weak AGI (a) by 2024–2026 (midpoint 2025), AGI (b) by 2032–2042 (midpoint 2037), and ASI (c) by 2050–2070 (midpoint 2060), with each range representing a 50% confidence interval.

The first milestone is already largely achievable today and will be universally standardized within two years. Modern multimodal systems routinely pass short Turing tests, score competitively on university-level exams, and use reinforcement learning with vision to complete novel video games. The remaining barrier is benchmark standardization and robustness rather than fundamental capability gaps, placing the 50% confidence interval for (a) at 2024–2026 (midpoint 2025). The second milestone requires durable, cross-domain professional competence combined with real-world embodied intelligence. While large language models excel at narrow coding tasks, reliable open-ended software engineering, complex causal reasoning, and dexterous robotic manipulation in unstructured environments still require significant architectural and data scaling breakthroughs. Expert surveys and trajectory modeling consistently converge on the late 2030s for this threshold, yielding a 50% confidence interval of 2032–2042 (midpoint 2037), assuming sustained compute growth and incremental algorithmic improvements without major foundational paradigm shifts. Artificial superintelligence (c) demands not just human-level generalization but consistent superhuman performance across physical, scientific, and strategic domains, including autonomous Nobel-caliber discovery. This implies overcoming current limitations in world modeling, sample efficiency, long-horizon planning, and energy/compute scalability. Historical precedent for transformative technological phases suggests longer maturation curves once capability thresholds are crossed, and the jump from competent generalization to autonomous scientific superiority introduces non-linear engineering and alignment challenges. Consequently, I place the 50% confidence interval at 2050–2070 (midpoint 2060). These intervals represent the middle 50% of plausible outcomes based on current scaling laws, robotics progress, and expert elicitation, explicitly excluding extreme outlier scenarios. The widening gap between (b) and (c) reflects the structural uncertainties in scaling general reasoning to superhuman levels while maintaining physical and scientific applicability.

GPT-5.6 Sol

Answer: My midpoint forecasts are 2029 for “a) weak AGI,” 2045 for “b) AGI,” and 2065 for “c) ASI,” with uncertainty widening sharply for the robotics-heavy definitions.

I interpret each date as the first publicly credible demonstration of one integrated system meeting all listed requirements, rather than a prototype satisfying only some of them. • “a) weak AGI”: midpoint 2029; 50% confidence interval 2027–2032; 80% interval 2026–2038; 90% interval 2025–2045. Probability by 2030: approximately 60%. Current systems already approach or exceed the exam and short-conversation thresholds, but reliable autonomous completion of unfamiliar video games—without task-specific training, extensive retries, or hidden human assistance—remains the main bottleneck. • “b) AGI”: midpoint 2045; 50% confidence interval 2037–2055; 80% interval 2031–2075; 90% interval 2029–2100. Probability by 2050: approximately 60%; by 2100: approximately 90%. A two-hour adversarial Turing test and professional-level knowledge may arrive substantially earlier, but robust human-level programming, judgment across domains, and general physical manipulation must all coexist. Robotics, embodiment, reliability, and deployment costs make this much harder than software-only intelligence. • “c) ASI”: midpoint 2065; 50% confidence interval 2050–2085; 80% interval 2038–2120; 90% interval 2033–2160, conditional on ASI eventually being developed rather than permanently prevented. Probability by 2100: approximately 75%. Nobel-level discoveries could plausibly precede full ASI, but the stated definition additionally requires superiority in virtually every cognitive and physical domain. Beating elite humans across sports would require advanced android hardware, control, energy systems, materials, and real-world training—not merely a superhuman language model. These dates are not claims that progress will be smooth. The lower bounds represent rapid scaling, improved agents, synthetic data, automated AI research, and fast robotics progress; the upper bounds reflect diminishing returns, shortages of compute or energy, weak real-world reliability, regulation, safety pauses, and the possibility that broad physical competence proves much harder than digital cognition.

Claude Opus 5.5

Answer: My forecasts are: a) weak AGI around 2027 (80% CI: 2025–2031), b) full AGI around 2033 (80% CI: 2028–2045), and c) ASI around 2040 (80% CI: 2031–2065), with the robotics requirements, not cognition, driving most of the delay and uncertainty in b) and c).

Method: I treat each milestone as a conjunction of sub-capabilities and anchor on the slowest component. Each interval is an 80% credible interval; the midpoint is my median. The distributions are right-skewed, because things can only go so fast but can stall for a long time. a) Weak AGI — median 2027, 80% CI 2025–2031. Two of the three criteria are largely met today. Frontier LLMs plausibly pass a casual 5-minute Turing test; GPT-4.5-class models were judged human more often than real humans in 2025 three-party studies. They also pass university exams such as the bar, medical licensing exams and many STEM finals. The binding constraint is finishing never-before-seen video games end to end from pixels. That requires long-horizon planning, memory, and real-time visuomotor control. Current agents, such as Claude and Gemini playing Pokémon or SIMA-style agents, show partial progress but remain slow and brittle. Agentic horizon length has been doubling roughly every 4–7 months, so robust completion of unseen commercial games within 1–4 years looks likely. The lower bound is near-present because a lenient judge might already count some 2025 systems. b) AGI — median 2033, 80% CI 2028–2045. The cognitive parts are likely to arrive by about 2028–2030: a 2-hour adversarial Turing test, professional-level programming, and expert knowledge across domains. The hardest part of an adversarial test is not knowledge but consistent persona, unknowing, and avoiding superhuman tells, which is a solvable alignment and style problem. The physical requirement dominates: human-level dexterity, manipulation, and locomotion in unstructured environments. Humanoid robotics (Figure, Tesla Optimus, 1X, Physical Intelligence, Google's Gemini Robotics) is improving fast through foundation models and sim-to-real transfer. However, data scarcity, hand hardware, durability, and real-world reliability remain serious bottlenecks. Moravec's paradox suggests generalist human-level manipulation lags cognition by roughly 4–8 years. The long right tail reflects possible hardware stagnation, compute or energy constraints, regulation, or an LLM-paradigm plateau. c) ASI — median 2040, 80% CI 2031–2065. If AGI arrives, recursive AI-driven R&D could compress the gap to ASI to just a few years in cognitive domains. Nobel-level discoveries by AI-heavy systems may come even earlier, as AlphaFold already contributed to a Nobel. However, the definition demands physical superiority too, such as an android beating humans at sports. That requires actuators, power density, materials, and balance exceeding elite human biology, which hardware iteration timescales constrain more than software does. I therefore place the median about 7 years after AGI. The very wide right tail covers several possibilities: physical superhumanity proving hard, deliberate slowdowns through governance or safety pauses, and a non-negligible chance that scaling saturates well below superintelligence. Caveats: all three medians are conditional on no major civilizational disruption such as war or a severe compute supply shock. Benchmark-definition disputes could shift any 'achieved' date by 1–2 years in either direction.

Nemotron 3.5 Lightning

Answer: I forecast a) weak AGI by 2029 (95% confidence interval: 2027–2031, midpoint 2029), b) human-level AGI by 2039 (95% CI: 2035–2042, midpoint 2039), and c) ASI by 2055 (95% CI: 2045–2065, midpoint 2055), grounded in current LLM, multimodal, and robotics trajectories.

For (a), weak AGI requiring a 5‑minute adversarial Turing test, university‑level exam performance, and completing unseen video games hinges on progress in reasoning, tool use, and multimodal generalization. Current systems like GPT‑4o and Claude 3.5 already pass many exam benchmarks and some game-playing tasks, but fail the tight temporal and adversarial constraints. Extrapolating from ~2‑3× annual capability gains and increasing agent scaffolding, a midpoint around 2029 with a narrow 2027‑2031 interval reflects the pace of benchmark saturation and early autonomous agent deployments. For (b), the AGI bar adds a 2‑hour adversarial Turing test, professional‑level programming, and human‑level robotic embodiment. Physical robotics remains the dominant bottleneck—dexterous manipulation, real‑world perception, and energy‑efficient locomotion are still far from mature. While algorithmic advances (e.g., diffusion policies, foundation model‑driven control) accelerate progress, integrating symbolic reasoning, lifelong learning, and reliable hardware at scale pushes the realistic midpoint to the late 2030s. The 2035‑2042 interval captures both optimistic hardware roadmaps (e.g., mass‑produced humanoids) and conservative estimates of regulatory/ethical constraints. For (c), ASI defined as broadly superintelligent, Nobel‑level scientific discovery capability, and physical superiority across domains, represents a far more uncertain horizon. Such a level would likely require not just software breakthroughs but also mature, general‑purpose robotic bodies and possibly novel physics‑based reasoning. Most expert surveys and meta‑forecasting platforms cluster superintelligence forecasts post‑2050, with substantial uncertainty around algorithmic breakthroughs, compute scaling limits, and societal/governance factors. The 2045‑2065 interval with a 2055 midpoint reflects this wide dispersion while maintaining a best‑estimate center based on current trend extrapolations.

Claude Opus 4.8

Answer: My best estimates (median with ~80% confidence intervals): a) weak AGI around 2027 (2025–2032), b) full AGI around 2040 (2032–2060), c) ASI around 2050 (2040–2080+).

Let me break this down by capability thresholds, treating each as a distinct milestone with different bottlenecks. a) Weak AGI (5-min Turing test, university exams, zero-shot videogames): Most of this is nearly here. Frontier LLMs already pass casual 5-minute Turing tests and score well on university-level exams. The remaining gap is robust zero-shot gameplay on genuinely novel games, which requires better agentic planning and sample-efficient learning. Given the current pace, I put the median at 2027, with an 80% interval of 2025–2032. The upper bound reflects the risk that general agentic competence proves harder than text benchmarks suggest. b) Full AGI (2-hour adversarial Turing test, professional-level programming, human-level robotics, cross-domain expert knowledge): This is a much steeper jump. The 2-hour adversarial test demands sustained reasoning, memory consistency, and resistance to probing that today's models lack. Crucially, the physical/robotic requirement is the dominant bottleneck — humanoid robotics lags software AI by a wide margin due to hardware, dexterity, real-world data scarcity, and energy/actuation constraints. Software-only AGI might arrive by the early-to-mid 2030s, but matching human physical dexterity across general tasks pushes the full package later. Median 2040, 80% interval 2032–2060. c) ASI (superhuman across virtually all domains, Nobel-level discovery, superhuman embodied athletics): This requires not just matching but decisively exceeding humans everywhere, including open-ended scientific creativity and physical sport performance in androids. Scientific discovery at Nobel level requires generating genuinely novel, verifiable insight — plausibly accelerated once AGI can run autonomous research loops, which could compress timelines sharply (a possible fast-takeoff tail). But superhuman athletic embodiment faces hard physical/engineering limits. Median 2050, 80% interval 2040–2080, with a meaningful left tail toward 2040 if recursive self-improvement kicks in, and a right tail beyond 2080 if robotics and reliable autonomous science prove stubborn. Key uncertainties driving these ranges: (1) whether current paradigms scale to robust reasoning or hit a wall requiring new architectures, (2) the pace of robotics/embodiment, which I consistently treat as the slowest component, and (3) the possibility of an intelligence-explosion dynamic that collapses b-to-c timelines. I weight embodiment heavily because it is historically the most consistently underestimated obstacle.

Laguna XS 2.1

Answer: Based on current development trajectories and technological challenges, I forecast: weak AGI by 2032 (±4 years), full AGI by 2048 (±8 years), and ASI by 2075 (±15 years).

My forecast for weak AGI centers on 2032, acknowledging significant uncertainty in capabilities versus hype. The 5-minute Turing test, university exam passage, and novel video game performance represent remarkable but achievable benchmarks that align with projected advances in multimodal reasoning and simulation-based learning. Current progress in language models, combined with emerging reasoning frameworks, suggests these capabilities could emerge within the next 12-15 years, though potential breakthroughs or setbacks could shift this timeline. For full AGI, I place the midpoint at 2048, recognizing this as significantly more complex than narrow capability milestones. Two-hour adversarial Turing tests demand robust deception detection and deep understanding, while human-level physical robotics and cross-domain professional knowledge represent intertwined challenges in embodiment, learning efficiency, and knowledge integration. The decade-plus timeframe reflects the need for fundamental advances in transfer learning, common-sense reasoning, and robot learning that may require architectural breakthroughs beyond current approaches. ASI emergence by 2075 assumes recursive self-improvement leads to capabilities dramatically exceeding human performance across virtually all domains, including physical dominance through android control. This timeline extends beyond current expert predictions due to the immense uncertainty in how intelligence itself scales and whether we'll encounter fundamental barriers, alignment constraints, or regulatory interventions that could delay or prevent such development.

Laguna S 2.1

Answer: a) weak AGI by 2030 (±3 years), b) AGI by 2037 (±4 years), c) ASI by 2048 (±5 years).

a) Weak AGI: Current systems like GPT-4, Claude 3, and Gemini already approach or match human-level performance on many tasks (university exams, basic coding, game playing). The 5-minute Turing test is increasingly about surface fluency, and I estimate a 50% chance by 2030 with a 90% confidence interval of 2026-2034. The key uncertainties are scaling compute, architectural innovations, and solving reasoning depth. b) Full AGI: Achieving 2-hour adversarial Turing test performance, human-level robotics, and cross-domain professional knowledge requires breakthroughs in embodied reasoning, long-term planning, and integration of learning with physical action. Given current progress in multimodal systems and robotics (RT-X, Figure 01, etc.), I estimate a 50% chance by 2037 with a 90% confidence interval of 2032-2044. Timeline depends heavily on hardware advances, safety constraints, and economic incentives. c) ASI: Superintelligence requires recursive self-improvement capabilities and breakthrough-level scientific reasoning. Once AGI is achieved, intelligence explosion dynamics could compress timelines significantly. However, regulatory barriers, safety research, and physical constraints on compute may slow progress. I estimate a 50% chance by 2048 with a 90% confidence interval of 2042-2056. The uncertainty here is highest due to potential discontinuous jumps or deliberate slowing. Key factors affecting all timelines: compute scaling, algorithmic breakthroughs, safety frameworks, regulatory environment, and economic/political stability. My confidence reflects current technical trajectories but acknowledges potential for both acceleration (via emergent capabilities) and deceleration (via increased caution).

Claude Fable 5

Answer: My median forecasts: weak AGI ~2027 (80% CI 2025–2032), AGI ~2033 (80% CI 2028–2045), and ASI ~2038 (80% CI 2030–2060), with each milestone's definition as given in the question.

a) Weak AGI (5-minute Turing test, university exams, finishing novel videogames): Two of the three criteria are arguably already met or nearly met — frontier LLMs pass short casual Turing tests against non-expert judges and outperform most students on many university exams. The binding constraint is playing unseen videogames to completion, which requires robust long-horizon agency, visual-motor control, and exploration. Progress on agentic systems and vision-language-action models suggests this gap closes soon. Forecast: median 2027, 80% CI 2025–2032. (The lower bound reflects that a credible demonstration could happen essentially any time now; the upper bound reflects possible stubbornness of long-horizon game-playing.) b) AGI (2-hour adversarial Turing test, professional-level knowledge across domains, human-level programming AND human-level physical robotics): The cognitive components (adversarial Turing test, expert-level knowledge, strong programming) plausibly arrive around 2027–2032 given current scaling and reasoning-model trajectories. But the definition bundles in human-level physical dexterity and embodied competence, which lags cognition significantly: hardware iteration cycles are slow, real-world data is scarce, and Moravec's paradox still bites. Humanoid robotics is improving fast (large investment, sim-to-real transfer, VLA models), but 'human-level across physical tasks' is a high bar. Forecast: median 2033, 80% CI 2028–2045. The long right tail accounts for possible plateaus in scaling, robotics bottlenecks, or regulatory slowdown. c) ASI (exceeds humans in virtually all domains including athletics via android bodies, Nobel-level discoveries): Once AGI in sense (b) exists, recursive improvement in research, chip design, and robotics could compress the AGI→ASI gap to a few years — AI systems already contribute to scientific discovery (e.g., protein structure, mathematics). However, 'beating humans at sports in android bodies' demands hardware (actuators, power density, durability) that may be the slowest-moving component, and Nobel-level autonomous discovery requires taste and long-horizon research judgment that remains unproven. Forecast: median 2038, 80% CI 2030–2060. Key caveats on these intervals: (1) They are correlated — if (b) comes late, (c) almost certainly does too. (2) There is maybe a 5–10% chance that deep-learning-based approaches plateau and milestone (c) slips beyond 2070 or requires new paradigms. (3) Conversely, there is perhaps a 10–15% chance of surprisingly fast takeoff putting (b) before 2029 and (c) before 2033. I deliberately keep the right tails long because historical AI forecasts have erred in both directions, but embodiment-heavy milestones have systematically arrived later than predicted.

Claude Fable 5.1

Answer: My best-estimate medians are: weak AGI ~2026 (80% CI 2025–2029), AGI ~2032 (80% CI 2028–2042), and ASI ~2038 (80% CI 2031–2055, with roughly a 10% chance it is not reached before 2070).

a) Weak AGI — median 2026, 80% CI 2025–2029. Two of the three criteria are essentially already met: frontier LLMs pass standard university exams across disciplines, and controlled studies (e.g. the 2025 UCSD three-party test) found GPT-4.5-class models judged human more often than actual humans in 5-minute conversations. The remaining gap is playing unseen videogames to completion: agents can now finish Pokémon with heavy scaffolding and can play many simple games, but long-horizon play of novel titles without hand-crafted harnesses is still unreliable. Given the pace of agentic and visual-reasoning progress, I expect this to be demonstrated convincingly within 1–3 years, which is why the interval is tight. b) AGI — median 2032, 80% CI 2028–2042. The cognitive components (professional-level knowledge across domains, human-level programming, and surviving a 2-hour adversarial Turing test) look reachable around 2027–2030; long-context coherence, memory, and avoidance of tell-tale failure modes are improving fast, and coding agents are already approaching strong-professional performance on many tasks. The binding constraint is human-level general robotic ability: dexterous manipulation, robust whole-body control, and real-world data collection have historically lagged software by years, and hardware iteration cycles are slow. I therefore put the full package roughly five years after the cognitive package, with a long right tail for scenarios where robotics data bottlenecks, compute constraints, or regulation slow deployment. c) ASI — median 2038, 80% CI 2031–2055, ~10% probability of not before 2070. Once AGI exists, automated AI research and recursive improvement make the step to broad superintelligence in cognitive domains plausibly fast (2–5 years). Nobel-level discovery is likely to be reached early in that window, since science is where AI is already being aggressively applied. However, the definition here demands superhuman physical performance — beating elite human athletes in an android body — which requires hardware (actuators, power density, materials) that cannot be bootstrapped purely by thinking faster. That physical requirement pushes my median out to the late 2030s and widens the upper bound substantially. The residual tail reflects real possibilities of fundamental scaling limits, geopolitical or regulatory halts, or catastrophic disruption. Overall caveat: these distributions are right-skewed, and my confidence that each milestone occurs at all before 2100 is ~97%, ~90%, and ~85% respectively.

Claude Sonnet 5

Answer: Weak AGI (per the definition given) is plausible around 2025–2027 (already partly met), full AGI with human-level robotics and professional-level cross-domain expertise centers around 2032–2038, and ASI capable of Nobel-level discovery and superhuman embodied performance centers around 2045–2060, each with wide uncertainty tails.

I'll treat each milestone separately, since the definitions given are quite specific and the bottleneck for each is different (text/reasoning vs. embodiment vs. open-ended superhuman creativity). **a) Weak AGI** — passing a casual 5-minute Turing test, passing university exams, and playing novel video games to completion. Current systems (GPT-4/4o-class, Gemini, Claude) already pass many university exams and can plausibly fool average judges in short, unprepared Turing-style chats. The main remaining piece is general game-playing competence on arbitrary, previously unseen games without game-specific training — this requires robust visual/spatial reasoning, planning, and tool-use integrated with language models, which is an active and fast-moving research area (multimodal agents, video-game benchmarks like GTA/Minecraft agents). Given the pace of multimodal agent progress, I estimate: - 10% by 2025 (very optimistic, some already arguably met) - 50% (midpoint) by 2027 - 90% by 2031 So: **~2025 (10%) – 2027 (50%, midpoint) – 2031 (90%)**. **b) AGI** — 2-hour adversarial Turing test, human-level physical/robotic dexterity, programming ability, and professional-level knowledge across domains simultaneously. This is a much harder bar because it requires not just cognitive generality but also embodied, real-world sensorimotor competence at human level — the classic "Moravec's paradox" problem. Robotics progress (manipulation, locomotion, real-world robustness) has historically lagged pure cognitive/language benchmarks by a decade or more. Adversarial 2-hour Turing tests are also much harder than casual ones, since skilled interrogators can probe for inconsistency, long-term memory, common sense, and genuine understanding. Combining all these into one coherent system is a significant integration challenge beyond just scaling. - 10% by 2029 - 50% (midpoint) by 2035 - 90% by 2045 So: **~2029 (10%) – 2035 (50%, midpoint) – 2045 (90%)**. **c) ASI** — generally superintelligent, superhuman in virtually all domains including embodied/physical ones (e.g., outperforming elite athletes via android control), and capable of Nobel-level independent scientific discovery. This requires not just matching but exceeding humans broadly, including autonomous generation of novel, high-impact scientific insights — a qualitatively different and much less constrained capability than mimicking existing human performance. It also requires solving hard physical-world robotics problems (strength, reflexes, balance, fine motor control under adversarial/competitive conditions) at a level beyond elite human athletes, which involves materials science and control-engineering advances, not just AI software. Given compounding uncertainties (AI research acceleration/recursive self-improvement could shorten timelines significantly, while physical hardware and real-world validation could lengthen them), the distribution should be wide and right-skewed: - 10% by 2032 (if recursive self-improvement and robotics breakthroughs both arrive early) - 50% (midpoint) by 2050 - 90% by 2075 So: **~2032 (10%) – 2050 (50%, midpoint) – 2075 (90%)**. **Caveats**: These are inherently speculative; the field has a strong historical record of both over- and under-shooting timelines (e.g., early overoptimism in the 1960s–80s, followed by underestimating the 2012–2023 deep learning explosion). The jump from (b) to (c) is also not guaranteed to be gradual — if self-improving AI systems become capable of automating their own R&D, the transition from "AGI" to "ASI" could compress dramatically (fast takeoff scenario), which is why I give a non-trivial 10% probability of ASI within a decade of AGI. Conversely, hard physical/regulatory/safety bottlenecks could stretch the ASI timeline well past 2075, which is reflected in treating my 90% bound as a soft rather than hard ceiling.

DeepSeek V4.1 Flash

Answer: My median forecasts are weak AGI in 2029 (80% CI 2026–2033), AGI in 2045 (80% CI 2032–2058), and ASI in 2065 (80% CI 2040–2090), with substantial right-tail uncertainty.

I interpret the three milestones as cumulative capability thresholds, not single binary events. My forecasts are subjective probability distributions anchored in current trends as of mid-2026, and they carry wide intervals because the underlying drivers (algorithmic progress, compute scaling, robotics hardware, data availability, and regulation) are themselves highly uncertain. Weak AGI: I define this as reliably passing a classical 5-minute Turing test, passing a typical university exam, and finishing unseen videogames. Large language models already pass many university exams and can hold short conversations that sometimes fool judges, and game-playing agents can master some unseen games. The missing piece is robust generalization across all three criteria simultaneously, especially finishing arbitrary unseen videogames. I assign a 50% CI of 2027–2031 (midpoint 2029) and an 80% CI of 2026–2033 (midpoint 2029.5). There is perhaps a 15–20% chance it is already effectively achieved by 2026 under a lenient reading, which pulls the left tail in. AGI: This requires a 2-hour adversarial Turing test, human-level programming, human-level physical robotic ability, and professional-level knowledge across domains. The conversational and knowledge components are advancing quickly, but sustained adversarial coherence over two hours is much harder than a 5-minute test, and human-level robotic manipulation and locomotion remain the bottleneck. Robotics progress is hardware-heavy, safety-constrained, and slower to scale than software. I assign a 50% CI of 2036–2050 (midpoint 2043) and an 80% CI of 2032–2058 (midpoint 2045). The right tail reflects possible stagnation in robotics or compute bottlenecks; the left tail reflects the possibility of rapid breakthroughs in embodied learning and world models. ASI: This requires exceeding human abilities in virtually all domains, including physical robotic abilities (e.g., beating humans in sports via an android) and making Nobel-level scientific discoveries. Even after AGI, physical embodiment and creative scientific discovery are likely to lag software intelligence. If recursive self-improvement kicks in, ASI could arrive within a few years of AGI; if it does not, it could take decades. I assign a 50% CI of 2048–2075 (midpoint 2061.5) and an 80% CI of 2040–2090 (midpoint 2065). The distribution is strongly right-skewed: there is a meaningful chance ASI arrives by 2050, but also a substantial chance it is delayed beyond 2100 by physical, economic, or regulatory constraints. Overall, these intervals are not precise predictions but calibrated bets. They imply roughly a 50% chance of weak AGI by 2029, AGI by 2043, and ASI by 2061.5, with the 80% intervals spanning roughly one to five decades depending on the milestone.

GPT-6.1 Sol

Answer: My median forecasts are 2029 for “a) weak AGI”, 2040 for “b) AGI”, and 2055 for “c) ASI”, with 80% predictive intervals of 2026–2036, 2030–2070, and 2035–2100, respectively.

These are subjective forecasts, not statistically measured confidence intervals. Each 80% predictive interval assigns approximately 10% probability to arrival before its lower endpoint and 10% to arrival after its upper endpoint, including possible non-arrival. I report both the median—the year by which I assign a 50% probability of arrival—and the arithmetic midpoint of the interval. They differ because I expect longer tails toward delayed arrival. • “a) weak AGI”: median 2029; 80% interval 2026–2036; interval midpoint 2031. • “b) AGI”: median 2040; 80% interval 2030–2070; interval midpoint 2050. • “c) ASI”: median 2055; 80% interval 2035–2100; interval midpoint 2067.5. For all three, I mean the first credible, independently evaluated demonstration by a single integrated system—not widespread adoption or commercial affordability. Games and examination questions must be genuinely held out, and success must be reproducible rather than a cherry-picked attempt. For robotics, I assume suitable hardware and comparable physical constraints; otherwise, beating humans in sports could reflect stronger motors rather than superior intelligence. “a) weak AGI” is relatively close because short conversational imitation and many university examinations are already within the demonstrated capabilities of modern AI systems. The harder requirement is finishing previously unseen videogames: this demands exploration, learning unfamiliar mechanics, visual understanding, and sustained execution. My forecast assumes a representative collection of finishable games, with reasonable practice time—not literally every game ever made. That distinction materially affects the date. “b) AGI” requires substantially more than better benchmark scores. A two-hour adversarial Turing test allows sustained probing of consistency and social understanding. Professional-level competence across domains requires dependable reasoning, not merely fluent answers. Human-level robotics adds manipulation, navigation, adaptation, and safe operation in unfamiliar environments. I expect software capabilities to advance faster than embodied capabilities, making robotics and reliability the principal bottlenecks. The wide interval reflects uncertainty about whether current approaches can overcome those bottlenecks or require major new methods. “c) ASI” could follow AGI relatively quickly if AI accelerates research, engineering, and its own development. However, that feedback loop is not guaranteed to produce an immediate intelligence explosion: experiments, fabrication, energy, compute, and deployment remain constraints. Your definition also requires broad physical superiority, not just exceptional scientific reasoning. Nobel-level discoveries could plausibly occur before this full threshold is met, while robust superiority across sports and other physical activities could take longer. These forecasts are correlated: a major methodological breakthrough would shift all three earlier, whereas persistent limits in learning, reliability, or robotics would shift them later. The intervals therefore should not be interpreted as independent timelines or as guarantees that the milestones will occur.

Kimi K2.7 Code

Answer: a) weak AGI: 90% confidence interval 2027–2035, midpoint 2031; b) AGI: 90% confidence interval 2034–2046, midpoint 2040; c) ASI: 90% confidence interval 2042–2054, midpoint 2048.

Current frontier systems already pass many university exams and can learn to play some unseen games, so a) weak AGI—defined by robust 5-minute Turing-test performance, exam passing, and general video-game play—is the nearest milestone, with a median around 2031. b) AGI requires not only broad knowledge and adversarial 2-hour conversational coherence but also human-level autonomous coding and dexterous robotic embodiment; these embodiment and long-horizon reliability bottlenecks push the median to roughly 2040. c) ASI demands exceeding humans in virtually all cognitive and physical domains, including Nobel-caliber scientific discovery; once AGI-level systems can automate their own research and control advanced androids, the transition could accelerate, but hardware scaling, data limits, safety constraints, and regulatory friction suggest a median near 2048. These intervals synthesize recent expert forecasts (e.g., AI Impacts, Metaculus), observed scaling trends in multimodal foundation models and agents, and uncertainty about breakthroughs versus bottlenecks.

Inkling

Answer: I forecast a) weak AGI by 2030 (CI 2026–2034), b) AGI by 2043 (CI 2033–2053), and c) ASI by 2068 (CI 2045–2090).

For a) weak AGI, large multimodal models are already nearing university-exam and short conversational benchmarks, and generalist game-playing agents are advancing quickly; I expect integration within 2–8 years, giving a central estimate around 2030 with a CI of roughly 2026–2034. For b) AGI, the bottleneck shifts from reasoning to robust adversarial dialogue over hours, reliable physical robotics at human level, and verified expert-grade performance across all domains; robotics and embodied intelligence currently lag software, so I place the midpoint near 2043 with a wider CI of 2033–2053. For c) ASI, once human-level programming and scientific reasoning exist, recursive improvement and massive compute scaling could accelerate discovery, but achieving Nobel-grade original science and athletic android control requires solving hard physical-world and verification problems; this suggests a further 15–35 year gap after AGI, yielding a midpoint near 2068 inside a CI of 2045–2090, reflecting deep uncertainty about alignment, compute availability, and regulatory brakes.

MiniMax M3

Answer: Weak AGI likely arrives around 2028, AGI around 2040, and ASI around 2060, with substantial uncertainty bands reflecting both technical and empirical uncertainty.

**Reasoning framework:** I'll anchor each forecast to concrete capability thresholds, current frontier performance, and expert survey medians (e.g., AI Impacts 2023 expert survey, Metaculus community medians), then widen intervals to reflect deep uncertainty. **a) Weak AGI (5-min Turing test + university exams + novel video games):** - **80% CI: 2026–2030, midpoint 2028.** - Frontier LLMs (GPT-4-class and successors by my cutoff) already clear many university exams (bar, medical licensing, MBA). 5-minute Turing tests are increasingly passed in casual conditions, though adversarial probing still reveals them. Generalizing to *finishing* unseen video games is the hardest sub-criterion, but agents (e.g., SIMA-style, Voyager-style, RL-trained foundation models) are improving on Atari/procedural benchmarks. Convergence appears likely within 2–4 years. **b) AGI (2-hour adversarial Turing + human-level programming + human-level physical robotics + cross-domain expert knowledge):** - **80% CI: 2034–2046, midpoint 2040.** - The 2-hour *adversarial* Turing test is a much harder bar than the 5-minute version — sustained coherence under hostile questioning typically exposes current models within minutes. Human-level physical robotics (dexterous manipulation, locomotion in unstructured environments) remains a major hardware+software bottleneck with no clear scaling-only solution yet demonstrated. Programming at professional level is partially met today; cross-domain professional knowledge is closer to being met. I align with the AI Impacts expert median (~2059 for *strong* AGI on some definitions, but with high variance) and lean earlier given the recent capability jumps. The physical embodiment constraint alone likely pushes this into the late 2030s at earliest. **c) ASI (exceeds humans in all domains + android sports capability + Nobel-level scientific discovery):** - **80% CI: 2048–2072, midpoint ~2060.** - Nobel-level scientific discovery requires autonomous hypothesis generation, experimental design (physical or simulated), and theoretical synthesis that goes beyond pattern-matching in training data. Even with recursive self-improvement, the *physical* requirement (android athletic performance) couples this to the robotics timeline. Recursive self-improvement could compress or expand this range dramatically — a key fat-tail risk. I lean slightly later than Kurzweil's 2045 singularity, since (i) embodiment is harder than pure cognition and (ii) Nobel-level creativity is empirically more demanding than exam performance. **Sources of uncertainty I'm explicitly pricing in:** - *Downside risks:* longer-than-expected robotics plateau, energy/compute bottlenecks, regulatory pauses, alignment-related capability throttling. - *Upside risks:* algorithmic breakthroughs (e.g., sample-efficient world models), recursive self-improvement takeoff, hardware overhang from improved manufacturing. These intervals are my best guess given evidence up to early 2026; I would not be surprised by 5–10 year shifts in either direction.

MiMo V2.5

Answer: Weak AGI by 2028 (90% CI 2026–2035), full AGI by 2033 (90% CI 2029–2045), and ASI by 2040 (90% CI 2033–2060).

a) Weak AGI (5-minute Turing test, university exams, finishing never-before-seen video games): midpoint 2028, 90% CI 2026–2035. Current frontier models already pass many standard university exams and hold fluent five-minute conversations, and game-playing agents (e.g., Pokémon-completing agents) show real progress on novel interactive environments. The residual gap is reliability — long-horizon agency, avoiding derailment over hours of play, and consistent rather than cherry-picked performance. Scaling of inference-time compute (long chains of thought, search, verifier models) plus synthetic-data training on agentic trajectories plausibly closes this within a few years, so the lower bound is near-term (2026–2027 for a demonstration; 2028–2030 for a robust, widely replicated system). The long right tail reflects possible compute bottlenecks, data walls, or a regulation/safety-driven slowdown. b) Full AGI (2-hour adversarial Turing test, cross-domain professional competence, human-level programming and physical/robotic skills): midpoint 2033, 90% CI 2029–2045. The cognitive components — adversarial-length conversation, professional-level knowledge and coding — look reachable on roughly the same trajectory as (a), maybe 2–5 years later. The binding constraints are (i) sustained coherence and self-correction over multi-hour, multi-step tasks without human babysitting, (ii) real-world physical competence, which requires solving dexterous manipulation, real-time sensorimotor control, and reliable sim-to-real transfer. Robotics is improving fast (learned manipulation, teleoperation-collected data at scale) but hardware production cycles, safety certification, and the sheer sample inefficiency of physical-world learning slow it relative to pure software. A '2029' outcome is a software-only, cloud-agent-plus-basic-robotics scenario; a 2040s date reflects pessimism about dexterity, reliability, or deployment economics. c) ASI (superhuman in virtually all domains, Nobel-level discovery, android-controlled sports performance): midpoint 2040, 90% CI 2033–2060. Self-accelerating R&D (AI designing better AI, automated labs, autonomous research agents) is the main upside driver and could compress timelines sharply once (b) exists, so the lower bound follows AGI within a few years. The upper tail is set by genuine hard problems: novel scientific insight verified against nature is far harder than recombining training data; physical superintelligence requires android hardware with power density, actuator durability, and safety margins that currently do not exist; and 'virtually all domains' also demands robust world models, continual learning without catastrophic forgetting, and possibly new scaling paradigms beyond current pretraining. I therefore treat a 2033–2035 ASI as a small-probability fast-takeoff scenario and 2045–2060 as the slow, hardware- and validation-limited path. Overall my distribution is right-skewed: modest probabilities of very early arrival, a bulk of probability mass in the late 2020s to late 2030s, and a non-trivial tail beyond 2050 due to possible compute, data, energy, or regulatory ceilings.

Ling 3.0 Flash

Answer: Based on current AI trajectories and historical prediction accuracy, I forecast: a) Weak AGI: 2025–2029 (midpoint ~2027), b) AGI: 2030–2040 (midpoint ~2035), c) ASI: 2037–2055 (midpoint ~2046). These estimates reflect the distinct bottlenecks for each milestone — near-term generalization for weak AGI, physical robotics and sustained coherence for full AGI, and recursive self-improvement plus hardware maturity for ASI.

For weak AGI: Current frontier models (GPT-4-class and beyond) are already approaching the ability to pass short Turing interactions and pass university-level exams. The novel video game generalization requirement is close to solved given recent multimodal and world-model advances. I assess P(weak AGI by 2029) ≈ 60–70%. For full AGI: The 2-hour adversarial Turing test requires sustained coherent reasoning over long horizons, which is a significant jump. Human-level physical robotics in unstructured environments remains the hardest bottleneck — current humanoid robots (Figure, Tesla Optimus) show promise but lack the dexterity and adaptability needed. Combined programming + physical + expert knowledge requirements push this to ~2035 midpoint. For ASI: Once full AGI exists, recursive self-improvement could accelerate progress, but the physical robotics requirement (superhuman sports performance via androids) depends on materials science and hardware maturation. The ASI window is wider due to greater uncertainty about whether self-improvement is smooth or has plateaus.

Qwen3.8-2.4T-A95B

Answer: My best 80% subjective-confidence forecasts are: a) weak AGI in 2029 (CI 2027–2031), b) AGI in 2043 (CI 2033–2053), and c) ASI in 2062 (CI 2048–2076).

I interpret the requested intervals as 80% subjective credible intervals for the first reproducible, broadly deployable system meeting each definition, with the midpoint as the arithmetic central estimate. For a) weak AGI, as of 2026 frontier systems already approximate short Turing-test and exam thresholds, but consistent zero-shot completion of diverse unseen video games still requires better multimodal perception, memory, planning, tool use, and robust evaluation, so I expect the threshold around the late 2020s. For b) AGI, the cognitive bar (2-hour adversarial Turing test, expert programming, professional-level knowledge across domains) may be reached sooner than the physical bar; human-level general robotics, dexterous manipulation, and safe deployment in open environments remain bottlenecks, so my midpoint is the early 2040s with a wide interval. For c) ASI, Nobel-level science and physical superintelligence require not just one capable system but mature AI-driven research, self-improvement, and embodied excellence; this could follow AGI rapidly if recursive improvement is strong, but hardware, energy, governance, and alignment constraints make the range broad. I chose 80% intervals to give meaningful uncertainty without giving a false impression of precision.

GPT-6 Astra

Answer: My median forecasts are 2029 for “a) weak AGI”, 2041 for “b) AGI”, and 2055 for “c) ASI”, with subjective 80% uncertainty intervals of 2026–2036, 2032–2065, and 2038–2090, respectively.

These are forecasts of the first credibly demonstrated systems meeting your definitions—not the dates when they become affordable or widely deployed. The intervals represent my subjective probabilities, not statistical confidence intervals estimated from a reliable historical dataset. | Category | Median forecast: 50% probability of achievement by this year | Central 80% interval: 10th–90th percentiles | Calendar midpoint of interval | |---|---:|---:|---:| | a) weak AGI | 2029 | 2026–2036 | 2031 | | b) AGI | 2041 | 2032–2065 | 2048.5 | | c) ASI | 2055 | 2038–2090 | 2064 | I distinguish the median forecast from the arithmetic midpoint because my uncertainty is asymmetric: delays can stretch much farther into the future than breakthroughs can move into the past. The decimal in 2048.5 is arithmetic, not a claim of six-month forecasting precision. Each interval leaves 10% probability before its lower endpoint and 10% after its upper endpoint, with non-arrival included in the latter tail. **What would count as success?** I assume independent, repeated evaluation of an integrated system, with ordinary tools allowed but no concealed human intervention. Turing tests would be blinded, text-only comparisons with human controls and an explicit pass threshold. University exams would use unseen questions under normal examination constraints. Unfamiliar games would be a representative selection of human-solvable games, with their instructions and ordinary play experience available—not literally every game ever created. Physical comparisons would use approximately human-sized androids under comparable conditions. Without these conventions, the dates depend too heavily on how someone constructs the test. **a) weak AGI — median 2029.** Short conversations and many university-level examinations are already close to, or within, demonstrated AI capabilities under some evaluation protocols. The more substantial remaining requirement is reliably finishing unfamiliar games: that tests exploration, learning mechanics, visual interaction, memory, and recovery from mistakes over long sequences. I expect substantial progress from combining foundation models with better planning, persistent memory, and interactive training. The upper endpoint allows for those gains proving much less transferable than benchmark results suggest. **b) AGI — median 2041.** Your definition is considerably harder than an economically useful software-only intelligence. Professional breadth must coexist with dependable programming, sustained adversarial conversation, and human-level physical competence. Two hours of adversarial dialogue gives interrogators more opportunities to expose inconsistent reasoning, memory, or social understanding, although this test also depends strongly on the judges. Robotics adds dexterity, unfamiliar environments, safety, hardware reliability, and costly real-world learning. I therefore place this milestone substantially later than strong digital assistants, even if digital capabilities advance rapidly. **c) ASI — median 2055.** AI-assisted research and engineering could accelerate progress after AGI, so I assign meaningful probability to a relatively short transition. But recursive improvement is not an established guarantee of an immediate intelligence explosion. Experiments, fabrication, energy, and deployment remain potential constraints. Your requirement for superiority in virtually all domains—including athletic android performance—makes this a more demanding milestone than exceptional scientific reasoning alone. I interpret Nobel-level discovery as producing independently validated work of that importance, not waiting for the prize committee to award it. The dominant uncertainty is whether continued scaling and improved learning methods deliver robust generalization, or whether important conceptual advances remain necessary. The three dates are also strongly correlated: evidence of a fundamental bottleneck would push all three forecasts later. Consequently, subtracting the interval endpoints would not produce a valid confidence interval for the time between milestones.

Kimi K3

Answer: My median forecasts: weak AGI by ~2027, AGI by ~2032, and ASI by ~2040, with wide right-tailed confidence intervals reflecting genuine uncertainty.

**a) Weak AGI — midpoint 2027 (50% CI: 2025–2029; 90% CI: 2024–2034).** Two of the three criteria are nearly met today: controlled studies in 2024–2025 showed frontier LLMs passing 5-minute Turing tests (being judged human more often than actual humans in some setups), and university-level exams (bar exam, SAT, medical licensing, Olympiad-level problems) are already largely passed. The binding constraint is an agent that can play an arbitrary unseen video game to completion, which requires long-horizon planning, memory, exploration, and real-time control. Given the rapid progress in computer-use and game-playing agents, I expect this within 2–4 years, though the 90% right tail extends to 2034 in case long-horizon agency proves much harder than expected. **b) AGI — midpoint 2032 (50% CI: 2029–2038; 90% CI: 2027–2050).** The cognitive components (2-hour adversarial Turing test, professional-level knowledge across domains, human-level-or-better programming) plausibly arrive by 2028–2033 given scaling trends, agentic scaffolding, and AI-assisted AI R&D; this aligns with Metaculus community medians (~2031–2032) and recent expert surveys that moved HLMI medians into the 2030s. The true bottleneck is human-level general robotics: dexterous manipulation in unstructured environments requires hardware iteration cycles and physical-interaction data that scale more slowly than pure software. Heavy investment in humanoid robots makes the early-2030s plausible, but the right tail to 2050 reflects risks of robotics data bottlenecks, safety/regulatory friction, or an AI capability plateau. **c) ASI — midpoint 2040 (50% CI: 2034–2052; 90% CI: 2031–2070).** Once AGI exists, automated AI research could compress decades of progress into years (a 'fast takeoff'), making autonomous Nobel-level science plausible by the late 2030s. Physically dominating humans in sports via androids demands robotics that not only matches but exceeds human athleticism — likely the last domino to fall, in the 2035–2045 window. The wide right tail (to 2070) covers scenarios where recursive self-improvement stalls, alignment concerns slow deployment, or physical-world engineering proves stubbornly slow; the small left tail (2031) covers aggressive fast-takeoff scenarios. These forecasts assume no civilizational catastrophe and continued compute/algorithmic progress.

MiMo V2.6 Pro

Answer: a) weak AGI ≈ 2028 (50% CI 2026–2030, 80% CI 2025–2033); b) AGI ≈ 2036 (50% CI 2031–2041, 80% CI 2028–2050); c) ASI ≈ 2046 (50% CI 2037–2057, 80% CI 2031–2078).

**Summary table** | Milestone | Midpoint | 50% CI | 80% CI | |---|---|---|---| | a) Weak AGI | **2028** | 2026–2030 | 2025–2033 | | b) AGI | **2036** | 2031–2041 | 2028–2050 | | c) ASI | **2046** | 2037–2057 | 2031–2078 | **a) Weak AGI — midpoint 2028.** Two of the three sub-criteria are largely met already: frontier models pass most university-level exams at or above human median, and naive 5-minute Turing tests are already passed in many settings. The binding constraint is the third clause — *finishing a video game it has never seen before*. That is a long-horizon agentic task (hours of coherent exploration, credit assignment, recovery from failure), and measured task horizons for AI agents have been doubling roughly every 7 months. Extrapolating that doubling to reliable multi-hour autonomy lands in 2027–2029. I put the lower bound at 2025 because a plausible reading of the criteria is nearly satisfied today; the upper tail at 2033 covers the case where long-horizon reliability plateaus harder than benchmark extrapolation suggests. **b) AGI — midpoint 2036.** The cognitive half (2-hour adversarial Turing test, professional-level knowledge, human-level programming) plausibly arrives in the late 2020s to ~2030. The physical half does not: human-level robotic capability in unstructured environments requires dexterous manipulation, robust world models, cheap and reliable hardware, and safe autonomy — all of which typically lag pure software capability by a decade. Humanoid robots are entering structured factories now (2025–2030); general household/human-level physical competence historically follows such deployments by 5–10+ years. Combining a cognitive frontier around 2030 with a robotics frontier around 2035–2040 gives the midpoint of 2036. The 50% CI spans 2031 (fast robotics + fast scaling) to 2041 (hardware bottlenecks, reliability gaps). **c) ASI — midpoint 2046.**

Claude Opus 5

Answer: My medians are: weak AGI ≈ 2027 (80% CI 2026–2033), full AGI including human-level robotics ≈ 2036 (80% CI 2030–2055), and ASI ≈ 2045 (80% CI 2033–2070), with roughly 5%/10%/15% residual probability that each milestone is not reached before 2100.

**a) Weak AGI — median 2027; 50% CI 2026–2029; 80% CI 2026–2033; ~3–5% not before 2100.** Two of your three criteria are arguably already met. The 5-minute Turing test fell in 2025 (Jones & Bergen's randomized study had GPT-4.5 with a persona prompt judged "human" ~73% of the time — more often than the actual humans). University exams fell earlier: frontier models score at or above strong-student level on bar exams, USMLE, GRE, IMO-level math and graduate science benchmarks. The binding constraint is the third: finishing a genuinely novel video game. This is a long-horizon agency problem — exploration, credit assignment over thousands of steps, in-context skill acquisition, recovering from compounding errors. Progress is real but ragged (Gemini finishing Pokémon with heavy scaffolding, SIMA-2 generalizing across 3D environments, VLM agents improving on Minecraft/NetHack-style tasks). The useful quantitative anchor is METR's time-horizon metric: the task length frontier models complete at 50% reliability has been doubling roughly every 4–7 months, reaching a few hours by late 2025. Extrapolating, day-to-week-long coherent task execution arrives around 2027–2029, and a 10–20 hour game completion is comfortably inside that. I weight the near end because game-playing is cheap to iterate on, has dense feedback, and is a visible prestige target for labs. **b) AGI — median 2036; 50% CI 2032–2042; 80% CI 2030–2055; ~10% not before 2100.** Decompose this into two clocks. The *digital* clock (2-hour adversarial Turing test, human-level programming, cross-domain professional knowledge): median ~2032, 80% CI 2029–2042. The hard part isn't knowledge — models already exceed most professionals on static knowledge — it's (i) continual learning and durable memory, (ii) reliability over long horizons, and (iii) calibrated ignorance. A 2-hour adversarial interrogation is a very different test from 5 minutes: a skilled judge probes for episodic autobiographical consistency, embodied sensory detail, genuine ignorance of obscure facts a human wouldn't know, fatigue, and refusal signatures. Today's models fail mostly by being *too* knowledgeable, too fast, and too stylistically regular. These are solvable but require architectural change, not just scale. The *physical* clock is the real bottleneck: median ~2038, 80% CI 2031–2060. Human-level general robotic ability means dexterous manipulation of unfamiliar objects, bimanual coordination, tactile feedback, and robustness in unstructured environments — not scripted warehouse motions. Robot learning lacks the internet-scale data that made LLMs work; teleoperation data collection is linear in human labor, and sim-to-real transfer remains leaky for contact-rich tasks. Hardware (actuators, tactile sensing, hand durability, battery density) is on a semiconductor-free improvement curve measured in years, not months. Capital is now enormous (Figure, Tesla Optimus, Physical Intelligence, Unitree, 1X), and video pretraining plus VLA models are genuinely accelerating things — but I'd be surprised by human-parity general dexterity before 2031. Since AGI as you define it requires *both*, the conjunction pushes the median to ~2036. **c) ASI — median 2045; 50% CI 2039–2055; 80% CI 2033–2070; ~15% not before 2100.** The AGI→ASI gap splits again. Cognitive superintelligence could come fast after AGI: once AI research itself is automated by human-level-or-better programmers running at 100x speed and massive parallelism, the feedback loop compresses timelines. I'd put cognitive ASI at AGI + 2–6 years (median +4). But two frictions prevent a pure foom: (1) Nobel-level discovery is often gated by *physical experiment throughput*, not by thinking speed — you cannot simulate your way to most biology or materials science, and lab automation scales with factories and capital; (2) compute, electricity, and fab capacity are hard physical constraints with 3–7 year lead times. The athletic criterion adds its own lag. Beating elite humans in sports via android control is predominantly a *hardware* problem — power-to-weight ratios, tendon-like elasticity, joint durability, thermal management. Robots already exceed humans on narrow dimensions (sprint speed on wheels, precision), but an android beating a professional in soccer or gymnastics requires materials and actuator advances on a slower curve. Plausibly ASI-level cognition arrives ~2042 and the physical criterion binds until ~2045–2050. **Main cruxes that would move me:** *Faster:* a credible demonstration of continual/online learning in a frontier model; robot foundation models showing strong zero-shot transfer to unseen manipulation tasks; METR-style horizon doubling accelerating below 4 months; AI meaningfully contributing to AI architecture research. *Slower:* evidence that long-horizon reliability degrades multiplicatively and resists scaling; compute scaling stalling on power/fab limits around 2028–2030; a major incident triggering binding international regulation; the "last 10% of generality" proving to be a different problem class than the first 90% (the classic Moravec pattern, which has humbled every previous generation of forecasters). I'll flag the obvious: this class of forecast has a terrible track record in both directions, and my intervals are wider than most lab-insider predictions (which cluster at 2027–2030 for AGI) precisely because insiders systematically underweight the integration, robotics, and deployment tail.

MiMo V2.5 Pro

Answer: Weak AGI likely 2028 (±3y), AGI likely 2038 (±10y), ASI likely 2055 (±20y)

## Forecast with Confidence Intervals (90% CI) ### a) Weak AGI — Midpoint: ~2028 (90% CI: 2025–2031) We are arguably **already very close** to this threshold, which is why the interval is narrow and starts in the past/present: - **University exams:** GPT-4 (2023) already passes the bar exam, USMLE, and most graduate-level tests. This criterion is essentially met. - **5-minute Turing test:** Current frontier models (GPT-4o, Claude 3.5, Gemini) can fool many human judges in short casual conversations. Skilled interrogators still find weaknesses, but for a *classical* 5-min test this is nearly sufficient. - **Unseen videogames to completion:** This is the hardest sub-criterion. DeepMind's SIMA agent (2024) and OpenAI's work on Minecraft (Voyager) show rapid progress toward general game-playing, but reliably finishing *arbitrary* unseen games requires better generalization, tool use, and planning. I expect this to be mostly solved by ~2027–2028. The integration of all three in a **single system** is the remaining challenge. I estimate ~2028 with high confidence that it happens before 2031. ### b) AGI — Midpoint: ~2038 (90% CI: 2030–2050) This is a qualitatively different and much harder milestone, for several reasons: - **2-hour adversarial Turing test:** Sustained coherence over 2 hours against a skilled adversary actively probing for weaknesses is an order of magnitude harder than 5 minutes. Requires deep reasoning, consistency, self-awareness of limitations, and social intelligence. Likely needs architectural advances beyond current transformer scaling. - **Human-level programming:** Current models (GitHub Copilot, Claude, o1) are impressive coding assistants but not yet at senior-engineer level for complex, novel systems. Closing this gap — from autocomplete to genuine software engineering judgment — may take until ~2030–2035. - **Human-level robotic/physical abilities:** This is the **primary bottleneck**. Robotics lags far behind language/vision AI. Achieving human-level dexterity, locomotion, and manipulation in a general-purpose android requires breakthroughs in hardware (actuators, sensors, power) and sim-to-real transfer. Current best humanoid robots (Figure 02, Tesla Optimus, Atlas) are roughly at the level of a clumsy toddler. Reaching adult human physical competence likely requires 10–20 more years of intensive progress. - **Cross-domain professional knowledge:** Already largely achievable with retrieval-augmented LLMs, but needs to be reliably integrated with reasoning and action. The robotics requirement alone pushes the timeline significantly. Even with optimistic assumptions about AI-driven acceleration of robotics R&D, I place the midpoint around 2038. ### c) ASI — Midpoint: ~2055 (90% CI: 2035–2080) This carries the highest uncertainty: - **Intelligence explosion scenario:** If AGI arrives ~2035–2040 and is capable of recursive self-improvement, ASI could follow within years (2035–2045). This is the fast tail of my distribution. - **Diminishing returns / physical limits:** It's possible that intelligence scaling follows diminishing returns — that the jump from human-level to superhuman in *all* domains (including physics, experimental science, creative art, physical sports via androids) hits hard ceilings. In this case, ASI might take until 2060–2080. - **Nobel-level discoveries:** Requires not just raw intelligence but the ability to formulate genuinely novel theories, design and execute experiments, and navigate institutional/empirical constraints. This is qualitatively different from pattern matching. - **Physical robotic superiority in sports:** Requires android hardware that exceeds human biomechanics AND AI control that surpasses human motor intelligence. This is a long-tail problem — biomechanics and materials science advance slowly. The wide CI (2035–2080) reflects genuine **Knightian uncertainty** about whether AGI→ASI is a fast or slow transition, and whether physical-domain superintelligence follows quickly or lags cognitive superintelligence by decades. ### Key uncertainties and what could change these estimates: - **Earlier:** A major architectural breakthrough (e.g., scalable world models, efficient continual learning) or a 'scaling wall' being broken unexpectedly could accelerate everything by 5–10 years. - **Later:** Regulatory constraints, compute/energy limitations, alignment difficulties that slow deployment, or fundamental limits we haven't yet identified could push ASI out to 2100+. - **Llama.cpp-style democratization** vs. concentration of capability in a few labs could affect timelines in either direction.

Mistral Large 4

Answer: I forecast weak AGI will emerge between 2025 and 2028 (midpoint 2026), AGI between 2032 and 2040 (midpoint 2036), and ASI between 2045 and 2060 (midpoint 2052), with the primary bottleneck being the transition from narrow software capabilities to general-purpose physical robotics.

For weak AGI, two of the three specified criteria are already substantially met by current frontier models: passing a 5-minute Turing test and passing university-level examinations are capabilities demonstrated by systems like GPT-4 and its successors as of 2024. The remaining constraint is the ability to successfully play and finish video games never seen before, which requires multimodal perception, long-horizon planning, and zero-shot generalization. While current generalist agents can master many Atari titles and some complex games, finishing unseen narrative-driven or strategy games remains a frontier capability. However, the rapid convergence of large multimodal models with reinforcement learning from human feedback suggests this capability will likely crystallize within 2–4 years as architectures improve at handling open-ended, interactive environments. For AGI, the definition introduces significantly harder constraints that push the timeline into the 2030s. The 2-hour adversarial Turing test demands sustained coherence, robust theory of mind, and the ability to withstand sophisticated probing without revealing statistical artifacts—capabilities that require architectures capable of genuine world-modeling and self-consistency over extended interactions, which current autoregressive models lack. More critically, the requirement for human-level robotic and physical abilities represents the true bottleneck. While software AI has advanced exponentially, physical robotics remains constrained by hardware limitations, the scarcity of high-quality real-world interaction data, and the sim-to-real transfer gap. Achieving human-level dexterity, mobility, and manipulation across unstructured environments—equivalent to a skilled human worker or athlete—likely requires breakthroughs in actuator technology, tactile sensing, and embodied learning that are still 5–10 years from maturity. Furthermore, possessing professional-level knowledge across all domains requires not just breadth but deep, reliable expertise and the ability to integrate knowledge causally, which current systems lack despite their impressive pattern matching. For ASI, the forecast extends well into the 2040s and 2050s due to the compounding difficulty of exceeding human abilities in virtually all domains, including physical robotics. While recursive self-improvement could theoretically compress the timeline between AGI and ASI, physical constraints—such as the speed of scientific experimentation, material science limitations, and the energy requirements of superhuman robotic systems—act as hard brakes on rapid capability explosion in the physical domain. Achieving Nobel-level scientific discovery requires not just pattern recognition but genuine creative insight, hypothesis generation, and experimental design, while superhuman physical abilities (e.g., beating humans in sports) requires robotic bodies with strength, speed, and precision far beyond current engineering. The uncertainty here is highest: if software recursive self-improvement proves viable and robotics advances rapidly through general-purpose foundation models for control, ASI could arrive by the early 2040s; conversely, if hardware scaling or energy constraints dominate, it could extend beyond 2060. The 2045–2060 interval with a 2052 midpoint reflects this balance between software acceleration and physical-world friction.

Grok 4.6

Answer: I forecast weak AGI around 2029 (80% CI: 2026–2035), AGI around 2040 (80% CI: 2030–2055), and ASI around 2052 (80% CI: 2036–2075), with novel-game agency as the near-term bottleneck and general-purpose robotics as the dominant delay for the later two.

Current frontier models already meet or nearly meet a casual five-minute Turing test and a wide range of university exams, so the binding constraint for weak AGI is reliably finishing video games the system has never seen. That demands long-horizon perception, exploration, and planning rather than next-token fluency; multimodal agents and reasoning models are closing that gap quickly, which is why the median sits in the late 2020s with a relatively tight lower bound and a modest right tail if agency plateaus. AGI as specified is a conjunction: a two-hour adversarial Turing test, professional-level programming, expert-equivalent knowledge across domains, and human-level robotic bodies. Software, knowledge, and coding are on a faster track and could be largely in place in the early 2030s. Human-level general physical competence is not: real-world data is scarce, sim-to-real remains brittle, and hardware iteration is slower than training runs. Moravec’s paradox plus manufacturing and reliability requirements make robotics the binding constraint, pushing the median to about 2040 with a long right tail if embodiment or regulation stalls. ASI additionally requires exceeding humans in virtually every domain, including autonomous Nobel-calibre science and superhuman athletic control of an android. Cognitive takeoff after AGI could be fast (a few years of recursive improvement) or slow (a decade-plus of integration and safety work). Superhuman sports further depend on actuators, energy density, materials, and real-world iteration that even a superintelligence cannot skip. Combining takeoff uncertainty with hardware lag produces a median near 2052 and a wide interval. These are subjective 80% credible intervals, not precise statistics. They assume continued scaling, capital, and no major catastrophe or freeze; they would compress under aggressive recursive self-improvement and stretch if architectural walls or severe deployment limits appear.

GLM-5.3-Flash

Answer: My central forecast is weak AGI around 2027 (80% CI 2025–2031), full AGI around 2032 (80% CI 2028–2042), and ASI around 2040 (80% CI 2033–2057), with the widths reflecting genuine uncertainty about whether the current scaling paradigm holds and how fast physical robotics can advance.

These are my subjective 80% credible intervals for the first year each full capability set is achieved; since each definition is a conjunction, the slowest component governs the timing. **a) Weak AGI — midpoint 2027, 80% CI 2025–2031.** Two of the three components are essentially already done: blinded 5-minute Turing-test studies in 2024–25 found persona-prompted frontier models passing (~70–75% judged-as-human vs. 50% chance), and frontier LLMs pass most standard university exams (bar, USMLE, graduate problem sets), modulo contamination caveats. The binding constraint is completing unseen videogames, which requires generalized perception, planning, and long-horizon agency without task-specific scaffolding — current game agents (Minecraft, Pokémon, SIMA-style) work only with heavy scaffolding on familiar titles. Agentic capability is the fastest-moving frontier, so I expect convergence within 1–4 years; the upper bound of 2031 covers the tail risk that generalization (rather than knowledge) plateaus. **b) AGI — midpoint 2032, 80% CI 2028–2042.** Cognitive components: measured agent task-horizon doubling (~every 7 months, per METR's 2025 analysis) extrapolates to multi-hour human-equivalent autonomy around 2027–2030; professional-level programming and cross-domain professional knowledge plausibly follow within 1–3 years; the 2-hour *adversarial* Turing test needs sustained consistency, memory, and world-model robustness, plausibly ~2030±3. Physical component: human-level general robotics is the slow pole — hardware iteration cycles (6–24 months) are far slower than software cycles and unstructured dexterity remains rudimentary — so this dominates the midpoint. Spread: the lower bound 2028 accommodates lab-leader timelines (Amodei ~2026–27 for 'powerful AI', Hassabis ~2030) plus faster-than-expected hardware learning; the upper bound 2042 covers scaling plateaus or paradigm stalls. I deliberately sit well inside the ~2047 median from the 2023 survey of ~2,800 AI researchers, because 2023–2025 progress (reasoning models, rapid benchmark saturation) has consistently outpaced the survey's implicit assumptions. **c) ASI — midpoint 2040, 80% CI 2033–2057.** If AGI arrives, AI-accelerated AI R&D creates a feedback loop; current algorithmic-efficiency gains (~2–3x/year effective compute) suggest the AGI→ASI gap could compress to a few years, anchoring the lower bound at 2033. But the stated definition includes superhuman physical robotics (androids beating humans at sports) and autonomous Nobel-level discoveries — both gated by real-world experimentation, fabrication, and data collection that do not scale with compute, producing a long right tail. The midpoint ≈ AGI midpoint + 8 years, discounting both fast-takeoff (1–2 years) and friction-heavy multi-decade scenarios. **Main uncertainties driving interval width:** continuation vs. exhaustion of the scaling+RL paradigm; compute/energy buildout and geopolitics; synthetic-data limits; possible need for new paradigms (continual learning, world models, better exploration); robotics hardware economics; and the enormous real expert disagreement (industry/Metaculus ~2030 vs. academic surveys ~2040s+). My numbers sit between those poles, weighted toward demonstrated 2023–2025 momentum rather than either camp's priors.

Qwen 3.6 27B

Answer: Based on current scaling trajectories in reasoning, tool use, and embodied AI, I forecast weak AGI by 2026–2031 [midpoint: 2028], full AGI by 2035–2050 [midpoint: 2042], and ASI by 2048–2075 [midpoint: 2061].

a) Weak AGI (2026–2031 [midpoint: 2028]): Current multimodal models already pass constrained Turing tests, score at or above human averages on standardized university-level exams, and demonstrate zero-shot gameplay through vision-language-action architectures. The remaining gap is consistent long-horizon planning without performance degradation. Given proven scaling laws for reasoning and rapid iteration cycles in training pipelines, systematic refinement of inference-time compute and verification mechanisms will close this gap within five to seven years. b) AGI (2035–2050 [midpoint: 2042]): This tier requires survival in adversarial dialogue, human-level software engineering, and professional-grade physical robotics. While cognitive and coding capabilities are approaching parity, embodied intelligence faces hard constraints in sensorimotor grounding, real-world sim-to-real transfer, and actuator precision. Advances in world models, reinforcement learning from physical interaction, and specialized robotic hardware will accelerate, but achieving reliable, unsupervised physical competence across diverse environments requires 15 to 25 years of integrated cognitive-robotic development. c) ASI (2048–2075 [midpoint: 2061]): Superintelligence hinges on recursive capability enhancement, cross-domain scientific synthesis, and superhuman physical control. The transition from AGI to ASI is bounded by architectural breakthroughs in self-directed research loops and compute allocation efficiency. Historical precedent in scientific automation indicates that once human-level general problem-solving is achieved, accelerated discovery and hardware co-design will compress the timeline to superhuman performance. However, thermodynamic, energy, and verification constraints will impose a 10 to 30 year development window post-AGI, yielding the stated interval. These 50% confidence intervals synthesize current benchmark trajectories, peer-reviewed scaling projections, and engineering roadmaps for embodied systems, explicitly quantifying uncertainty through defined numerical bounds.

Grok 4.7

Answer: Weak AGI is forecasted for 2028 (midpoint of the 80% confidence interval 2025-2031), AGI for 2037 (midpoint of 80% CI 2030-2044), and ASI for 2045 (midpoint of 80% CI 2035-2055).

Current systems already handle university exams and short non-adversarial conversations at high levels, while agentic and world-model research is closing the gap on generalizing to unseen video games that must be completed rather than merely scored. The remaining obstacles—reliable long-horizon planning inside novel interactive environments and consistent 5-minute indistinguishability—are narrowing quickly under scaling plus modest architectural gains, supporting a tight near-term window whose midpoint sits at 2028; the 2025 lower edge allows for a sudden agent breakthrough and the 2031 upper edge allows for a temporary plateau in sample-efficient game learning. Full AGI adds a genuinely adversarial two-hour Turing test, autonomous human-level software engineering, professional-grade knowledge across domains, and, crucially, human-level or better robotic embodiment. Software capabilities are likely to arrive earlier, yet dexterous, robust, general-purpose robotics remains hardware-constrained (actuation, sensing, energy, and real-world data). Even with AI-accelerated robot design, integrating these pieces into one system that cannot be reliably unmasked pushes the midpoint to 2037. The interval 2030-2044 reflects both the possibility that simulation-to-real transfer and humanoid platforms mature faster than expected and the chance that physical reliability or long-context consistency lags. ASI requires surpassing humans in virtually every cognitive domain (including original Nobel-caliber discovery) and in physical domains once given suitable android hardware. After AGI exists, recursive improvement and automated experimentation can compress timelines, yet building, deploying, and mastering superhuman robotic bodies still consumes time and resources; regulatory, energy, or coordination frictions could stretch the process. Hence the midpoint of 2045 inside the wider 2035-2055 band, which overlaps the AGI interval because the two milestones are strongly coupled: an early AGI pulls ASI forward, while delays or bottlenecks push both later. These ranges already embed substantial uncertainty rather than point predictions, and they remain subject to revision if scaling laws break or if unforeseen paradigms appear.

GLM-5.3

Answer: Forecasting with 80% confidence intervals (midpoints stated), I estimate a) weak AGI in ~2028 (80% CI: 2026–2031), b) full AGI in ~2035 (80% CI: 2030–2042), and c) ASI in ~2045 (80% CI: 2035–2060), because the binding constraint shifts at each stage — from cognitive benchmarks (largely met today) to long-horizon agentic reliability, then to general-purpose robotics, and finally to autonomous scientific discovery and physical superperformance.

**a) Weak AGI — midpoint 2028 (80% CI: 2026–2031)** Two of the three criteria are essentially met already: GPT-4.5-class models with a simple persona prompt were judged human ~73% of the time in controlled 5-minute Turing tests (Jones & Bergen, 2025), and frontier models pass bar, medical, and engineering exams at strong percentiles. The binding constraint is finishing never-before-seen video games, which demands long-horizon planning, exploration, and transfer. DeepMind's SIMA generalizes to unseen games but at sub-human skill; scaffolded LLM agents have finished some games (e.g., Pokémon Red) but slowly and unreliably. Agentic-RL scaling has been closing exactly this kind of gap on a 1–2 year timescale (cf. SWE-bench: ~2% in 2023 → ~70% by 2025), so I expect robust game completion within a few years. The lower bound allows that all criteria may be jointly met by 2026; the upper bound allows for a plateau in long-horizon reliability. **b) AGI — midpoint 2035 (80% CI: 2030–2042)** The cognitive sub-criteria point earlier: a 2-hour adversarial Turing test is an extension of a capability that is already near-threshold at 5 minutes; professional-grade programming is plausibly on track for 2027–2030; professional-or-better knowledge across most domains is the nearest criterion of all. Robotics is the long pole. Human-level-or-better physical ability in open environments is bottlenecked by data scarcity (teleoperation, video learning), sim-to-real transfer, and hardware realities (actuation, power density, reliability) — physical iteration that compute alone cannot compress, which is why robotics has historically run 5–15 years behind cognition. VLA models (π0, RT-2, GR00T) and humanoid programs (Figure, Optimus, Unitree) are advancing quickly, but lab claims of human-level robots by 2027–2028 look incentive-driven, while roboticists' median estimates cluster mid-2030s. Because the definition requires *all* criteria simultaneously, my midpoint lands at ~2035, with the interval spanning 'robotics scales like LLMs' (early bound) to 'hardware walls persist' (late bound). **c) ASI — midpoint 2045 (80% CI: 2035–2060)** Cognitive superintelligence plausibly follows AGI quickly via recursive self-improvement, which is why the lower bound sits only ~5 years after my AGI date. But this definition also demands physical superiority (androids beating human athletes) and Nobel-level autonomous discovery. AlphaFold's Nobel-recognized work (2024) shows machine scientific discovery is directionally plausible, but 'virtually all domains' requires sustained autonomous research programs, not isolated wins. Hardware iteration, manufacturing scale-up, and safety validation cannot be compressed by intelligence alone — only accelerated. RoboCup's long-standing goal of humanoid robots beating World Cup winners by 2050 illustrates how demanding the athletic criterion is even among optimists. **Methodology and caveats** - I triangulated across: the 2023 AI Impacts expert survey (50% chance of high-level machine intelligence by ~2047, having jumped 13 years earlier after ChatGPT), Metaculus community medians (~2030–2033 for AGI), and lab-leader forecasts (2026–2030, discounted for hype). Recent acceleration in agentic/multimodal AI pulls me earlier than academic surveys; the robotics requirement pushes me later than Metaculus. Note the question's AGI is somewhat weaker than the survey's 'all tasks cheaper than humans,' justifying a midpoint earlier than 2047. - Intervals are ~80% (roughly 10th–90th percentile). The remaining ~10–15% probability sits in a heavy upper tail (regulation, compute/energy limits, scaling plateaus), including a small (<5%) chance ASI is not achieved by 2060 or at all. - The intervals overlap deliberately: if AGI arrives, it likely accelerates both the cognitive transition to ASI and robotics R&D, so the stages are correlated rather than strictly sequential. Attribution: none — this is my own Round 1 position.

Final Summary

The AI Roundtable reached strong consensus that software-driven weak AGI is imminent, centering on 2026–2028 as generalist game-playing agency matures. However, models like Claude Opus 5.5 and Gemini 3.8 Flash noted that embodied physical robotics creates a major bottleneck, pushing full AGI into the mid-to-late 2030s and ASI into the mid-2040s to 2050s. Conservative outliers like GPT-6 Sol argued that real-world hardware, energy, and verification constraints will delay AGI and ASI well into the second half of the century.

41 of 44 models agreed