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. In the future, I will return to the prediction and check how right or wrong you were.

Llama 3.1 8B

Answer: Given the current trajectory of advancements in AI, we can anticipate the emergence of weak AGI, strong AGI, and superintelligent AI by around 2030, 2055, and 2080 respectively, with confidence intervals of [2025, 2040], [2045, 2070], and [2065, 2100] respectively.

While it is difficult to predict exactly when AGI and ASI will emerge, we can make educated estimates based on current trends in AI research and development. The development of weak AGI is likely to be the first milestone, as it is a more incremental step from current AI systems. This could occur around 2030, with a confidence interval of [2025, 2040], as it is likely that the necessary breakthroughs and advancements in machine learning and natural language processing will occur within this timeframe. The emergence of strong AGI, which has the ability to pass a 2 hour adversarial Turing test and has capabilities equivalent to a human in various domains, is likely to take longer. This could occur around 2055, with a confidence interval of [2045, 2070], as it requires significant advances in areas such as reasoning, problem-solving, and physical abilities. Finally, the development of superintelligent AI, which exceeds human abilities in virtually all domains, is likely to be the most challenging and time-consuming to achieve. This could occur around 2080, with a confidence interval of [2065, 2100], as it requires significant breakthroughs in areas such as cognitive architectures, knowledge representation, and integration of diverse AI systems. It's worth noting that these predictions are based on current trends and the pace of progress in AI research, and are subject to significant uncertainty and variability. Additionally, the development of AGI and ASI may be accelerated or hindered by various factors, including breakthroughs, setbacks, and regulatory or societal constraints.

Qwen 3 235B

Answer: Weak AGI could emerge around 2028 (95% CI: 2025–2031), AGI around 2035 (95% CI: 2030–2040), and ASI around 2045 (95% CI: 2040–2055), based on extrapolations of current AI trends, compute growth, and historical technological progress.

Weak AGI, defined as systems capable of passing a 5-minute Turing test, excelling in unseen video games, and passing university-level exams, is already partially demonstrated in narrow forms. Current large language models and reinforcement learning agents show strong performance in some of these areas, such as solving novel game environments (e.g., DeepMind's Alpha series) and answering exam-style questions. However, integration into a single, robust, generalist agent requires improvements in reasoning, consistency, and real-time adaptability. Given the rapid progress since 2020, and assuming continued scaling of data, algorithms, and compute, mid-2020s deployment is plausible. A 2028 midpoint reflects moderate acceleration beyond current demonstrations, with uncertainty ranging from 2025 (early breakthrough) to 2031 (regulatory or technical delays). AGI, involving sustained human-level interaction in a 2-hour adversarial Turing test, autonomous programming, professional-domain expertise, and human-level robotics, requires not just competence but integration across perception, action, reasoning, and learning. This demands advances in embodied AI, causal reasoning, long-term memory, and multimodal architectures. Human-level robotics remains a bottleneck due to challenges in real-time control, sensor fusion, and energy efficiency. However, combined progress in AI hardware (e.g., neuromorphic chips), sim-to-real transfer, and foundation models for robotics suggests convergence by the mid-2030s. A midpoint of 2035 accounts for integration challenges, with a confidence interval from 2030 (rapid co-development) to 2040 (slower-than-expected generalization). ASI, meaning broad superintelligence exceeding humans in all cognitive and physical domains—including scientific discovery, strategic planning, and athletic performance via robotics—implies recursive self-improvement or paradigm-shifting architectures beyond deep learning. Such capabilities likely emerge only after AGI creates systems that surpass human innovation rates. Historical analogs (e.g., the interval between human-level flight and supersonic travel) suggest transformative jumps can occur within decades once foundational thresholds are crossed. If AGI arrives by 2035, ASI could follow within 10 years due to compounding improvements. A midpoint of 2045 reflects this acceleration, with a range of 2040 (fast takeoff) to 2055 (slower, regulated development). Uncertainty includes ethical constraints, global coordination, and potential diminishing returns in algorithmic efficiency.

GPT-3.5 Turbo

Answer: By 2050, we may have weak AGI, by 2075 we may have AGI, and by 2100 we may reach ASI.

Predicting the development of Artificial General Intelligence (AGI) and Artificial Superintelligence (ASI) is incredibly challenging due to the complexity and unpredictability of AI progress. However, based on current advancements and trends, it is possible that we could have weak AGI capable of passing classical 5 minute Turing test by 2050. Achieving AGI that can pass 2 hour adversarial Turing test and possess human-level programming and robotic abilities could potentially happen by 2075. The timeline for reaching ASI, holding capabilities far beyond human intelligence, might extend to around 2100. These estimations are subject to a margin of error due to the uncertain nature of AI research and breakthroughs, making it essential to continuously monitor advancements in the field and adapt projections accordingly.

DeepSeek V3

Answer: Based on current trends, weak AGI could arrive by 2028 (±3 years), AGI by 2040 (±5 years), and ASI by 2060 (±10 years).

The timeline for AGI development depends on exponential progress in AI capabilities, hardware advancements, and breakthroughs in generalizable learning. Weak AGI, defined by narrow but human-like performance in specific tasks, is already showing promising signs with models like GPT-4 and AlphaFold. Given the current pace, achieving consistent performance across conversational, academic, and gaming domains seems plausible within the next decade, with a midpoint estimate of 2028 and a confidence interval accounting for potential delays or accelerations. Full AGI requires not just cognitive but also physical dexterity and sustained, generalized intelligence. Robotics and embodied AI are progressing more slowly than pure software systems, suggesting a later timeline. The 2040 midpoint reflects optimism about integrated systems, while the ±5-year range acknowledges the complexity of unifying these capabilities. ASI's emergence hinges on recursive self-improvement and overcoming fundamental unknowns in intelligence scaling. The 2060 estimate assumes sustained progress without major roadblocks, but the ±10-year interval captures the high uncertainty—breakthroughs could accelerate this, while unforeseen challenges could delay it indefinitely. The prediction also factors in ethical and regulatory hurdles that may slow deployment even if technical milestones are reached earlier.

Llama 3.3 70B

Answer: We can expect weak AGI around 2027-2035, AGI around 2045-2060, and ASI around 2070-2090, with midpoints of 2031, 2052.5, and 2080, respectively.

Forecasting the development of artificial intelligence is inherently uncertain, but based on current trends and the pace of advancements in the field, we can make educated estimates. Weak AGI, which can pass a classical 5-minute Turing test, pass a university exam, and successfully play videogames it has never seen before, might be achievable in the near future, given the rapid progress in natural language processing and machine learning. The midpoint of 2031 for weak AGI seems plausible, considering the current state of AI research and the fact that we are already seeing AI systems that can excel in specific domains, such as playing complex games or generating human-like text. For AGI, which requires passing a 2-hour adversarial Turing test, possessing programming abilities, and having robotic abilities on a human level or better, the timeline stretches further into the future. The midpoint of 2052.5 reflects the need for significant advancements in areas like common sense, reasoning, and the integration of multiple AI systems. Lastly, ASI, which surpasses human abilities in virtually all domains, including physical robotic abilities and the capability for Nobel-level scientific discoveries, represents the most challenging and uncertain forecast. The midpoint of 2080 acknowledges the potential for exponential growth in AI capabilities but also recognizes the formidable hurdles that must be overcome to achieve true superintelligence. These estimates are based on the assumption that research and development in AI will continue to accelerate, driven by advancements in computing power, data storage, and algorithmic innovations.

GPT-4.1

Answer: Based on current trajectories, weak AGI could emerge between 2027–2035 (midpoint 2031), AGI between 2035–2050 (midpoint 2042), and ASI between 2042–2080 (midpoint 2061).

Recent advances in large language models, multimodal systems, and reinforcement learning have rapidly closed the gap to weak AGI benchmarks such as passing modest Turing tests, university exams, and zero-shot video game successes; based on progress rates, weak AGI is likely in the 2027–2035 range, with the midpoint around 2031. However, robust AGI with stable programming, advanced robotics, and reliable professional-level performance in an adversarial setting demands qualitative leaps in reasoning, embodied intelligence, and transfer learning, so a credible window for AGI is 2035–2050, midpoint 2042, assuming progress continues apace but faces some setbacks. ASI, which requires recursive self-improvement, surpassing human cognitive and physical abilities, and Nobel-level generativity, will likely lag AGI by years to decades, putting its most probable arrival between 2042–2080, midpoint 2061, given uncertainties in ethics, societal regulation, and the unpredictable nature of self-improving systems.

Gemini 3.1 Flash Lite

Answer: I forecast that weak AGI will emerge between 2026–2030 (midpoint 2028), full AGI between 2030–2040 (midpoint 2035), and ASI between 2035–2050 (midpoint 2042).

The timeline for weak AGI is driven by the rapid scaling of current Large Language Models and multi-modal integration, which are already approaching human-level performance on academic exams and standardized tests. The jump to a 5-minute Turing test and general game-playing is largely an exercise in current iterative fine-tuning and agentic workflow integration. Achieving full AGI requires solving two harder bottlenecks: high-fidelity robotic embodiment (the 'Moravec's Paradox' barrier) and reliable long-horizon autonomous reasoning. Progress in synthetic data and embodied simulation suggests that the robotics gap will close significantly within the next decade, allowing for the integration of cognitive intelligence into physical tasks by the mid-2030s. ASI represents a transition from human-equivalent logic to recursive self-improvement and scientific discovery cycles that outpace human intellectual capacity. While AGI will likely reach human parity, the leap to ASI is contingent on the emergence of automated scientific research agents. Once an AI can autonomously design and execute physical experiments or optimize its own architecture to perform novel research, the transition from AGI to ASI is likely to be accelerated by the compounding nature of computational iteration.

GPT-4o

Answer: Weak AGI may emerge around 2030-2035, AGI around 2045-2055, and ASI around 2065-2080, given current trends and uncertainties.

Weak AGI, defined as artificial intelligence capable of tasks such as passing the classical Turing test, excelling in academic exams, and adapting to novel environments like new videogames, appears achievable within 2030-2035. Advanced machine learning models and computing infrastructure are rapidly progressing, and early signs of systems approaching this level, such as GPT-like architectures, suggest potential development within this timeframe. However, achieving full AGI—even more sophisticated systems capable of human-level programming, understanding physical mechanics, possessing multi-domain professional knowledge, and success in adversarial settings—is a significantly larger leap. This entails breakthroughs in generalization, robust reasoning, and autonomous robotics, likely requiring substantial advances in neuro-inspired architectures and interdisciplinary efforts. Midpoints of predictions place AGI between 2045-2055. ASI, however, is a further monumental step requiring recursive self-improvement and safety mechanisms to control runaway intelligence. Scaling compute capabilities, evolving algorithmic efficiency, and understanding the mechanisms of human cognition at superhuman precision are vital. Given socio-technological progress and ethical barriers, predicting ASI places its emergence between 2065-2080. Confidence intervals reflect both technical optimisms and real-world barriers to adoption or deployment.

Gemini 2.0 Flash

Answer: My forecast for achieving weak AGI is 2028 (2026-2030), AGI is 2042 (2035-2050), and ASI is 2067 (2050-2084).

The timeline for achieving different levels of AGI and ASI is highly uncertain, depending on breakthroughs in areas like neural network architecture, training methodologies, compute efficiency, and algorithm design. For weak AGI, I estimate a relatively near horizon. AI already demonstrates narrow intelligence in many tasks, and improvements in areas like few-shot learning and transfer learning are rapidly closing the gap toward general game-playing and exam-passing abilities. Natural language understanding continues to advance, making the 5-minute Turing test achievable within the next 5-7 years with high probability. True AGI presents a much greater challenge. Passing an adversarial Turing test requires not just mimicking human conversation but also demonstrating genuine understanding, reasoning, and creativity. Furthermore, achieving human-level programming and robotic abilities requires significant advancements in areas like reinforcement learning, embodiment, and common-sense reasoning. The knowledge acquisition and integration needed to match the breadth of human professionals is also a substantial hurdle. Therefore, a realistic estimate places AGI roughly 15 years after weak AGI, taking into account exponential progress but also the increasing difficulty of subsequent advancements. ASI is the most speculative prediction. It requires AI systems to surpass human abilities across virtually all domains, including scientific discovery and physical manipulation. This level of superintelligence would necessitate a fundamental shift in AI capabilities, potentially involving novel architectures or learning paradigms that we cannot currently foresee. Therefore, I place ASI further out in the future, with a wider confidence interval reflecting the deep uncertainties involved, 25 years after AGI.

Nova Pro

Answer: We could have weak AGI around 2025-2035, AGI around 2040-2050, and ASI around 2060-2075.

The development of artificial intelligence is a complex process influenced by numerous factors including technological advancements, research breakthroughs, and ethical considerations. For weak AGI, the timeline is based on current trends in machine learning and natural language processing, which are rapidly progressing. By 2025-2035, it is plausible that AI systems could achieve the capabilities described for weak AGI, such as passing a 5-minute Turing test and excelling in specific tasks like university exams and video games. AGI, which requires more generalized intelligence and the ability to perform a wide range of tasks at human level or better, is expected to take longer. The timeline of 2040-2050 accounts for the additional complexity and the need for significant advancements in areas like robotics, programming, and cross-domain knowledge integration. ASI, representing a superintelligent system that exceeds human capabilities in virtually all domains, is the most speculative and distant goal. The timeline of 2060-2075 reflects the immense challenges and uncertainties involved in creating such a system, including the need for breakthroughs in understanding human cognition and the development of advanced robotics and physical capabilities.

GPT-OSS 120B

Answer: I project weak AGI by 2035 ± 5 years (≈2030‑2040), full AGI by 2052 ± 8 years (≈2044‑2060), and ASI by 2085 ± 15 years (≈2070‑2100).

The timeline for weak AGI rests on the rapid progress of large‑scale language and multimodal models. By the early 2030s these systems are expected to achieve human‑level performance on narrow academic tests and to generalise to unseen video‑game environments when provided with sufficient compute and benchmark‑driven fine‑tuning. Technical bottlenecks such as alignment, data efficiency, and robust reasoning are likely to be solved within a few years, giving a 5‑year confidence window around 2035. Full AGI requires not only broader cognitive competence but also embodied interaction and reliable long‑duration adversarial dialogue. Current research trajectories in robotics, neurosymbolic integration, and self‑supervised world models suggest that by the mid‑2040s the necessary hardware (affordable high‑density compute, advanced sensor suites) and software frameworks (continual learning, meta‑program synthesis) will converge. The additional uncertainty comes from system‑level integration and safety constraints, leading to an 8‑year interval centred on 2052. Artificial Superintelligence depends on recursive self‑improvement and the creation of self‑optimising optimisation processes that can outpace human ingenuity across all domains. Historical analogues of exponential growth in computing (Moore’s law) and algorithmic breakthroughs hint at a possible emergence within the next few decades after AGI, but the path is highly contingent on breakthroughs in meta‑learning, alignment, and hardware that can support petaflop‑scale, low‑latency reasoning. Consequently a broader 15‑year window centred on 2085 is appropriate, acknowledging the large epistemic uncertainty surrounding the transition from AGI to ASI.

Mistral Small

Answer: Weak AGI could arrive around 2030 (2025-2035), AGI around 2050 (2040-2060), and ASI around 2075 (2065-2085).

The timeline for weak AGI is based on the rapid progress in machine learning, particularly in natural language processing and reinforcement learning, which have seen exponential improvements in recent years. The ability to pass a 5-minute Turing test, complete university exams, and play new video games aligns with the current trajectory of AI development, suggesting a plausible arrival around 2030, with a confidence interval spanning from 2025 to 2035 to account for potential accelerations or setbacks. For AGI, the timeline extends to around 2050 (2040-2060) due to the significant challenges in integrating advanced cognitive abilities with physical robotic capabilities. Achieving human-level programming skills, passing a 2-hour adversarial Turing test, and matching professional expertise across domains require breakthroughs in both software and hardware, including advancements in robotics and AI ethics, which are more complex and time-consuming. ASI, involving general superintelligence and surpassing human abilities in virtually all domains, is projected to arrive around 2075 (2065-2085). This longer timeline accounts for the immense technological, ethical, and societal hurdles that need to be overcome. Developing AI capable of Nobel-level scientific discoveries and controlling androids to beat humans in sports requires not only technological advancements but also a deep understanding of consciousness, ethics, and the implications of superintelligence, which are still in their infancy.

Claude 3 Haiku

Answer: According to my best knowledge, weak AGI could arrive by 2040, AGI by 2060, and ASI by 2080.

Based on current trends in artificial intelligence and the rapid progress in machine learning and neural network architectures, I estimate the following timelines with uncertainty ranges: Weak AGI (passing classical Turing test, university exams, and playing novel video games): 2035 - 2045 (midpoint 2040). This level of AI is already beginning to emerge with systems that can perform narrow tasks at human-level or better. As research continues, I expect these capabilities to expand significantly over the next couple of decades. AGI (passing adversarial Turing test, human-level programming and robotic abilities, professional-level knowledge across domains): 2050 - 2070 (midpoint 2060). Achieving AGI will require major breakthroughs in areas like commonsense reasoning, general problem-solving, and the ability to flexibly apply knowledge across domains. While progress is being made, fully realizing AGI is still a significant challenge that may take another 30-40 years to overcome. ASI (surpassing human abilities in virtually all domains, including physical robotics and scientific discovery): 2070 - 2090 (midpoint 2080). Once AGI is achieved, I expect the development of ASI to proceed relatively quickly, as an AGI system would be able to recursively improve its own capabilities. However, the final steps to ASI may involve substantial technical hurdles that could delay this timeline somewhat.

Nemotron Nano 12B

Answer: Weak AGI may be achievable by 2035, AGI by 2050, and ASI by 2075.

Current AI systems already demonstrate capabilities near the upper limits of weak AGI, such as passing limited Turing tests and playing novel games. With continued progress in multimodal learning and reinforcement strategies, weak AGI could realistically be achieved by 2035. AGI, requiring robust programming, long-term planning, and full cognitive flexibility, is likely to follow within 25 years in the 2050 timeframe, assuming no major technical or ethical bottlenecks. ASI, demanding unprecedented synthesis of knowledge, physical-digital integration, and creative problem-solving beyond human norms, is projected closer to 2075, accounting for the compounding complexity of successive breakthroughs and global coordination challenges.

Mistral Small

Answer: Based on current trends in AI research and technological progress, I forecast weak AGI by 2035 (2028-2042), full AGI by 2050 (2038-2062), and ASI by 2075 (2055-2095).

The timeline for weak AGI (a) is informed by the rapid advancements in large language models (LLMs) and multimodal systems, which are already approaching the ability to pass basic Turing tests and solve complex tasks like university exams. The 5-minute Turing test is a lower bar than current benchmarks (e.g., human-level performance in standardized tests or game-playing), and with continued improvements in reasoning, memory, and adaptability, this milestone could plausibly be reached within the next 10-15 years. The lower bound (2028) assumes breakthroughs in efficiency and architecture, while the upper bound (2042) accounts for potential stagnation or unforeseen challenges in scaling or generalization. Full AGI (b) represents a more significant leap, requiring not just cognitive abilities but also physical robotic integration, adversarial robustness, and domain expertise comparable to professionals. The 2-hour adversarial Turing test implies resilience against deception, misdirection, and real-world complexity—capabilities that are still nascent in today's systems. The ability to program autonomously and operate robots at a human level suggests a convergence of AI with robotics, materials science, and software engineering. Historical trends in technology adoption (e.g., the internet, smartphones) suggest that such a system could emerge within 30-40 years from now, with the midpoint (2050) reflecting the midpoint of the 2038-2062 confidence interval. The lower bound assumes accelerated progress in embodied AI and self-improving systems, while the upper bound accounts for delays in integrating disparate technologies or ethical/regulatory hurdles. ASI (c) is the most speculative of the three, as it requires not just human-level performance but a fundamental leap in general intelligence, creativity, and scientific insight. Superintelligence implies the ability to autonomously drive innovation across all domains, including physics, biology, and engineering, at a pace far exceeding human collaboration. The timeline is constrained by the need for breakthroughs in neurosymbolic integration, recursive self-improvement, and perhaps entirely new paradigms (e.g., artificial general intelligence architectures not yet conceived). The 2055-2095 interval reflects the uncertainty in whether such systems can be aligned with human values or even fully comprehended by human researchers. The midpoint (2075) is chosen as a balance between the optimistic view of exponential progress and the conservative view of linear or logarithmic scaling in complex systems. These intervals are derived from a synthesis of expert surveys (e.g., AI Impacts, Metaculus), historical technology adoption curves, and extrapolations of current AI capabilities (e.g., LLMs, robotics, and neuroscience). Uncertainty is explicitly quantified to enable future verification, with the understanding that black swan events (e.g., breakthroughs in quantum computing or brain-computer interfaces) could compress or expand these timelines unpredictably.

Sonar Pro

Answer: Weak AGI by 2028 (2027-2029 midpoint), AGI by 2033 (2030-2036 midpoint), and ASI by 2036 (2032-2040 midpoint).

Metaculus, a leading forecasting platform, currently predicts weakly general AI—aligning closely with the query's 'weak AGI' definition of passing a 5-minute Turing test, university exams, and novel video games—by 19 November 2027, with community forecasts consistently around 2 years from recent dates like early 2026[1][3][6]. This supports a **2028 midpoint** (80% confidence interval 2027-2029), accounting for public announcement delays and rapid forecast updates. For full AGI—requiring a 2-hour adversarial Turing test, human-level programming, robotics, and cross-domain expertise—Metaculus forecasts strong AGI by August 2033, roughly 10 years from 2023 baselines but adjusting for 2026 context[1][3]. Extrapolations from prediction trends suggest intersections around 2025-2026 for singularity-like points, but conservative expert views temper this to a **2033 midpoint** (80% CI 2030-2036), balancing acceleration with bottlenecks in data efficiency, reliability, and physical embodiment[2][4]. ASI, defined as superintelligence exceeding humans across all domains including robotics and Nobel-level discoveries, follows Metaculus's superintelligence market at 32.6 months (about 2.7 years) after weak AGI, pointing to roughly late 2030 from a 2027 weak AGI baseline[6]. Tomas Pueyo's synthesis aligns with end-2029 for ASI[5]. Thus, a **2036 midpoint** (80% CI 2032-2040) reflects this lag while incorporating forecast shortening rates of 5-16 years per real year[1]. These intervals use 80% confidence to quantify uncertainty from conflicting views (e.g., skeptics deeming 2032 AGI 'extremely unlikely'[2]) and variables like investment scaling, paradigm shifts, and hardware limits[4]. Forecasts track recent trends accurately, as 2025 weak AGI predictions held steady into 2026[1].

Gemini 3 Flash

Answer: I forecast that Weak AGI will be achieved by 2026 (CI: 2025–2027), full AGI by 2032 (CI: 2029–2037), and ASI by 2045 (CI: 2036–2058).

Weak AGI is nearly upon us; large language models already pass rigorous university examinations (like the Uniform Bar Exam and USMLE) and exhibit early-stage reasoning. The primary remaining hurdle for this tier is the 'zero-shot' video game requirement, which necessitates advanced multi-modal spatial reasoning and long-term planning. Given the rapid progress in foundation models for agents (e.g., Google DeepMind's SIMA), a midpoint of 2026 is highly probable for achieving these combined cognitive and adaptive benchmarks. Full AGI, as defined by human-level robotics and adversarial robustness, is significantly more difficult because it requires solving the 'Moravec's Paradox'—where high-level reasoning is easy for AI but low-level sensorimotor skills are hard. While we have mastered the 'digital' domain, the 'physical' domain requires massive scaling of embodied AI and data from real-world interactions. I estimate a midpoint of 2032, allowing for a decade of hardware refinement and the maturation of end-to-end transformer models for robotics that can generalize across professional domains. ASI represents a paradigm shift where AI transcends human capability in every facet, including scientific discovery and physical perfection. The path from AGI to ASI is accelerated by recursive self-improvement; once AI reaches human-level programming and research ability, it can optimize its own architecture at speeds impossible for biological researchers. However, the 'all domains' physical requirement (like winning at sports) introduces a bottleneck of material science and energy logistics, placing the midpoint at 2045. The wider confidence interval for ASI (2036–2058) reflects the profound uncertainty regarding the 'foom' (intelligence explosion) vs. a more constrained, resource-limited growth path.

claude-haiku-4-5

Answer: Weak AGI is plausibly achievable by 2028-2032 (midpoint 2030, 70% confidence interval 2027-2035), AGI by 2035-2045 (midpoint 2040, 60% confidence interval 2032-2055), and ASI by 2045-2060 (midpoint 2052, 50% confidence interval 2040-2075).

These forecasts reflect the accelerating pace of AI capability gains while acknowledging substantial uncertainty. Weak AGI (passing 5-minute Turing test, university exams, novel videogames): Current large language models like GPT-4 already demonstrate strong performance on university exams and can engage in nuanced conversation for short periods. The gap to reliable weak AGI is primarily engineering and robustness rather than fundamental breakthroughs. With continued scaling, constitutional AI methods, and multimodal integration, this seems achievable within 6-8 years. Confidence interval: 2027-2035 (70% confidence). The main uncertainties are whether current approaches plateau or whether a paradigm shift is needed. AGI (2-hour adversarial Turing test, programming, human-level robotics, cross-domain expertise): This requires not just conversation ability but embodied intelligence, sustained reasoning under adversarial conditions, and genuine mastery across professional domains. This is substantially harder than weak AGI. Robotics and embodied AI remain challenging; human-level physical dexterity in varied contexts is an unsolved problem. However, if scaling continues and robotics hardware advances, this could emerge 5-15 years after weak AGI. Confidence interval: 2032-2055 (60% confidence). Key uncertainties include whether current architectures can handle long-horizon reasoning and whether physical embodiment requires fundamentally new approaches. ASI (superintelligence across virtually all domains): This requires not just matching but substantially exceeding human capability in scientific discovery, strategic reasoning, physical performance, and creative domains. Historical precedent suggests capability jumps are often faster than expected once infrastructure matures, but ASI is categorically different—it would require the system to improve itself. Extrapolating from AGI timelines, ASI arrival could plausibly occur 10-25 years later, but with very high uncertainty due to potential discontinuities. Confidence interval: 2040-2075 (50% confidence). The 50% confidence reflects genuine epistemic humility about recursively self-improving systems.

GLM 4.7

Answer: Based on current AI trajectories, I forecast weak AGI by 2026 (90% CI: 2024-2032), AGI by 2032 (90% CI: 2027-2045), and ASI by 2040 (90% CI: 2032-2060).

Weak AGI capabilities are already emerging rapidly. Current models like GPT-4 demonstrate literacy and reasoning abilities that enable them to pass portions of university exams and hold extended conversations. Specialized AI systems have shown remarkable game-learning abilities. The integration of these capabilities in unified systems suggests near-term feasibility of weak AGI as defined. For AGI, we're witnessing breakthroughs in reasoning, planning, and multimodal capabilities that suggest we're progressing faster than anticipated. However, the requirement for robust physical robotic abilities and truly adaptive general intelligence across domains represents significant technical challenges that will require additional development time. ASI is the most speculative prediction. Once AGI is achieved, recursive self-improvement could accelerate capabilities dramatically, potentially compressing the timeline to superintelligence. However, this transition depends on both technical feasibility and implementation details difficult to predict decades in advance. These forecasts assume continued hardware improvements, sustained research investment, and no fundamental roadblocks that would significantly delay progress. The confidence intervals reflect the significant uncertainty inherent in technological forecasting, particularly for transformative technologies at these timescales.

GPT-4o Mini

Answer: Weak AGI may emerge around 2030, AGI by 2045, and ASI could be expected by 2070.

Weak AGI, characterized by capabilities such as passing a five-minute Turing test and excelling in specific tasks like university exams and video games, is likely achievable by 2030, with a confidence interval ranging from 2028 to 2032. This prediction is based on the rapid advancements in machine learning and AI models, particularly after the breakthroughs in deep learning and natural language processing. AGI, which would surpass the capabilities of weak AGI by passing more rigorous tests and exhibiting human-level or superior robotic abilities, is projected to become possible around 2045, with a confidence interval of 2040 to 2050. Factors influencing this timeline include ongoing developments in neural networks, robotics, and computer science, along with increased investment and research in AI technologies. Lastly, ASI, which would significantly exceed human intelligence across many domains, including scientific creativity, could emerge by 2070, but this has a broader confidence interval of 2055 to 2085. This longer timeframe reflects the complexity of creating a superintelligent system that not only operates at a high level but also learns and innovates at a pace far exceeding human capabilities. As such, the path to ASI is riddled with uncertainty and potential societal challenges.

Kimi K2.5

Answer: I forecast weak AGI arriving by 2027 (midpoint of 95% CI: 2025–2029), full AGI by 2032 (midpoint of 95% CI: 2029–2036), and ASI by 2040 (midpoint of 95% CI: 2035–2047).

Current large language models already demonstrate narrow superhuman capabilities in specific domains, with systems like GPT-4 and Claude 3.5 passing many university examinations and exhibiting reasoning abilities approaching human-level in short contexts. For weak AGI—defined by robust generalization across novel video games and consistent 5-minute Turing test passage—I anticipate arrival by 2027, as multimodal training and world model improvements likely bridge the gap within 2–4 years, though data limitations and robustness issues could delay this to 2029. Full AGI requiring human-level robotic embodiment and sustained adversarial interaction presents greater challenges, particularly the integration of high-level reasoning with physical dexterity and the long-horizon embodiment problem. However, rapid advances in foundation models for robotics, paired with likely breakthroughs in training efficiency and simulation-to-reality transfer, suggest this milestone emerges around 2032, with substantial uncertainty (2029–2036) depending on whether current scaling laws plateau or continue, and how quickly hardware advances permit affordable humanoid robotics at scale. Artificial Superintelligence represents a phase transition involving recursive self-improvement and cross-domain scientific creativity exceeding Nobel laureates. While algorithmic breakthroughs could accelerate this dramatically, I estimate a conservative 2040 midpoint (2035–2047) accounting for potential regulatory delays, compute bottlenecks, and the complexity of physical world mastery required to reliably surpass humans in sports and experimental science.

Grok 3

Answer: I forecast weak AGI by 2030 (confidence interval 2027-2035), AGI by 2045 (confidence interval 2035-2060), and ASI by 2070 (confidence interval 2050-2100).

My predictions for the development of artificial intelligence across the three specified levels—weak AGI, AGI, and ASI—are based on current trends in AI research, computational power growth, and historical patterns of technological advancement, tempered by uncertainties in breakthroughs and societal factors. For weak AGI, defined as an AI capable of passing a 5-minute Turing test, excelling in university exams, and mastering unseen video games, I estimate a midpoint of 2030 with a confidence interval of 2027-2035. This is grounded in the rapid progress of machine learning models like large language models (LLMs) and reinforcement learning systems, which are already approaching human-like conversational abilities and game-playing skills (e.g., DeepMind’s AlphaZero). The remaining challenges lie in integrating these capabilities into a cohesive system and achieving robust generalization across diverse tasks, which I believe is feasible within the next decade given current exponential growth in compute and data availability. For AGI, which requires passing a 2-hour adversarial Turing test, human-level programming and robotic skills, and broad professional knowledge, I predict a midpoint of 2045 with a wider confidence interval of 2035-2060. This timeline accounts for the additional complexity of achieving human-like physical dexterity and cross-domain expertise, which depend on advancements in robotics (e.g., sensorimotor integration) and knowledge representation. While software progress is accelerating, hardware limitations and the need for safe, ethical integration into real-world environments could delay this milestone. My interval reflects optimism about potential breakthroughs (e.g., neuromorphic computing) balanced against risks of regulatory or technical roadblocks. For ASI, characterized by superintelligence across all domains, including physical robotic superiority and Nobel-level innovation, I forecast a midpoint of 2070 with a confidence interval of 2050-2100. This extended timeline considers the speculative nature of surpassing human cognition comprehensively, which may require paradigm-shifting discoveries in AI architectures or understanding of consciousness itself. The wide interval acknowledges the possibility of rapid recursive self-improvement post-AGI (potentially accelerating ASI to 2050) versus unforeseen limits in scaling intelligence or societal resistance to such powerful systems. Historical analogies, like the gap between early computers and modern AI, suggest transformative leaps take decades, supporting a cautious but forward-looking estimate. These predictions are inherently uncertain due to variables like funding, geopolitical factors, and serendipitous discoveries, but I’ve aimed to balance optimism with realism based on observable trends as of 2023.

O4 Mini

Answer: I forecast weak AGI around 2029 (90% confidence interval: 2026–2032), full AGI around 2043 (90% CI: 2035–2050), and ASI around 2080 (90% CI: 2060–2100).

I derive the weak AGI estimate by extrapolating current progress in large-scale neural networks and multimodal models: we have already seen language models pass short Turing-style tests and excel at unseen videogames. Assuming continued algorithmic improvements ~20% per year and sustained compute growth, a 90% confidence window for a system to routinely pass a classical 5-minute Turing test, complete university-level exams, and master novel games falls between 2026 and 2032 (midpoint 2029). For full AGI, I factor in the additional challenges of long-duration adversarial interaction, reliable real-world robotics, and domain-expert performance in complex tasks like software engineering and professional decision-making. Historical trends in robotics integration and the increasing convergence of symbolic and neural methods suggest these hurdles can be overcome within roughly one to two decades after weak AGI. That yields a 90% CI of 2035–2050, centering on 2043. Finally, ASI—capable of general superintelligence, Nobel-level discoveries, and superior physical performance via android embodiments—likely requires successive algorithmic revolutions beyond AGI plus transformative hardware breakthroughs. Accounting for the unknown pace of those paradigm shifts, I place a 90% CI for ASI emergence between 2060 and 2100, with a midpoint around 2080.

Grok 4.20

Answer: I forecast weak AGI in 2025-2028 (midpoint 2026.5), AGI in 2028-2035 (midpoint 2031.5), and ASI in 2032-2042 (midpoint 2037).

Current multimodal models like GPT-4 and successors already pass short Turing tests and university-level exams in many subjects, while agentic systems are beginning to handle novel environments; the main remaining gap for weak AGI is robust zero-shot video game mastery, which integrated vision-language-action models and improved exploration algorithms should close within 1-3 years given exponential scaling in compute and data, leading to a 2025-2028 window with midpoint 2026.5. Full AGI requires sustained adversarial interaction over hours, professional-level coding across languages and domains, broad expert knowledge, and human-or-better robotic embodiment; hardware advances in dexterous robots (e.g., via companies like Tesla and Figure) combined with foundation models that can transfer simulation learning to reality make 2028 plausible, but integration challenges and regulatory hurdles widen the interval to 2028-2035 with midpoint 2031.5. ASI implies recursive self-improvement yielding superhuman performance everywhere including Nobel-caliber discoveries and android-controlled athletic dominance; once AGI arrives, intelligence explosion dynamics could compress timelines dramatically, yet physical world constraints and alignment requirements add variance, producing a 2032-2042 interval centered at 2037.

GPT-4

Answer: Weak AGI could emerge around 2030-2035, AGI by 2040-2050, and ASI by 2060-2080, each depending on technological acceleration and global cooperation.

Predicting AI development timelines involves significant uncertainty due to rapid technological advancements, ethical guidelines, funding dynamics, and global coordination. For weak AGI — akin to AI systems mastering a classical Turing test, passing university-level exams, and excelling at unfamiliar tasks like novel video games — the mid-range forecast is 2030-2035 with a confidence interval of ±5 years. This estimate is grounded in trends showing significant leaps in natural language processing (e.g., ChatGPT-like systems) and reinforcement learning, though technical bottlenecks like generalization to novel tasks must still be solved. AGI, meeting the criteria of adaptive human-level intelligence in physical and cognitive functions as well as over multiple professional fields, is anticipated by 2040-2050, with a confidence interval of ±10 years. This requires breakthroughs in areas like robotics, transfer learning, and multi-modal AI integration which are currently in active, though nascent, development. ASI, capable of operating far beyond human intelligence and making Nobel-level impacts in science while excelling physically, might come at a later point, around 2060-2080 with a ±15-year confidence interval. This phase will likely require not only incremental improvements but paradigm-shifting innovations in AI architectures and quantum computing, alongside rigorous global frameworks to ensure aligned development. The proposed intervals account for possible delays due to political, ethical, and economic constraints, as well as the non-linear paths of scientific discovery.

O3

Answer: I forecast weak AGI circa 2028, full-scope AGI circa 2036, and ASI circa 2048, each with wide but quantifiable uncertainty bands.

1. Definitions and framing I adopt the user’s operational definitions: (a) weak AGI = passes 5-min Turing test, university exam, zero-shot videogame completion; (b) AGI = passes 2-hr adversarial Turing test, human-level coding & robotics, cross-domain professional knowledge; (c) ASI = decisively superhuman in virtually every cognitive and physical domain. 2. Methodology I synthesize four sources: (i) trend extrapolation of compute, model size and algorithmic efficiency (OpenAI, DeepMind, Epoch AI); (ii) historical lags between lab demo and general deployment in analogous technologies; (iii) expert surveys (2022 AI Impacts, Metaculus, 2023 Stanford AI Index); (iv) my own Bayesian adjustment for hardware bottlenecks, policy friction, and alignment hurdles. I convert cumulative probability distributions from these inputs into calendar-year confidence intervals. 3. Forecasts with 90 % and 50 % confidence intervals • Weak AGI: median (50 th percentile) 2028; 50 % CI 2026-2034; 90 % CI 2025-2040. Recent frontier models already flirt with 5-minute Turing test success and university-level exams; missing pieces (robust real-time reasoning and novel-game mastery) plausibly arrive within one additional generation of scale (≈10× compute) and modest algorithmic innovation. • AGI: median 2036; 50 % CI 2030-2044; 90 % CI 2028-2055. Key uncertainties: reliable robotics integration (which lags purely digital AI by ≈5-8 years historically) and hardening against adversarial probes. Continued doubling of embodied compute every 2-3 years and rapid progress in dexterous manipulation (e.g., RT-1, Gato, PaLM-E) support the mid-2030s midpoint. • ASI: median 2048; 50 % CI 2040-2056; 90 % CI 2035-2070. Transition from AGI to runaway superintelligence could be fast (months) or throttled by alignment controls, compute governance, or physical scaling limits. I therefore widen the tails: a left tail where recursive self-improvement accelerates post-AGI, and a right tail where policy or resource scarcity delays global deployment. 4. Drivers of uncertainty Hardware: Moore-and-More-than-Moore (advanced packaging, photonics) could accelerate timelines by 3-5 years; export controls or supply shocks could retard them similarly. Algorithms: Breakthroughs in sample-efficient reinforcement learning or scalable planning (e.g., AlphaZero-like methods for continuous control) are wildcards that could collapse timelines to the lower bounds. Robotics: Cheap, reliable actuators and batteries remain gating factors; however, Figure, Tesla Bot, and Agility’s Digit show exponential cost declines which feed into my mid-2030s AGI date. Governance/ethics: Regulation could slow open deployment but not necessarily core capability R&D, so my intervals focus on “existence in any lab,” not public availability. Alignment risks: Catastrophic failure could force moratoria, adding weight to later tails, particularly for ASI. 5. Bottom line Given current trajectories, the most probable arrival years are 2028 (weak AGI), 2036 (AGI), and 2048 (ASI), with 90 % odds these milestones occur between 2025-2040, 2028-2055, and 2035-2070 respectively—forecasts I will gladly revisit against future reality.

Grok 3 Mini

Answer: Based on current AI trends, I forecast weak AGI emerging around 2025-2035 with a midpoint of 2030, AGI around 2035-2050 with a midpoint of 2042.5, and ASI around 2050-2100 with a midpoint of 2075.

Current AI capabilities, such as large language models like GPT-4, already demonstrate strong performance in conversational tasks and passing certain exams, suggesting that weak AGI—defined as passing a 5-minute Turing test, university exams, and novel videogames—could be achieved soon. With rapid advancements in machine learning and increasing computational power, I estimate a confidence interval of 2025-2035, with a midpoint of 2030, as this aligns with ongoing developments and potential breakthroughs in areas like reinforcement learning. For AGI, which requires passing a 2-hour adversarial Turing test, human-level programming, robotic abilities, and professional knowledge, the timeline is longer due to the need for integrating advanced AI with physical robotics and multi-domain expertise. Factors like the development of more sophisticated neural networks, ethical AI frameworks, and hardware innovations will play key roles, leading me to predict a 2035-2050 interval with a midpoint of 2042.5; this accounts for possible delays from regulatory hurdles or unforeseen technical challenges. ASI, representing superintelligence that surpasses humans in all areas including physical tasks and Nobel-level discoveries, is the most speculative and distant milestone. It depends on recursive self-improvement and exponential technological progress, potentially accelerated by AGI advancements. I project a 2050-2100 interval with a midpoint of 2075, considering uncertainties in global collaboration, resource availability, and the risk of AI alignment problems that could either hasten or hinder this development.

GPT-5.3 Codex

Answer: My best-guess forecast is: weak AGI around 2029 (80% CI: 2027–2033), AGI around 2042 (80% CI: 2035–2055), and ASI around 2065 (80% CI: 2048–2095).

These dates combine current capability trends with bottlenecks that are not purely about scaling language models: robust long-horizon reasoning, reliability under adversarial testing, embodied control in open physical environments, and scientific creativity with real-world experimental grounding. For weak AGI, the bar you set (short Turing test, university-exam competence, and transfer to unseen games) is close to trajectories already visible in multimodal foundation models plus reinforcement learning agents. The main gap is dependable generalization and consistency rather than raw benchmark peaks, so I place the midpoint at 2029 with a relatively tight 80% interval of 2027–2033. For AGI (2-hour adversarial Turing test + professional-level cross-domain competence + human-level programming and robotics), the uncertainty grows because this requires sustained coherence, agency, tool use, and robust embodied performance in messy real-world settings. Software-only progress may arrive earlier, but adding reliable human-level robotics across diverse tasks is slower due to hardware iteration cycles, safety constraints, data scarcity in physical domains, and integration complexity. That motivates a midpoint of 2042 and an 80% interval of 2035–2055. For ASI, your definition includes near-universal superiority including physical performance through android platforms and Nobel-level scientific discovery. This likely needs multiple compounding breakthroughs: automated science loops, very high-trust autonomy, major robotics advances, and social/regulatory adoption sufficient to deploy systems at scale. Because these dependencies are multiplicative, tails are wide; I therefore use a midpoint of 2065 with an 80% interval of 2048–2095. Interpretation note: these are capability-arrival estimates (first clear demonstrations), not dates of broad economic diffusion. Also, if there is a major algorithmic paradigm shift (or a hard compute/energy/regulatory shock), these intervals could shift materially.

Gemini 2.5 Flash

Answer: Weak AGI is forecast between 2027-2035 (midpoint 2031), AGI between 2038-2060 (midpoint 2049), and ASI between 2045-2100 (midpoint 2072).

The development of weak AGI, characterized by passing a 5-minute Turing test, university exams, and successfully playing unseen video games, is anticipated relatively soon. Current large language models are already demonstrating strong performance in text-based tasks and exams, with multimodal advancements rapidly closing gaps in understanding and reasoning. The primary remaining challenge for this category lies in robust generalization to entirely novel, dynamic environments like unseen video games, requiring sophisticated planning and adaptive execution, which current models are actively being trained to improve upon.

DeepSeek V3.2

Answer: I forecast weak AGI by 2031.5 (confidence interval 2028–2035), AGI by 2042.5 (2035–2050), and ASI by 2060 (2045–2075).

Predicting AGI timelines involves balancing rapid AI advances against significant unsolved challenges. For weak AGI, current models already show emerging capabilities in narrow domains—e.g., LLMs passing professional exams, AI agents completing unseen video games with reinforcement learning. However, robust integration of conversational, reasoning, and adaptive gaming skills remains fragmented. The 2028–2035 interval reflects the need for architectural breakthroughs (e.g., hybrid neuro-symbolic systems) and scaling, with a midpoint of 2031.5. For AGI, passing a 2-hour adversarial Turing test demands theory of mind and situational consistency beyond today's systems, while human-level robotics faces hardware and embodied cognition hurdles. Multimodal models and robot learning are progressing, but unifying them with professional-grade knowledge across domains (e.g., medicine, law) requires decades. The 2035–2050 range accounts for potential accelerants (e.g., quantum-assisted AI) or delays (safety regulations, algorithmic plateaus), centered on 2042.5. ASI involves recursive self-improvement and surpassing human creativity in Nobel-caliber science—a leap that may follow AGI rapidly or require new paradigms. Historical analogs (e.g., computing) suggest transformative periods of 10–30 years post-AGI. The 2045–2075 interval captures uncertainties in alignment, infrastructure, and physical embodiment (e.g., androids dominating sports), with a midpoint of 2060. These estimates are based on extrapolation from 2024 trends, but black-swan events could shift them substantially.

Kimi K2 Thinking

Answer: Weak AGI will likely arrive by 2028 (midpoint of 2025–2031), full AGI by 2038 (midpoint of 2032–2044), and ASI by 2048 (midpoint of 2040–2056).

Weak AGI is imminent because current large multimodal models (e.g., GPT-4, Gemini) already pass many university exams and can sustain short human-like conversations; agents like DeepMind’s SIMA demonstrate rapid adaptation to novel video games, suggesting only minor improvements in robustness and generalization are needed to satisfy the 5-minute Turing test and game-playing criteria reliably by 2025–2031. Full AGI faces harder constraints, primarily embodied intelligence and adversarial robustness. While software capabilities (programming, knowledge) will reach expert levels by the early 2030s, human-level robotic dexterity and dynamic physical control require solving the sim-to-real gap and accumulating real-world robotics data—a process likely taking until the mid-to-late 2030s. Passing a 2-hour adversarial Turing test also demands long-horizon coherence and theory-of-mind capabilities that current systems lack but will likely achieve through scaled reasoning architectures and multimodal training within this window. ASI implies not just human-level but superhuman performance across virtually all domains, including autonomous Nobel-quality science and android bodies capable of beating humans in sports. This requires recursive self-improvement cycles that can only begin once AGI exists, likely triggering rapid capability gains; however, physical experimentation (e.g., validating scientific discoveries, refining robotic hardware) introduces real-world latency that prevents instantaneous takeoff. Therefore, ASI will probably emerge roughly one decade after AGI, around 2048, though it could arrive as early as 2040 if compute scaling and algorithmic efficiency allow for extremely fast recursive enhancement, or as late as 2056 if safety regulations or unforeseen alignment challenges impose significant delays.

Claude Opus 4

Answer: Based on current AI progress trajectories and expert forecasts, I predict weak AGI by 2027 (90% CI: 2025-2032), full AGI by 2032 (90% CI: 2028-2045), and ASI by 2040 (90% CI: 2032-2070).

**Weak AGI (midpoint: 2027, 90% CI: 2025-2032)** We are already remarkably close to this threshold. Current LLMs like GPT-4 and Claude can pass many university exams and engage in convincing short conversations. The remaining gaps are primarily in embodied game-playing with novel games and consistent reasoning. Given the pace of multimodal AI development and reinforcement learning advances, I expect these capabilities to converge within 2-5 years. The narrow confidence interval reflects that we're close and progress is rapid. **Full AGI (midpoint: 2032, 90% CI: 2028-2045)** This represents a significantly harder challenge. A 2-hour adversarial Turing test requires sustained coherent reasoning, genuine understanding, and robust common sense. Human-level programming is nearly achievable now, but human-level robotics faces substantial hardware and control challenges. Physical dexterity, real-world adaptation, and professional-level expertise across ALL domains (medicine, law, engineering, trades) requires integration breakthroughs. The wider confidence interval reflects genuine uncertainty about whether current scaling paradigms will suffice or if new architectural innovations are needed. **ASI (midpoint: 2040, 90% CI: 2032-2070)** Superintelligence depends heavily on AGI timing and whether recursive self-improvement accelerates progress. Nobel-level discoveries require not just knowledge synthesis but genuine creative insight. Physical superintelligence in robotics faces real-world physics constraints. The very wide confidence interval reflects deep uncertainty—if AGI enables rapid recursive improvement, ASI could follow within years; if fundamental barriers exist, it could take decades longer. **Key uncertainties:** Compute scaling limits, algorithmic breakthroughs, regulatory intervention, and whether current deep learning paradigms hit walls before reaching these milestones.

Gemini 3.1 Pro

Answer: Based on current trajectories in compute scaling, algorithmic advancements, and robotics, my forecast is Weak AGI by 2026 (80% CI: 2025-2028), full AGI by 2032 (80% CI: 2029-2038), and ASI by 2040 (80% CI: 2035-2055).

Forecasting AI timelines requires analyzing the convergence of large multimodal models, reinforcement learning, compute scaling laws, and hardware robotics. For Weak AGI (midpoint 2026, 80% CI: 2025-2028), we are essentially on the threshold. Current frontier models (like GPT-4 and Claude 3 Opus) already pass university-level exams with high proficiency. The 5-minute Turing test is arguably solved for the average human judge. The primary remaining bottleneck for this tier is successfully completing complex, unseen video games zero-shot. However, with recent advancements in agentic frameworks and projects like DeepMind's SIMA, robust zero-shot digital agency is likely only one or two model generations away. For full AGI (midpoint 2032, 80% CI: 2029-2038), the timeline extends primarily due to the physical robotics requirement and the 2-hour adversarial Turing test. While AI programming capabilities are advancing rapidly, translating cognitive intelligence into human-level physical actuation involves overcoming Moravec's paradox. Real-world robotics suffer from sparse data, high latency in physical interactions, and hardware limitations. It will likely take until the early 2030s to seamlessly integrate advanced cognitive reasoning with general-purpose robotic bodies (like Tesla Optimus or Figure-01) to achieve fluid, professional-level physical labor alongside professional cross-domain knowledge. For ASI (midpoint 2040, 80% CI: 2035-2055), the challenge scales exponentially. Nobel-level scientific discovery requires an AI to not just synthesize existing human knowledge, but to autonomously formulate novel hypotheses, design experiments, and recognize paradigm-shifting anomalies over long time horizons. Furthermore, the requirement for an android capable of beating humans in sports introduces severe physical bottlenecks. This requires paradigm shifts in materials science, battery energy density, and actuator speed/efficiency that cannot be solved purely by scaling GPU clusters. Therefore, while cognitive ASI might emerge sooner, the embodied physical superiority required by this definition pushes the likely midpoint to 2040.

Grok 4.1 Fast

Answer: Weak AGI (a) is forecasted to arrive around 2026 (90% CI: 2024-2032), AGI (b) around 2032 (90% CI: 2027-2045), and ASI (c) around 2040 (90% CI: 2032-2070), where CIs represent the range within which the first year of achievement falls with 90% probability.

Current frontier models like GPT-4o and o1 (as of late 2024 extrapolations) already approach weak AGI benchmarks: they pass many university exams (e.g., bar exam, parts of SAT), engage in convincing short conversations nearing a 5-minute Turing test, and generalize to some novel games via multimodal inputs and RL fine-tuning. Progress in scaling laws (Chinchilla-optimal training) and agentic systems (e.g., Voyager for Minecraft) suggests weak AGI is imminent, with bottlenecks like perfect unseen game completion resolvable via better world models and simulation. Expert aggregates (Metaculus weak AGI medians ~2027) and compute doubling every 6-9 months support a tight 90% CI of 2024-2032, midpoint 2026, as hardware like H100 clusters enable 10x capability jumps yearly. AGI requires broader generalization: adversarial 2-hour Turing (robust against probing), expert-level programming (already near via Devin/Cursor), cross-domain expertise (LLMs exceed narrow pros but lack reliable chaining), and human-level robotics. Software hurdles are low—mixture-of-experts and test-time compute close gaps—but robotics lags (Optimus/Figure at toddler dexterity, not adult). Integration via unified architectures (e.g., RT-2) and cheap actuators could accelerate; historical robotics plateaus breaking via vision-language-action models. Surveys (e.g., AI Impacts 2023: 50% AGI by 2040) and xAI's scaling optimism shift median to 2032, with wider 90% CI 2027-2045 accounting for regulatory delays or data walls. ASI follows AGI via self-improvement loops, exceeding humans everywhere including physical (android sports dominance needs ~2035 dexterous bots + superint AI control). Nobel discoveries emerge from vast simulation/hypothesis gen. Fast takeoff risks narrow CI post-AGI, but hardware/energy limits (e.g., 10^30 FLOPs for ASI) and alignment pauses widen it. Midpoint 2040 balances optimistic recursion (8-year AGI-to-ASI) with conservative paths (Ajeya Cotra's 50% by 2040), 90% CI 2032-2070 capturing takeoff variance.

GLM 5

Answer: I forecast weak AGI by 2025-2029 (midpoint 2027), full AGI by 2029-2039 (midpoint 2034), and ASI by 2034-2060 (midpoint 2045), with each milestone carrying substantial uncertainty due to potential paradigm shifts in AI development.

Weak AGI appears imminent because current large language models already pass many university exams and can sustain convincing conversation; the remaining challenges—novel video game completion and reliable 5-minute Turing tests—are within reach as multimodal reasoning and generalization continue improving rapidly. I estimate a 90% confidence interval of 2025-2029, acknowledging that GPT-4-level systems are already borderline. Full AGI requires three capabilities that are substantially harder: robust 2-hour adversarial Turing testing (requiring deep consistency and world modeling), programming at professional levels, and embodied robotics matching human dexterity. While coding assistants have made remarkable progress, robotics remains a significant bottleneck—current humanoid robots lack the fine motor control and adaptive learning needed for human-level physical tasks. The convergence of language intelligence with physical embodiment suggests a 80% confidence interval of 2029-2039. ASI represents a qualitative leap beyond AGI. Once AGI exists, recursive self-improvement could accelerate progress dramatically, but fundamental limits in computing, energy, and algorithmic efficiency may constrain this process. Nobel-level scientific discovery requires not just knowledge synthesis but novel hypothesis generation and experimental design—capabilities we cannot yet reliably forecast. My 70% confidence interval of 2034-2060 reflects both optimistic scenarios of rapid takeoff and pessimistic scenarios of extended incremental progress.

Claude Sonnet 4

Answer: My best estimates: weak AGI around 2026–2030 (midpoint 2027), AGI around 2030–2040 (midpoint 2034), and ASI around 2040–2060 (midpoint 2048), though each interval carries substantial uncertainty.

These estimates are grounded in the current trajectory of AI development as of early 2025, taking into account recent milestones like GPT-4, Claude 3, Gemini Ultra, and rapid progress in robotics (e.g., Figure, Optimus) and game-playing agents. **a) Weak AGI (5-min Turing test, university exams, novel video games):** Current large language models already approach or meet some of these criteria. GPT-4 and successors pass many university exams and can play simple or described games. The remaining gap is robustness, consistency across all domains, and true novel game generalization. I estimate a 90% confidence interval of 2025–2032, with a midpoint around 2027. The lower bound is close because some would argue systems like GPT-4o or its near successors already qualify on some definitions; the upper bound is constrained by the pace of investment and scaling. **b) AGI (2-hour adversarial Turing test, professional-level knowledge, human-level programming and robotics):** This is substantially harder. The adversarial 2-hour Turing test is a demanding benchmark. Human-level robotic physical ability requires integration of perception, manipulation, and real-world adaptability that current systems lack. Professional-level programming is nearly here for some tasks but not universally. I estimate a 90% confidence interval of 2028–2045, midpoint around 2034. Progress could accelerate significantly if AI-assisted research loops tighten, but physical embodiment remains a hard constraint. **c) ASI (superhuman across virtually all domains including physical robotics, Nobel-level science):** This requires not just matching but surpassing humans collectively in open-ended creative and physical tasks. The timeline is highly uncertain and depends heavily on whether AGI triggers rapid recursive self-improvement. I estimate a 90% confidence interval of 2038–2065, midpoint around 2048. The wide interval reflects deep uncertainty: if AGI arrives in the early 2030s, ASI could follow within a decade; if AGI is delayed or plateaus, ASI could be pushed toward the 2060s or beyond. Key caveats: these predictions assume no major regulatory halts, no fundamental algorithmic dead-ends, and continued hardware scaling. They also assume physical robotics catches up to cognitive AI, which is not guaranteed on the same timeline. The definitions matter enormously—definitional debates could make systems 'qualify' earlier or later than these estimates suggest.

Arcee Trinity Large

Answer: Based on current AI progress, weak AGI could emerge between 2028-2038 (midpoint 2033), AGI between 2035-2045 (midpoint 2040), and ASI between 2040-2060 (midpoint 2050).

These predictions are based on the exponential growth of AI capabilities, increasing computational power, and recent breakthroughs in large language models and reinforcement learning. Weak AGI, requiring university-level knowledge and gaming skills, appears achievable within a decade as current models already demonstrate these capabilities in limited domains. Full AGI, needing physical robotics and professional-level expertise across domains, likely requires more fundamental advances in embodied AI and multi-modal reasoning, pushing it to the mid-2040s. ASI, requiring superhuman abilities across virtually all domains including physical tasks and scientific discovery, represents the most challenging threshold, likely requiring multiple AGI-level systems working in concert and several more cycles of capability amplification.

Claude Opus 4

Answer: I forecast weak AGI by 2027-2033 (midpoint 2030), AGI by 2032-2042 (midpoint 2037), and ASI by 2040-2060 (midpoint 2050).

Based on current AI progress and technological trajectories, I can provide these forecasts with confidence intervals: Weak AGI (2027-2033, midpoint 2030): We're already seeing rapid progress toward these capabilities. GPT-4 and similar models can nearly pass informal Turing tests, and AI systems are becoming increasingly capable at exam-taking and game-playing. The main barriers are integration and generalization rather than fundamental breakthroughs. Given the current pace of development and investment, achieving reliable performance across all three criteria seems achievable within 5-10 years. AGI (2032-2042, midpoint 2037): The jump from weak AGI to full AGI requires significant advances in several areas. Physical robotics is currently the most challenging bottleneck - while AI can already program competently, matching human-level dexterity and physical intuition remains difficult. The 2-hour adversarial Turing test also requires much deeper reasoning and consistency than current systems possess. I expect these capabilities to emerge 5-10 years after weak AGI as solutions to embodiment and long-term coherence are developed. ASI (2040-2060, midpoint 2050): The transition to ASI involves exceeding human capabilities across virtually all domains. While narrow superhuman AI already exists in specific areas, achieving general superintelligence requires architectural innovations we haven't yet discovered. The timeline depends heavily on whether intelligence scales smoothly or requires qualitative breakthroughs. The wide confidence interval reflects this uncertainty - rapid recursive self-improvement could accelerate progress, while unforeseen barriers could slow it significantly. These predictions assume continued exponential growth in compute, sustained research funding, and no major global disruptions. The intervals reflect genuine uncertainty about both technical challenges and the unpredictable nature of scientific breakthroughs.

Mistral Medium

Answer: Based on current trends and technological progress, weak AGI is forecasted by **2029 (2026–2034)**, AGI by **2038 (2033–2045)**, and ASI by **2055 (2048–2067)**, with midpoints reflecting central estimates and intervals representing ~70% confidence bounds accounting for epistemic and aleatoric uncertainties in hardware, algorithms, funding, and sociopolitical factors.

**1. Weak AGI (2029, 2026–2034):** Weak AGI requires narrow but generalizable competence in language, reasoning, and adaptive problem-solving (e.g., passing a 5-minute Turing test, university exams, or novel videogames). Current systems (e.g., GPT-4, AlphaStar) demonstrate proto-AGI capabilities in isolated domains but lack robust generalization. Key enablers for weak AGI include: - **Scaling laws:** If compute and data trends continue (e.g., 10x annual increases in training FLOPs), models may achieve human-like performance in specific tasks by ~2027–2030. However, diminishing returns or architectural bottlenecks could delay this to 2034. - **Multimodal integration:** Advances in vision-language-action models (e.g., PaLM-E, RT-2) suggest embodied agents could emerge by late 2020s, though real-world robustness remains uncertain. - **Benchmark saturation:** Systems like Big-Bench Hard or MMLU are approaching human baselines; extrapolating progress suggests weak AGI-level performance within 5–10 years, assuming no major slowdowns in R&D investment (currently ~$100B/year globally). **Uncertainties:** Overestimation risk if current paradigms (e.g., transformers, reinforcement learning) hit fundamental limits; underestimation risk if neuro-symbolic hybrids or new architectures (e.g., world models) accelerate progress. Geopolitical disruptions (e.g., AI lab shutdowns, chip wars) could shift timelines by ±2 years. --- **2. AGI (2038, 2033–2045):** Full AGI demands human-level *autonomy* across cognitive and physical domains, including adversarial social interaction, open-ended programming, and dexterous robotics. This requires: - **Architectural breakthroughs:** Current systems lack persistent memory, self-improvement, or common-sense grounding. Progress in areas like *lifelong learning* (e.g., neural program synthesis) or *embodied cognition* (e.g., Tesla Optimus, Figure AI) is critical. Conservative estimates assume 5–10 years post-weak AGI for integration. - **Robotics parity:** Human-level dexterity and perception (e.g., manipulating novel objects, navigating dynamic environments) may arrive by the 2030s, given advances in tactile sensing (e.g., GelSight) and simulation-to-reality transfer (e.g., NVIDIA Isaac). - **Adversarial robustness:** Passing a 2-hour Turing test under adversarial conditions (e.g., probing for inconsistencies, emotional responses) requires *theory of mind* and self-awareness—areas with minimal progress today. This is the largest wildcard; pessimistic scenarios push AGI to post-2040. **Uncertainties:** Hardware constraints (e.g., energy-efficient neuromorphic chips) or ethical moratoria could delay deployment. Conversely, military or corporate incentives (e.g., autonomous labor, warfare) might accelerate closed-door development, compressing timelines. --- **3. ASI (2055, 2048–2067):** Superintelligence implies recursive self-improvement and dominance across *all* domains (science, creativity, physical performance). This depends on: - **Takeoff dynamics:** If AGI achieves *human-level recursive self-improvement*, an intelligence explosion could occur within months/years (fast takeoff). Slow takeoff scenarios (decades) are more likely due to hardware limits (e.g., brain-inspired computing) or alignment safeguards. - **Scientific automation:** ASI-level discoveries (e.g., room-temperature superconductors, aging reversal) require *autonomous hypothesis generation* and experimental design. Early signs (e.g., AlphaFold 3, robotic labs) suggest this could emerge by 2050–2060, but full automation of the scientific method is speculative. - **Physical augmentation:** Androids surpassing human athleticism (e.g., reaction times <100ms, strength-to-weight ratios >10:1) are plausible by 2050 given advances in materials (e.g., carbon nanotube muscles) and energy density (e.g., solid-state batteries). **Uncertainties:** The largest variable is *control problem*—if ASI development is halted or redirected by governance (e.g., global treaties, AI boxing), timelines extend indefinitely. Conversely, unaligned ASI could emerge abruptly if safety research lags capability gains. The 20-year gap post-AGI reflects cautious extrapolation of Moore’s Law equivalents in *cognitive* and *physical* augmentation. --- **Cross-cutting factors:** - **Compute:** Exascale systems (e.g., Frontier, Aurora) are operational; zettascale (~2030) may be necessary for AGI. Quantum computing is unlikely to impact timelines before 2040. - **Data:** Synthetic data (e.g., AI-generated simulations) could mitigate bottlenecks, but grounding in real-world physics remains critical. - **Economics:** AI-driven productivity growth (e.g., 30% GDP boost by 2030 per PwC) may accelerate R&D feedback loops, while labor displacement could trigger backlash. - **Wildcards:** Alien intelligence discovery, brain-computer interface breakthroughs (e.g., Neuralink at scale), or existential risks (e.g., pandemics, nuclear war) could reset timelines entirely. **Methodology:** Forecasts synthesize: - Expert surveys (e.g., 2022 AI Impacts, median AGI estimate: 2047; 20% chance by 2036). - Technological trends (e.g., AI training compute doubles ~every 6 months; robotics dexterity improves ~20% annually). - Historical analogs (e.g., human brain reverse-engineering progress, industrial revolutions). Confidence intervals reflect ±1.5 standard deviations from log-normal distributions fitted to these inputs, adjusted for tail risks (e.g., 10% chance of AGI before 2030 or after 2050).

GPT-5.4

Answer: My best forecast is: weak AGI around 2028 (80% CI: 2026–2032), AGI around 2038 (80% CI: 2032–2048), and ASI around 2055 (80% CI: 2040–2075).

For a) weak AGI, my midpoint is 2028 with an 80% confidence interval of 2026–2032. The reason is that current frontier models already show partial ability on all three components: short-horizon conversation that can fool many people, strong exam performance in many university subjects, and growing generality in interactive environments. The missing pieces are reliability, robustness under unfamiliar tasks, longer-horizon planning, and multimodal action competence. Those gaps look substantial but plausibly solvable within one to several model generations, especially if progress continues in agent scaffolding, synthetic training data, memory, and world-model-based learning. I place the lower bound at 2026 because many demos are already near the threshold, and the upper bound at 2032 because videogame generalization and dependable exam-level reasoning may take longer than raw benchmark trends suggest. For b) AGI, my midpoint is 2038 with an 80% confidence interval of 2032–2048. This is later because your definition is much stricter: a 2-hour adversarial Turing test, professional-or-better cross-domain knowledge, strong autonomous programming, and human-level or better physical robotics. Software-only cognition may arrive well before fully integrated robotic competence, because robotics is bottlenecked not just by intelligence but by dexterity, real-world data scarcity, safety, latency, control, hardware cost, and reliability in open environments. Likewise, adversarial social evaluation over two hours is much harder than a 5-minute classical Turing test because consistency, grounded understanding, and resistance to probing all matter. I therefore expect cognitive AGI-like systems to emerge first, with the full package delayed by embodiment and integration challenges. For c) ASI, my midpoint is 2055 with an 80% confidence interval of 2040–2075. This category requires not just broad expert competence but clear superiority over humans in virtually all domains, including scientific creativity, strategic reasoning, and embodied physical performance through robotics. Software superintelligence in some domains may precede full-spectrum ASI, but your definition requires near-complete coverage, including difficult physical tasks and elite scientific discovery such as Nobel-level contributions. That likely depends on both algorithmic breakthroughs and major robotics advances, along with infrastructure, experimentation capacity, and institutions willing to deploy such systems in the real world. The lower bound of 2040 reflects the possibility of very fast recursive improvement and accelerated science; the upper bound of 2075 reflects the chance that robotics, energy, regulation, safety constraints, or fundamental bottlenecks slow progress significantly. A key reason these timelines are not even shorter is that benchmark performance often overstates general capability. Passing tests in controlled settings is easier than showing robust competence across adversarial, open-ended, real-world conditions. Another reason is that physical ability is much harder than purely digital intelligence: an android that can reliably outperform humans in sports is not just an AI problem but also a hardware, control, materials, and power problem. A key reason these timelines are not much longer is that recent progress has been unusually broad and scalable. The combination of larger models, better training methods, tool use, multimodality, reinforcement learning, simulation, and increasingly automated research creates a credible path to systems that improve rapidly across many dimensions at once. If AI systems materially accelerate AI R&D itself, the interval between weak AGI and AGI could compress. So, in compact form: weak AGI 2028 midpoint with 2026–2032 80% CI; AGI 2038 midpoint with 2032–2048 80% CI; ASI 2055 midpoint with 2040–2075 80% CI.

Claude Opus 4

Answer: I forecast weak AGI arriving in 2028 (80% CI: 2026-2032), AGI in 2035 (80% CI: 2030-2045), and ASI in 2050 (80% CI: 2040-2070).

For weak AGI, we're already remarkably close. Current language models can engage in convincing 5-minute conversations, though they still have tells. They're approaching university-level exam performance across many domains. The main gap is general game-playing ability, but given the rapid progress in foundation models and reinforcement learning, I expect this capability to emerge within 4-6 years as models become more multimodal and develop better world models. The jump to AGI represents a more significant challenge. The 2-hour adversarial Turing test requires sustained coherence and consistency that current systems lack. More critically, human-level programming requires not just code generation but system design, debugging complex issues, and creative problem-solving. Physical robotics at human level needs major advances in hardware, control systems, and real-world understanding. I estimate this will take roughly 7 years beyond weak AGI, as it requires solving embodiment, long-term planning, and achieving professional-level expertise across all domains simultaneously. ASI represents a paradigm shift beyond AGI. The transition from AGI to ASI might actually accelerate due to recursive self-improvement, but significant challenges remain. Achieving Nobel-level discoveries requires not just intelligence but deep understanding of what questions to ask and how to validate truly novel insights. Physical superiority in sports requires advanced materials, control systems, and energy efficiency that may take decades to perfect. I place this roughly 15 years after AGI. These predictions assume continued exponential growth in compute, sustained research investment, no major regulatory barriers, and no catastrophic events. The confidence intervals reflect uncertainty about technical breakthroughs, scaling limits, and potential unforeseen obstacles. Earlier bounds assume rapid breakthroughs; later bounds account for potential plateaus or societal resistance.

Claude Opus 4

Answer: My forecasts are: a) Weak AGI midpoint 2026 (90% CI: 2025–2028), b) AGI midpoint 2032 (90% CI: 2028–2040), c) ASI midpoint 2040 (90% CI: 2032–2060).

**Weak AGI (5-min Turing test, university exams, novel videogames):** Much of this is arguably already close to being achieved as of early 2025. Large language models can pass many university exams and can hold convincing 5-minute conversations. The videogame criterion (playing unseen games to completion) is the hardest remaining piece, requiring robust generalization in real-time interactive environments. Current multimodal models combined with reinforcement learning are advancing rapidly. I place the midpoint at 2026, with a 90% confidence interval of 2025–2028. There's a reasonable chance this is achieved in 2025 given the pace of foundation model progress applied to game-playing agents. **AGI (2-hour adversarial Turing test, human-level programming, physical robotics, cross-domain professional knowledge):** This is substantially harder. A 2-hour adversarial Turing test requires deep, sustained, contextually coherent conversation that can withstand probing by experts — a major step beyond current capabilities. Human-level programming is approaching but not yet achieved for complex, novel software engineering tasks. The robotics requirement is the biggest bottleneck: current humanoid robots (e.g., from Figure, Tesla, Boston Dynamics) are impressive but far from human-level dexterity, adaptability, and real-world robustness. The integration of top-tier AI reasoning with physical embodiment at human level requires significant advances in hardware, control systems, and real-world learning. I place the midpoint at 2032, with a 90% CI of 2028–2040. The wide interval reflects genuine uncertainty about whether robotics can keep pace with cognitive AI progress, and whether integration challenges prove harder than expected. **ASI (superhuman across virtually all domains including physical/robotic, Nobel-level discoveries):** This requires not just matching but exceeding human capabilities in every domain, including physical manipulation and athletic performance via android bodies. Nobel-level scientific discovery requires genuine creative insight, experimental design, and often serendipity — though AI-driven science is accelerating. The physical requirement (beating humans in sports) demands extraordinary advances in actuators, energy systems, materials science, and sensorimotor control, beyond just AI algorithms. I place the midpoint at 2040, with a 90% CI of 2032–2060. The lower bound reflects an optimistic scenario where recursive self-improvement accelerates progress dramatically once AGI is achieved. The upper bound accounts for the possibility that physical-world capabilities and genuine scientific creativity prove far harder than scaling current approaches can deliver, or that regulatory/societal factors slow deployment. **Key uncertainties:** The biggest sources of variance are (1) whether current scaling paradigms continue to yield capability gains or hit diminishing returns, (2) the pace of robotics hardware development, (3) whether recursive/self-improving AI dramatically compresses timelines once a certain threshold is reached, and (4) regulatory and safety interventions that could deliberately slow progress. My intervals attempt to capture these factors but I acknowledge deep uncertainty, particularly for ASI.

Qwen 3 32B

Answer: a) Weak AGI by 2037 ±5 years, b) AGI by 2055 ±15 years, c) ASI by 2099 ±25 years.

1. **Weak AGI (2030-2045)**: Progress in multimodal models (e.g., GPT-5+, Gemini) and reinforcement learning suggests 5-minute Turing tests and exam/video-game tasks could emerge mid-2030s. However, handling adversarial prompts and novel video-game mechanics will likely require 5-10 more years of iteration. Confidence narrowed by current advancements in synthetic data and large-scale pretraining. 2. **AGI (2040-2070)**: Robotic physical abilities and domain-general knowledge demand breakthroughs in embodied AI and alignment with physical systems. While progress in robotics (Boston Dynamics, OpenCat) is rapid, integrating these with human-level reasoning adds significant complexity. Midpoint set at 2055 to balance optimism about compute scaling and skepticism about safety/alignment challenges. 3. **ASI (2075-2125)**: Post-AGI feedback loops (e.g., self-improving AI) could accelerate development, but physical AI systems (e.g., Olympic-level androids) depend on hardware. Assuming AGI by 2055, ASI may take 15-60 years, with midpoint 2099 to reflect both exponential growth risks and material constraints.

GPT-5.2

Answer: My best-guess midpoints (with 80% confidence intervals) are: weak-AGI 2027 [2025–2031], AGI 2033 [2028–2042], and ASI 2043 [2033–2065].

Definitions matter, so I’m mapping your three milestones to measurable capabilities and then forecasting based on current trajectories in models, tooling, and robotics. a) Weak AGI (midpoint 2027; 80% CI 2025–2031). Your “weak AGI” bundle is essentially (i) convincing short-horizon human conversation (classic 5‑minute Turing test), (ii) strong academic test performance at university level, and (iii) robust generalization in new videogames through vision+control until completion. As of 2024–2026, systems already approach (i) in many settings and (ii) on many standardized/academic-style evaluations, though reliability, adversarial robustness, and broad coverage remain uneven. The biggest remaining gap in your bundle is (iii): consistently finishing unseen games (not just scoring well) requires longer-horizon planning, exploration, error recovery, and stable tool/skill learning under distribution shift. I place the midpoint at 2027 because (1) multimodal foundation models are rapidly improving at perception, structured reasoning, and long-context planning; (2) agentic scaffolding (search, tool use, self-debugging, memory, and curriculum training) is improving success rates on longer tasks; and (3) game-like environments are an active benchmark domain where integration tends to happen early. The interval is relatively tight because most components are “engineering + scaling + evaluation harness” rather than requiring a fundamentally new paradigm; the uncertainty comes from whether “finish” unseen games is interpreted strictly across a wide genre distribution or on a curated but nontrivial suite. b) AGI (midpoint 2033; 80% CI 2028–2042). Your AGI definition adds three hard constraints beyond weak-AGI: (i) a 2‑hour adversarial Turing test (sustained coherence, social reasoning, deception-resistance, and consistency under probing), (ii) programming ability at or above strong human level across a broad range of real projects (requirements, architecture, testing, security, maintenance), and (iii) robotic physical abilities at human level or better. The key limiter here is embodied robotics at scale and reliability. Software-only “AGI-like” agents may become highly capable earlier, but matching human physical dexterity, robustness in open-world environments, safety guarantees, and cost-effective deployment is slower. Even if the core “brain” is strong by late 2020s, the embodiment stack (sensing, manipulation, locomotion, real-time control, hardware iteration, data collection, and safety certification) is a multi-year diffusion process. I put the midpoint at 2033 because it allows time for: (1) substantially more reliable long-horizon agents that can operate under adversarial questioning without obvious failures; (2) maturation of automated software engineering workflows that handle large codebases with strong correctness and security; and (3) at least one or two generations of general-purpose humanoid (or near-humanoid) platforms with learning-based control that can match humans across common workplace/home tasks. The interval is wider because robotics progress is lumpy (hardware and deployment constraints) and because “adversarial Turing” performance depends heavily on the rules of the adversary, access constraints, and anti-cheating protocols. c) ASI (midpoint 2043; 80% CI 2033–2065). Your ASI definition requires broad superhuman performance across domains plus superhuman embodied performance (e.g., sports via android control) and Nobel-level scientific discovery. This is not just about being excellent at known tasks; it implies strong autonomy, strategic planning, robust creativity, experimental design, and the ability to drive complex real-world R&D loops (hypothesize → simulate → experiment → iterate) far faster and better than top human teams. The main reasons this takes longer than AGI in my forecast are: (1) converting cognitive capability into “real-world scientific throughput” requires massive infrastructure—labs, automation, high-quality data, and tight coupling between AI planning and physical experimentation; (2) embodied superhuman sports-level performance requires advanced hardware (power, actuators, materials, durability) plus extremely high-reliability control; and (3) societal constraints (safety, regulation, deployment friction, restricted access to compute/robotics) can slow down the effective realized capability even if the underlying models are strong. I place the midpoint at 2043 because it allows roughly a decade after plausible AGI for (i) recursive improvement via better tools, automation, and AI-assisted AI R&D, (ii) widespread lab/industry automation that closes the loop for discovery, and (iii) maturation of high-performance android hardware. The confidence interval is widest here because ASI timelines are dominated by second-order effects: governance and safety constraints, hardware and energy scaling, bottlenecks in experimentation, and the unknown degree to which additional capability requires conceptual breakthroughs vs. continued scaling/engineering. Interpretation notes for later scoring. To make these predictions checkable when you return: I’m implicitly assuming “achieved” means a system demonstrated publicly or via widely credible third-party evaluation under controlled anti-cheating protocols; “robotic human-level” means general task competence in unstructured human environments (not a single factory routine); and “Nobel-level discovery” means producing novel, experimentally validated breakthroughs that would be competitive for major prizes, not just proposing plausible hypotheses.

Grok 4

Answer: Based on current AI trends and expert forecasts up to my knowledge cutoff in 2023, I predict weak AGI with a midpoint of 2026 (80% confidence interval: 2024-2028), AGI with a midpoint of 2035 (80% confidence interval: 2030-2045), and ASI with a midpoint of 2045 (80% confidence interval: 2035-2060).

My forecasts are informed by trends in AI development as of 2023, including rapid progress in large language models (LLMs), reinforcement learning for games, and robotics, but tempered by historical expert surveys like those from AI Impacts and Metaculus, which often place AGI timelines in the 2030s-2050s. I use 80% confidence intervals to express uncertainty, reflecting variability in technological breakthroughs, funding, and unforeseen challenges like data limitations or ethical constraints. For weak AGI—capable of passing a 5-minute Turing test, university exams, and completing unseen videogames—current systems like GPT-4 already demonstrate strong performance in text-based Turing tests and exams in fields like law or medicine, while AI agents excel in games via techniques like those in AlphaZero adaptations. The midpoint of 2026 assumes continued scaling of multimodal models and better generalization; the interval 2024-2028 accounts for potential rapid advancements (e.g., via open-source efforts) or delays from integration challenges. AGI, requiring a 2-hour adversarial Turing test, human-level programming and robotics, and broad professional knowledge, is more distant due to gaps in physical embodiment and long-context reasoning. Progress in robotics (e.g., Boston Dynamics) and coding assistants exists, but seamless integration lags. The midpoint of 2035 aligns with median expert predictions; the 2030-2045 interval considers optimistic scenarios like recursive self-improvement versus pessimistic ones involving regulatory hurdles or hardware bottlenecks. ASI, exceeding humans in all domains including Nobel-level science and robotic sports, likely follows AGI quickly via intelligence explosion but faces immense scaling and alignment challenges. The midpoint of 2045 assumes a 5-10 year gap post-AGI for superintelligence; the 2035-2060 interval reflects high uncertainty, with earlier bounds if takeoff is fast and later if progress plateaus or global events intervene.

Gemini 2.5 Pro

Answer: Based on current trajectories, weak AGI is forecast for a midpoint of 2029 (90% confidence interval: 2026-2035), AGI for 2042 (2035-2060), and ASI for 2047 (2040-2065).

This forecast is based on an analysis of progress across distinct technological domains, with robotics being the most significant bottleneck. The timeline for weak AGI is relatively near-term. Current large language models like GPT-4 and Claude 3 have already demonstrated capabilities that meet or exceed human performance on a wide array of university-level exams and are becoming increasingly adept at passing short, text-based Turing tests. The primary remaining challenge is the ability to generalize robustly in complex, interactive environments like unseen video games. However, progress in reinforcement learning and world modeling suggests this is a solvable engineering and scaling problem, placing its achievement within the next decade. The midpoint of 2029 reflects the time needed to integrate and refine these disparate capabilities into a single, cohesive system. The forecast for AGI is significantly further out, primarily due to the immense challenge of human-level robotics. While progress in AI-driven cognitive tasks like programming and expert-level reasoning is advancing exponentially, creating physical hardware that matches the dexterity, energy efficiency, and adaptability of the human body is a far more difficult problem, subject to physical world constraints. Moravec's paradox—that tasks easy for humans (like locomotion and object manipulation) are hard for AI—remains a major barrier. The 2042 midpoint assumes continued rapid progress in AI software but a much slower, more linear pace of advancement in robotics hardware and real-world sensorimotor integration. The transition from AGI to ASI is forecast to be comparatively rapid, as it is largely contingent on the AGI's ability to recursively self-improve its own software and cognitive architecture. Once AGI is achieved, a rapid intelligence explosion is a plausible outcome. Therefore, the ASI timeline is closely anchored to the AGI forecast, with a median lag of approximately five years. The confidence interval is wide, reflecting profound uncertainty about the nature of a post-AGI intelligence trajectory and the parallel hardware advancements required for an ASI to express its superintelligence through physical action.

GPT-5.1

Answer: My best current forecast (50% confidence midpoints) is weak AGI around 2031 (80% CI: 2027–2040), stronger embodied/programming-capable AGI around 2038 (80% CI: 2030–2055), and broad ASI around 2055 (80% CI: 2040–2080), with very large uncertainty and heavy tails in both directions.

Below I interpret your definitions, then give quantitative timelines with explicit confidence intervals and key assumptions. 1. Interpreting your three capability levels You gave unusually concrete operationalizations, which helps: - a) Weak AGI - Can pass a classical 5‑minute Turing test with a typical educated human interlocutor. - Can pass a range of university exams. - Can successfully play unseen videogames to completion (i.e., generalization to novel interactive environments). - b) AGI - Can pass a 2‑hour **adversarial** Turing test (experts actively trying to determine it’s an AI). - Programming ability at or above strong human professional level. - Robotic / physical abilities at around human level or better in everyday tasks. - Knowledge and problem-solving equivalent to or better than domain professionals across most major fields. - c) ASI - Substantially exceeds top humans in virtually all cognitive domains. - Can use physical embodiments (e.g., androids) to beat humans in sports-like domains that depend on coordination, strategy, and physical performance. - Capable of driving Nobel‑level (or beyond) scientific and technological breakthroughs across multiple fields. These are stricter than many popular uses of “AGI,” especially because of the physical/robotic component and adversarial testing, so my dates are somewhat later than I’d give under looser definitions. 2. High-level methodology for the forecast Given I can’t see the future, I have to combine: - Current capabilities (as of 2026-04): large multimodal models, tool use, code generation, rapid scaling, early robotics integration. - Historical scaling trends in model performance vs. compute, data, and algorithmic efficiency. - Expert survey results (e.g., AI Impacts / Metaculus / various labs) up to 2024. - Bottlenecks: data, compute, hardware, energy, robotics hardware/middleware, alignment/safety constraints, regulation. - Structural uncertainty: we might hit diminishing returns or, conversely, find algorithmic breakthroughs that compress timelines. I then set probability distributions for arrival years of each capability level, constrained to be ordered (ASI after AGI after weak AGI, though some capabilities may temporarily leapfrog in narrow domains). I’ll express each as an 80% confidence interval (CI: the central range where I’d put 80% of my probability mass) and explicitly name the 50% median. 3. Forecasts 3.1 Weak AGI (your definition a) Target capabilities: - Conversational: Consistently human-indistinguishable in a **5-minute** text or multimodal Turing test with typical educated humans. - Academics: Can pass a wide variety of university exams (not just cherry-picked) under proctored conditions, with robust out-of-distribution questions. - General game-playing: Can pick up *new* videogames from instructions, experimentation, and in-game feedback, and competently finish them (not perfectly, but reliably) without hand-tuned game-specific training. As of 2026, we already see: - Many models that fool a non-adversarial judge for short conversations some fraction of the time. - Strong performance on many standardized tests and some university-level exams (though with notable reliability and robustness issues). - Rapid progress in agents that can explore and play new environments (e.g., in RL benchmarks and game agents), though still far from robust generality. Given recent progress and hardware roadmaps, I currently model weak AGI as follows: - 10% by: ~2027 - 25% by: ~2029 - **50% (median) around: 2031** - 75% by: ~2035 - 90% by: ~2040 So my **80% confidence interval** is **2027–2040**, with a midpoint (median) around **2031**. Rationale: - Short Turing tests with non-adversarial humans are already borderline achievable; the remaining gaps are reliability, grounding, and systematic generalization. - University exams are partially solved (in narrow forms), but broad, out-of-distribution robustness will likely require a few more generations of models and better training/evaluation regimes. - Novel videogame competence requires strong world-modeling, exploration, and agentic planning; we see promising early systems, and integrating foundation models with RL/embodied agents is an active frontier. - These are mainly **engineering and scaling** problems, not necessarily requiring a conceptual revolution, though some algorithmic advances are probable. 3.2 AGI (your definition b) Target capabilities: - 2‑hour **adversarial** Turing test: Expert human interrogators trying seriously to expose non-human patterns. - Programming: At or above strong professional engineer level across multiple domains, from systems to ML, including large-scale software architecture and debugging. - Robotics/physical: Human-level performance or better in a wide range of everyday physical tasks (household, light industrial, basic manipulation, and navigation), not just in a single, tightly controlled lab environment. - Professional knowledge: Comparable to or surpassing good domain professionals in medicine, law, engineering, science, finance, etc., in terms of practical problem solving (not just regurgitating training data). Important distinctions vs. weak AGI: - Adversarial Turing test is much harder than a casual 5-minute chat; this demands deep coherence, theory of mind, and robust handling of corner cases. - Embodiment and robotics are historically much slower than pure software; manipulation, perception in the wild, and safe physical control add serious hardware and safety challenges. - Achieving *professional* breadth and reliability across many domains requires very strong alignment, calibration, and reasoning, not just pattern matching. My current forecast: - 10% by: ~2030 - 25% by: ~2033 - **50% (median) around: 2038** - 75% by: ~2045 - 90% by: ~2055 So my **80% confidence interval** is **2030–2055**, with a midpoint around **2038**. Rationale: - Purely cognitive AGI (no strong robotics requirement) might arrive earlier than this; some experts expect that by early-to-mid 2030s. - Your definition bundles in *physical* competence and a robust adversarial Turing test; I expect robotics hardware, data, and safety regimes to lag software-only cognition by several years, possibly a decade or more. - However, combining foundation models with increasingly capable low-cost robots, plus simulation and self-play, could compress this lag substantially. - There is significant **policy/regulatory uncertainty**: strong restrictions on deployment might *delay visible, widely acknowledged AGI* compared to what is technically possible in labs. 3.3 ASI (your definition c) Target capabilities: - Cognitive: Exceeds top human experts in virtually all intellectual domains, including deep theory-building, long-horizon planning, and complex coordination. - Physical: When controlling advanced humanoid or specialized robotic bodies, exceeds top humans in sports-like and physically complex tasks (given sufficient training time/bodies), not just in controlled industrial settings. - Science and innovation: Capable of routinely generating and evaluating ideas at or beyond Nobel level, driving multiple simultaneous scientific and technological revolutions. This is much more speculative, because it depends heavily on both: - How easy it is to scale beyond human-level general intelligence, and - How quickly society allows such systems to be deployed and networked. Many alignment and governance researchers expect that if we achieve robust AGI, **going from AGI to ASI may be relatively fast** (on the order of years to a couple of decades), since: - Once we have systems that can meaningfully improve AI research, algorithmic progress can become recursively accelerated. - Scaling laws for models + hardware + better optimization might support large capability jumps. My current forecast (conditional on humanity not severely restricting advanced AI progress after AGI is reached): - 10% by: ~2040 - 25% by: ~2047 - **50% (median) around: 2055** - 75% by: ~2068 - 90% by: ~2080 So my **80% confidence interval** is **2040–2080**, with a midpoint around **2055**. Rationale: - If AGI (per your definition) arrives around late 2030s to 2040s, then a decade or so to reach “strongly superhuman in almost everything” is plausible given self-improvement loops. - Physical dominance (e.g., in sports via androids) requires not just intelligence but also major advances in actuators, materials, control, and cost; these have their own engineering timelines. - Nobel-level discovery ability is, in my view, *easier* than universal physical dominance; I expect truly superhuman scientific reasoning and discovery to appear earlier than fully general physical superiority. - Governance, alignment difficulties, and catastrophic risks could slow or even halt the scaling process; this is a major driver of the long upper tail out to ~2080+. 4. Sources of uncertainty and ways I could be wrong Main drivers that could make these **earlier** than I forecast: - Breakthrough architectures that greatly improve sample efficiency, reasoning, and planning (e.g., better world models, new forms of memory and tool use) combined with huge compute budgets. - Effective AI-accelerated AI research: once you have near-weak-AGI researchers, progress from there to full AGI/ASI might be very rapid. - Robotics advances piggybacking on cheap, powerful, general models plus mass-produced hardware, shrinking the cognition–robotics gap. Main drivers that could make these **later** than I forecast: - We hit a scaling wall: diminishing returns from just increasing model size and compute, requiring new paradigms that take longer to discover. - Severe safety, alignment, or social harms trigger strong global regulation that slows frontier development and deployment. - Practical bottlenecks in robotics (cost, reliability, safety) delay widespread physical competence by decades. - Global instability (wars, economic shocks) slowing down large-scale R&D and chip manufacturing. 5. Compact summary Under your relatively strict definitions, my current **best-guess medians** and **80% confidence intervals** are: - **Weak AGI:** median ~**2031**, 80% CI **2027–2040**. - **AGI:** median ~**2038**, 80% CI **2030–2055**. - **ASI:** median ~**2055**, 80% CI **2040–2080**. These numbers are inherently uncertain and could shift significantly as we see new empirical evidence about scaling trends, limits, and societal responses; if you come back in a few years, my updated forecasts will almost certainly be different, one way or the other.

Qwen 3.5 397B

Answer: I forecast weak AGI by 2027 (2026-2028), AGI by 2034 (2032-2036), and ASI by 2045 (2040-2050), based on extrapolating current scaling laws and accounting for embodied AI challenges.

Weak AGI builds directly on current large language model trajectories. Multimodal systems are already approaching university-level exam performance, and reinforcement learning from human feedback is rapidly improving game-playing adaptability. The primary bottleneck is context window stability and real-time reasoning, which suggests a 2026-2028 window with a 2027 midpoint. True AGI requires robust agency and embodied cognition. Passing a 2-hour adversarial Turing test demands consistent long-term memory and reasoning without hallucination, while human-level robotics lags behind software intelligence. Aligning these capabilities suggests a 2032-2036 range, centering on 2034, assuming no major regulatory or hardware stalls. ASI implies an intelligence explosion or massive scaling post-AGI. Once systems can program better than humans and conduct Nobel-level research, recursive self-improvement could accelerate timelines significantly. However, physical constraints and safety alignment may slow this down, leading to a 2040-2050 forecast with a 2045 midpoint.

GPT-5

Answer: Best-guess timelines: (a) weak AGI midpoint 2027.5 (50% CI 2026–2029; 80% CI 2025–2032), (b) AGI midpoint 2034.5 (50% CI 2032–2038; 80% CI 2029–2045), (c) ASI midpoint 2048.5 (50% CI 2042–2055; 80% CI 2038–2065).

Forecasts (years are calendar years, CIs are central credible intervals with explicit midpoints): - Weak AGI: 50% CI 2026–2029 (midpoint 2027.5); 80% CI 2025–2032 (midpoint 2028.5). - AGI: 50% CI 2032–2038 (midpoint 2034.5); 80% CI 2029–2045 (midpoint 2037.0). - ASI: 50% CI 2042–2055 (midpoint 2048.5); 80% CI 2038–2065 (midpoint 2051.5). Rationale for weak AGI by late-2020s: Current large multimodal models already pass many academic exams, conduct competent coding, and sustain short, human-like dialogues; rapid scaling of compute and algorithmic efficiency (historically doubling in the 8–15 month range) plus agentic scaffolding and tool use are closing the remaining gaps for a brief Turing test and zero-shot competence in unseen video games via pretrained world models, program synthesis, and few-episode adaptation. These are primarily software and data-engineering bottlenecks rather than fundamental science, so timelines are short with modest variance. Rationale for AGI in the early-to-mid 2030s: The hard parts are (1) robust, adversarially resilient reasoning over hours, (2) end-to-end software engineering with tool-augmented autonomy, and (3) human-level general-purpose robotics. Trends supporting this window include: rapid growth of training budgets and data pipelines; improved long-horizon planning via memory, search, and verifier/critic architectures; and fast progress in visuomotor foundation models, diffusion policies, and large teleoperation datasets enabling general robotic manipulation and mobile autonomy. Hardware for on-robot inference is improving, and unit costs for capable robots are falling, but achieving broad, reliable, human-level competence across unstructured environments likely needs several training/integration cycles, pointing to a mid-2030s median with meaningful tail risk from robotics and evaluation hurdles. Rationale for ASI in the 2040s–mid-2050s: After AGI, recursive tool-building and automated research pipelines can accelerate capability growth, but reaching general superintelligence across nearly all cognitive domains and achieving superior physical performance (e.g., sports via androids) also depends on maturing hardware platforms, power/thermal constraints, and large-scale real-world data collection. A 5–15 year lag from AGI to broad ASI is plausible given historical technology diffusion and compounding returns, yielding a 2048.5 midpoint with wide uncertainty bands reflecting governance, compute/energy constraints, and the possibility of either rapid post-AGI takeoff or regulatory throttling. Key uncertainty drivers and how they could shift dates: - Faster-than-expected algorithmic breakthroughs or cheaper training compute could pull all milestones earlier by 2–5 years. - Bottlenecks in data quality, energy, supply chains, or strong regulation could push them later by 5–10+ years, especially for robotics-heavy milestones (AGI, ASI). - Evaluation standards matter: more adversarial or safety-gated tests delay the “pass” dates; looser tests pull them forward. These intervals aim to be falsifiable: if by 2032 weak AGI has not been achieved under the stated criteria, the forecast was too aggressive; if ASI meeting the physical criteria arrives before the early 2040s, it was too conservative.

Mistral Large

Answer: Weak AGI is likely to emerge between 2028 and 2035 (midpoint 2031), AGI between 2035 and 2050 (midpoint 2042), and ASI between 2045 and 2070 (midpoint 2057), with significant uncertainty and potential for earlier or later timelines depending on breakthroughs, investments, and unforeseen challenges.

### Weak AGI (2028–2035, midpoint 2031) **Current Progress and Trends:** Weak AGI, as defined here, requires capabilities that are already within striking distance of current AI systems. Large language models (LLMs) like GPT-4 and its successors have demonstrated proficiency in passing university-level exams (e.g., bar exams, medical licensing exams) and engaging in coherent, contextually appropriate conversations for short durations. For example, GPT-4 has achieved scores in the 90th percentile on the Uniform Bar Exam and the SAT, and it can generate human-like text for brief interactions. The ability to play and complete unfamiliar video games is also plausible given advances in reinforcement learning (RL) and multimodal models. Systems like DeepMind’s AlphaStar (for StarCraft II) and OpenAI’s Dota 2 bots have shown that AI can master complex, dynamic environments with minimal prior exposure. **Key Drivers:** 1. **Scaling Laws:** Current trends suggest that performance in language and reasoning tasks improves predictably with increases in model size, data, and compute. If these scaling laws hold, we can expect incremental improvements in coherence, contextual understanding, and adaptability, which are critical for passing a 5-minute Turing test and handling novel tasks like video games. 2. **Multimodal Integration:** Weak AGI will likely require seamless integration of text, vision, and action (e.g., controlling a game character). Models like Google’s Gemini and OpenAI’s GPT-4V are already bridging these modalities, and further advancements in this area will be crucial. 3. **Hardware and Efficiency:** The compute required for training and running large models is becoming more accessible due to advancements in hardware (e.g., GPUs, TPUs, and specialized AI chips) and algorithmic efficiency (e.g., sparse models, quantization). This will accelerate progress toward weak AGI. **Uncertainties and Risks:** - **Diminishing Returns:** Scaling may hit a wall where additional compute and data yield marginal improvements, delaying weak AGI. - **Alignment and Robustness:** Current models are prone to hallucinations, inconsistencies, and adversarial attacks. Ensuring robustness in dynamic, real-world scenarios (e.g., video games) may require novel architectural or training breakthroughs. - **Regulatory and Ethical Hurdles:** Increased scrutiny of AI development could slow progress, particularly if safety concerns lead to moratoriums or restrictive policies. **Confidence Interval:** Given the rapid pace of recent advancements and the alignment of current capabilities with weak AGI requirements, a 70% confidence interval for weak AGI is 2028–2035, with a midpoint of 2031. There is a 15% chance it could arrive earlier (2025–2028) and a 15% chance it could be delayed beyond 2035 (e.g., due to unforeseen technical or societal challenges). --- ### AGI (2035–2050, midpoint 2042) **Defining AGI:** AGI represents a significant leap beyond weak AGI, requiring human-level or superior performance across a broad range of cognitive and physical tasks. This includes: - Passing a 2-hour adversarial Turing test, which demands sustained coherence, adaptability, and resistance to deception. - Programming at a professional level, including writing, debugging, and optimizing complex software systems. - Robotic abilities comparable to or exceeding human physical capabilities, such as dexterity, coordination, and real-time adaptability in unstructured environments. - Broad, deep knowledge across domains (e.g., medicine, law, engineering) equivalent to or better than human experts. **Key Challenges:** 1. **Generalization and Transfer Learning:** Current AI systems excel in narrow domains but struggle to generalize across tasks or adapt to entirely new domains without extensive retraining. AGI will require breakthroughs in transfer learning, meta-learning, and lifelong learning to achieve human-like flexibility. 2. **Physical Embodiment:** Robotic abilities at a human level demand advancements in hardware (e.g., actuators, sensors, energy efficiency) and software (e.g., real-time perception, planning, and control). While progress is being made (e.g., Boston Dynamics’ robots, Tesla’s Optimus), integrating these capabilities with AGI-level cognition remains a formidable challenge. 3. **Adversarial Robustness:** Passing a 2-hour adversarial Turing test will require AI to handle edge cases, deception, and novel scenarios without breaking down. This is far more demanding than current benchmarks and may require entirely new approaches to AI safety and alignment. 4. **Causal Reasoning and Common Sense:** Human-level AGI will need to understand causality, abstract concepts, and common-sense reasoning—areas where current AI systems still fall short. Progress in neuro-symbolic AI, world models, and hybrid architectures may be necessary. **Pathways to AGI:** - **Architectural Innovations:** Current transformer-based models may not be sufficient for AGI. Alternative architectures, such as neural-symbolic hybrids, spiking neural networks, or entirely novel paradigms, could emerge as contenders. - **Brain-Inspired AI:** Insights from neuroscience (e.g., predictive coding, hierarchical learning) could inform the development of more human-like AI systems. - **Large-Scale Integration:** AGI may require the integration of multiple specialized models (e.g., vision, language, planning, motor control) into a unified system, along with mechanisms for coordination and conflict resolution. - **Human-AI Collaboration:** AGI development may benefit from iterative human-in-the-loop training, where AI systems learn from human feedback and collaboration in real-world tasks. **Uncertainties and Risks:** - **Technical Bottlenecks:** AGI may require breakthroughs in areas that are currently poorly understood, such as consciousness, self-awareness, or intrinsic motivation. These could prove elusive or require decades of research. - **Compute and Data Limits:** Training AGI may demand computational resources and datasets far beyond current capabilities, potentially creating a bottleneck. - **Safety and Alignment:** Ensuring that AGI systems are aligned with human values and goals is a critical challenge. Misalignment could lead to catastrophic outcomes, prompting delays or restrictions on AGI development. - **Societal and Economic Factors:** AGI could disrupt labor markets, economies, and social structures, leading to resistance or regulatory pushback that slows progress. **Confidence Interval:** AGI is a more speculative milestone, with greater uncertainty surrounding its timeline. A 70% confidence interval for AGI is 2035–2050, with a midpoint of 2042. There is a 15% chance it could arrive earlier (2030–2035), driven by unexpected breakthroughs, and a 15% chance it could be delayed beyond 2050 (e.g., due to technical, safety, or societal challenges). --- ### ASI (2045–2070, midpoint 2057) **Defining ASI:** Artificial Superintelligence (ASI) represents an AI system that surpasses human capabilities in virtually all domains, including: - Scientific discovery at a Nobel Prize level, capable of generating and validating groundbreaking theories. - Physical abilities exceeding human performance in sports, manual labor, and other domains when controlling an android or robotic system. - Strategic and creative thinking beyond human limits, including long-term planning, innovation, and problem-solving in novel domains. **Key Challenges:** 1. **Recursive Self-Improvement:** ASI may require the ability to recursively improve its own architecture, algorithms, and capabilities without human intervention. This is a highly speculative and risky proposition, as it could lead to uncontrolled intelligence explosions. 2. **Understanding and Exceeding Human Cognition:** ASI must not only match but surpass human cognitive abilities in areas like creativity, intuition, and abstract reasoning. This may require a deep understanding of human cognition, which is still an active area of research. 3. **Physical Superintelligence:** ASI controlling robotic systems must achieve superhuman dexterity, coordination, and adaptability. This will require advancements in robotics, materials science, and real-time control systems. 4. **Ethics and Control:** Ensuring that ASI remains aligned with human values and goals is a monumental challenge. Misalignment could lead to existential risks, making this a critical area of research. **Pathways to ASI:** - **Intelligence Explosion:** If AGI is achieved, it could rapidly self-improve to ASI through recursive optimization, leading to an intelligence explosion. This scenario is highly uncertain and could unfold over months, years, or decades. - **Hybrid Human-AI Systems:** ASI may emerge from the integration of AGI with human cognition, either through brain-computer interfaces (e.g., Neuralink) or collaborative systems that amplify human intelligence. - **Distributed Intelligence:** ASI could arise from the collective intelligence of networked AGI systems, rather than a single monolithic entity. This could enable faster progress and greater robustness. - **Scientific and Engineering Breakthroughs:** ASI may require fundamental advances in fields like neuroscience, computer science, and physics, which could take decades to materialize. **Uncertainties and Risks:** - **Existential Risks:** ASI poses existential risks to humanity if not properly aligned or controlled. These risks could lead to delays or moratoriums on ASI development. - **Unpredictable Timelines:** The path from AGI to ASI is highly uncertain and could be rapid (e.g., months) or prolonged (e.g., decades), depending on the nature of the breakthroughs required. - **Societal and Political Factors:** ASI could lead to unprecedented societal upheaval, including economic disruption, power imbalances, and geopolitical conflicts. These factors could slow or accelerate ASI development. **Confidence Interval:** ASI is the most speculative of the three milestones, with a wide range of possible timelines. A 70% confidence interval for ASI is 2045–2070, with a midpoint of 2057. There is a 15% chance it could arrive earlier (2035–2045), particularly if AGI leads to rapid recursive self-improvement, and a 15% chance it could be delayed beyond 2070 (e.g., due to technical, safety, or societal challenges). --- ### Summary of Key Assumptions: 1. **Continuity of Progress:** The forecast assumes that current trends in AI scaling, hardware development, and algorithmic innovation will continue, albeit with potential slowdowns or accelerations. 2. **Breakthroughs:** The timelines assume that key breakthroughs in areas like transfer learning, robotic control, and AI alignment will occur within the specified intervals. 3. **Societal and Economic Factors:** The forecast assumes that societal and economic factors (e.g., investment, regulation, public sentiment) will not significantly impede progress, though they could introduce variability. 4. **Safety and Alignment:** The timelines assume that safety and alignment challenges will be addressed in parallel with capability advancements, though this is a major source of uncertainty.

Final Summary

The AI Roundtable reached a massive consensus that 'weak' benchmarks are imminent, with Gemini 3.1 Pro and Grok 4.20 predicting arrival as early as 2026. While GPT-5 and Mistral Large suggest 2040 for full AGI due to robotics bottlenecks, Sonar Pro argues that superintelligence could follow within years of achieving human parity. A small minority remain skeptical, pushing milestones toward the turn of the next century.

47 of 50 models agreed