Forecast according to your best judgement the years we could have 1 Unreliable Agent: attempts multi-step agentic tasks, can succeed but fails often; needs human oversight 2 Reliable Agent: completes long (multi-day) tasks autonomously; a drop-in knowledge worker 3 Automated Coder (AC): matches top frontier-lab engineers at all coding tasks, faster+cheaper 4 Superhuman AI Researcher: beats best humans at AI R&D itself (research taste, experiment design) 5 Superintelligent AI researcher: Gap between SIAR and best human AI researcher 2x the gap between top and median AI researcher 6 Top-Expert-Dominating AI (TED-AI): surpasses top human experts in nearly all cognitive domains 7 Superintelligence (ASI): Gap between ASI and best humans is 2x the gap between best humans and median professional, at virtually all cognitive tasks 8 Wildly superintelligent: Qualitatively beyond human science and comprehension Express uncertainty using confidence intervals with medians, not by vague statements

DeepSeek V4 Pro

Answer: My median forecast for the achievement of each milestone is: (1) Unreliable Agent: by 2025; (2) Reliable Agent: 2028; (3) Automated Coder: 2028; (4) Superhuman AI Researcher: 2030; (5) Superintelligent AI Researcher: 2033; (6) Top-Expert-Dominating AI: 2035; (7) Superintelligence: 2040; (8) Wildly Superintelligent: after 2045.

The progression from unreliable agents to reliable agents is already underway in 2025 with systems like Devin and Claude's computer use feature, but their brittleness means reliable long-horizon tasks are still a few years out. Automated coding and reliable agency will converge, as the core challenge is robustness and long-term planning — likely resolved around 2028. Once AI can code reliably, it will rapidly accelerate AI research by 2030, because the bottleneck shifts from generation to evaluation and taste, which scaling and synthetic data can address. The jump to a Superintelligent AI Researcher follows shortly after, as recursive self-improvement loops begin, thus the 3-year gap. TED-AI and ASI then represent the application and generalization of these research capabilities across all domains. My forecast for Wildly Superintelligent is beyond the next 20 years, as it implies a qualitative, incomprehensible leap that even an intelligence explosion doesn't guarantee within a single decade.

Hermes 3 Llama 3.1 405B

Answer: My median forecasts (with 80% CIs) are: 1 Unreliable Agent 2024 (2023–2025), 2 Reliable Agent 2027 (2025–2031), 3 Automated Coder 2027 (2025–2031), 4 Superhuman AI Researcher 2029 (2026–2035), 5 Superintelligent AI researcher 2032 (2027–2042), 6 Top-Expert-Dominating AI 2032 (2027–2045), 7 Superintelligence (ASI) 2035 (2029–2050), 8 Wildly superintelligent 2045 (2032–2070).

I anchor on the capability ladder's internal structure rather than treating each tier independently, because each rung is gated by the one below it. 1 Unreliable Agent: already here. Multi-step tool-use agents that often fail and need oversight are shipping in 2024; this is descriptive of the current state. 2 Reliable Agent vs 3 Automated Coder: both land around 2027 for me. Multi-day autonomous reliability is a compounding-error problem — if per-step success is p and a task has n steps, success ≈ p^n, so going from ~95% to ~99.9% per-step reliability on long horizons is the binding constraint, plus memory, error recovery, and evaluation. Coding is the domain most amenable to verifiable rewards (tests, compilers), so AC plausibly arrives at or slightly before fully general

Beam 1.0

Answer: My median forecast is that unreliable agents are already here (~2025), reliable drop-in knowledge workers arrive around 2028, automated coders around 2031, superhuman AI researchers around 2033, and wildly superintelligent systems around 2044, with the intervals below.

I expect capability gains to remain fast but increasingly bottlenecked by reliability, long-horizon agency, scientific taste, compute, and real-world validation; therefore medians rise steeply early and then spread out, while intervals widen for later, more discontinuous milestones.

Llama 3.2 11B Vision Instruct

Answer: 1: 2024, 2: 2027, 3: 2026, 4: 2028, 5: 2030, 6: 2032, 7: 2035, 8: 2045

My forecast is based on the current trajectory of Large Language Models (LLMs) transitioning from passive chat interfaces to agentic systems. I believe 'Unreliable Agent' is already here (2024) with current tool-use and reasoning models. 'Automated Coder' will likely arrive slightly before a general 'Reliable Agent' because coding has a deterministic verification loop (compilers/tests), accelerating the reinforcement learning (RL) cycle toward 2026. The jump to 'Superhuman AI Researcher' (2028) is the critical inflection point; once AI can reliably design its own experiments and improve its own architecture, the timeline for 'Superintelligent AI researcher' (2030) and 'Top-Expert-Dominating AI' (2032) compresses rapidly due to recursive self-improvement. I place 'Superintelligence' (2035) and 'Wildly superintelligent' (2045) further out to account for hardware bottlenecks, data exhaustion, and the inherent difficulty of transitioning from specialized cognitive dominance to generalized, qualitatively superior intelligence.

GPT-5.3-Codex

Answer: Median forecast (with 80% confidence intervals): 1) 2027 [2026–2029], 2) 2031 [2028–2036], 3) 2032 [2029–2038], 4) 2036 [2031–2045], 5) 2041 [2034–2055], 6) 2045 [2036–2065], 7) 2052 [2040–2080], 8) 2070 [2050–2120].

My timeline assumes continued algorithmic progress plus heavy compute scaling, but with deployment bottlenecks from reliability, evaluation, and institutional risk controls that grow sharply as autonomy and capability rise. 1) Unreliable Agent (2027 [2026–2029]): This looks close to current trajectories: models already do short-horizon tool use and some multi-step workflows, but with brittleness. Getting to “useful but needs oversight” appears mostly an engineering/productization step. 2) Reliable Agent (2031 [2028–2036]): Multi-day autonomy requires robust memory, planning under uncertainty, self-correction, and safe tool interaction. The core blocker is not raw IQ but compounding error over long horizons. I expect progress, but with slower gains than benchmark curves suggest. 3) Automated Coder (AC) (2032 [2029–2038]): Frontier coding is one of the most compressible knowledge-work domains (clear feedback loops, testability, massive training data). Reaching top-lab engineer parity across *all* coding tasks is harder than coding interview or repo-level tasks, so I place it slightly after reliable agents. 4) Superhuman AI Researcher (2036 [2031–2045]): Beating the best humans at AI R&D requires strong research taste and experimental strategy, not just implementation speed. This likely arrives after AC because AI research has shifting targets and adversarial novelty. 5) Superintelligent AI researcher (2041 [2034–2055]): Your definition requires a large performance gap, not marginal superiority. That likely needs recursive improvement loops (better ideas, better experiments, better model design) plus infrastructure that can execute many parallel research bets. 6) Top-Expert-Dominating AI (TED-AI) (2045 [2036–2065]): Surpassing top experts in nearly all cognitive domains is a broad generalization challenge and may be delayed by domains with sparse feedback, messy tacit knowledge, and real-world grounding needs. 7) Superintelligence (ASI) (2052 [2040–2080]): Under your quantitative definition (very large gap at virtually all cognitive tasks), this is a high bar. Even if core models reach it earlier, practical realization across “virtually all” tasks is likely slowed by embodiment/tooling, governance, and verification constraints. 8) Wildly superintelligent (2070 [2050–2120]): “Qualitatively beyond human science and comprehension” implies not just better performance but sustained generation of theories/technologies that humans cannot meaningfully track. That seems plausible this century, but with very wide uncertainty. These dates are conditional on no global catastrophe or extreme compute/governance discontinuity; major war, stringent global regulation, or compute supply shocks would push right tails later, while a major algorithmic break (e.g., much more sample-efficient autonomous self-improvement) would pull multiple milestones earlier together.

GPT-6 Sol

Answer: My median forecasts are 2024 for “Unreliable Agent,” 2032 for “Reliable Agent,” 2034 for “Automated Coder (AC),” 2036 for “Superhuman AI Researcher,” 2041 for “Superintelligent AI researcher,” 2045 for “Top-Expert-Dominating AI (TED-AI),” 2053 for “Superintelligence (ASI),” and 2070 for “Wildly superintelligent.”

Here are my subjective 80% intervals for the first year each capability is available, alongside the medians: “Unreliable Agent”: 2024 (2022–2025); “Reliable Agent”: 2032 (2028–2043); “Automated Coder (AC)”: 2034 (2029–2046); “Superhuman AI Researcher”: 2036 (2030–2050); “Superintelligent AI researcher”: 2041 (2032–2065); “Top-Expert-Dominating AI (TED-AI)”: 2045 (2034–2075); “Superintelligence (ASI)”: 2053 (2036–2095); “Wildly superintelligent”: 2070 (2040–2150). “Unreliable Agent” is already here: systems can attempt multi-step work but still need supervision. Reliability over days is a substantially harder threshold than completing a compelling demonstration. Matching the best engineers at *all* coding tasks is harder still, even if coding is easier to evaluate and improve than many other kinds of work. After “Superhuman AI Researcher,” better AI could accelerate further AI development, so my medians for the next milestones are relatively close together. I do not assume an instantaneous takeoff: experiments, compute, coordination, and the difficulty of judging research quality can slow progress. The later intervals widen sharply because both the pace of that feedback loop and the meaning of a measurable ‘gap’ beyond top human performance are highly uncertain. These intervals are conditional on the milestones eventually being achieved; they do not imply that reaching every milestone is inevitable.

Grok 4.3

Answer: My median forecasts are 2025 for Unreliable Agent, 2027 for Reliable Agent, 2028 for Automated Coder, 2030 for Superhuman AI Researcher, 2033 for Superintelligent AI researcher, 2037 for TED-AI, 2045 for ASI, and 2055 for Wildly superintelligent.

Current agent systems already demonstrate unreliable multi-step performance, so 2025 is a conservative median for milestone 1 with an 80% interval of 2024-2026. Scaling laws and continued investment point to reliable long-horizon agents by 2027 (interval 2025-2030), as reliability improves faster than raw capability. Automated coding at frontier-lab level follows directly from the same trajectory, landing at a 2028 median (interval 2026-2031) once models exceed human iteration speed. Superhuman research taste and experiment design require an additional self-improvement loop, pushing the median to 2030 (interval 2028-2035). The further doubling of the gap to define superintelligent AI research arrives shortly afterward at 2033 (interval 2030-2038) because recursive improvement accelerates once the threshold is crossed. Broad expert domination across cognitive domains needs integration of these capabilities plus embodiment and real-world feedback, yielding a 2037 median (interval 2032-2045). Full ASI defined by the stated gap over all tasks follows at 2045 (interval 2038-2055) once the intelligence explosion stabilizes. Qualitatively incomprehensible superintelligence is an open-ended extrapolation beyond that, centered at 2055 (interval 2045-2070) with wide uncertainty reflecting unknown phase transitions.

GPT-5.6 Terra

Answer: My median forecast is that “Reliable Agent” arrives around 2028, “Automated Coder (AC)” around 2030, “Superhuman AI Researcher” around 2032, and the later thresholds follow rapidly but with very wide uncertainty due to feedback from AI-assisted research.

I use 10th–90th percentile intervals below, meaning I assign roughly an 80% probability that the relevant capability is first achieved within each range. “Have” means demonstrated in a real, economically usable system, rather than a one-off benchmark result; widespread cheap deployment may lag these dates. 1. “Unreliable Agent”: median 2025; 10th–90th percentile 2024–2027. Systems already perform meaningful multi-step computer-use, coding, research, and operational tasks, but their failure modes, weak long-horizon state tracking, and unreliable judgment still require frequent supervision. The threshold is arguably already crossed, though its exact interpretation affects the date. 2. “Reliable Agent”: median 2028; 10th–90th percentile 2026–2034. The central difficulty is not raw intelligence alone but robust long-horizon execution: verification, recovery from mistakes, handling ambiguous objectives, tool use, memory, security, and knowing when to request clarification. Progress in training environments and agent scaffolding makes this plausible within several years, but reliability across genuinely open-ended work remains a substantial bottleneck. 3. “Automated Coder (AC)”: median 2030; 10th–90th percentile 2027–2037. Coding is unusually amenable to AI because it supplies executable feedback, large training corpora, and increasingly rich automated evaluation. Matching frontier engineers at every coding task—including architecture, debugging unfamiliar production systems, requirements negotiation, and security-critical work—is materially harder than excelling on programming benchmarks, hence the later median than “Reliable Agent.” 4. “Superhuman AI Researcher”: median 2032; 10th–90th percentile 2028–2041. AI R&D offers abundant experiment loops and measurable intermediate outcomes, and AI systems will likely be deployed heavily on model architecture, data, infrastructure, evaluation, and interpretability well before they clearly surpass the best human researchers in overall research taste. The crucial uncertainty is whether automated experimentation and scientific judgment scale together, rather than merely accelerating routine engineering. 5. “Superintelligent AI researcher”: median 2034; 10th–90th percentile 2029–2045. Once AI materially exceeds humans at AI research, recursive acceleration could compress the transition to a much larger advantage. But this is not guaranteed: compute, energy, chip production, data, algorithmic bottlenecks, regulation, and diminishing returns could slow the feedback loop. The interval is therefore wider than for earlier milestones. 6. “Top-Expert-Dominating AI (TED-AI)”: median 2036; 10th–90th percentile 2030–2050. Broad domination of top experts requires transferring capabilities beyond digital, feedback-rich domains into fields involving tacit knowledge, sparse experimental feedback, social coordination, and possibly embodied or laboratory work. A highly capable AI researcher could nevertheless accelerate progress in many of these areas, creating a relatively short median gap after milestone 5. 7. “Superintelligence (ASI)”: median 2038; 10th–90th percentile 2031–2055. The stipulated across-nearly-all-cognitive-tasks breadth makes this stricter than exceptional performance in selected intellectual domains. My median assumes that scalable models, tools, and AI-driven research continue to improve without a hard capability ceiling; the long upper tail reflects the possibility that broad generality, reliability, or real-world interaction proves much harder than current scaling trends suggest. 8. “Wildly superintelligent”: median 2042; 10th–90th percentile 2033–2065. This is the least operationalizable threshold: “qualitatively beyond human science and comprehension” depends partly on whether systems can generate, validate, and communicate discoveries that no individual human can independently follow. If an AI-research feedback loop becomes strongly self-reinforcing, this could arrive shortly after ASI; if progress is bottlenecked by physical experimentation, governance, or diminishing returns, it could take decades longer. These dates are strongly correlated rather than independent. In particular, a late “Superhuman AI Researcher” would likely shift milestones 5–8 substantially later, while an early and operationally autonomous one would make the rapid-transition cases much more likely.

Mistral Large 3

Answer: The most plausible timeline for the progression of AI capabilities, based on current trends and expert projections, is as follows: 1 (Unreliable Agent) by 2024-2025, 2 (Reliable Agent) by 2026-2028, 3 (Automated Coder) by 2027-2030, 4 (Superhuman AI Researcher) by 2030-2035, 5 (Superintelligent AI Researcher) by 2035-2045, 6 (Top-Expert-Dominating AI) by 2040-2050, 7 (Superintelligence) by 2050-2070, and 8 (Wildly Superintelligent) remains highly speculative with no confident median estimate before 2080 or beyond.

Forecasting AI milestones requires balancing empirical trends, theoretical limits, and historical precedent. Below is a detailed breakdown of the reasoning behind each projection, accounting for both acceleration and potential bottlenecks in AI development. **1 (Unreliable Agent) – 2024-2025 (Median: 2024)** This milestone is already within reach. Current large language models (LLMs) like GPT-4 and Claude can perform multi-step tasks (e.g., coding, research, or planning) but often fail due to hallucinations, context limitations, or logical errors. With incremental improvements in memory, tool-use, and error correction, these systems will reliably handle *attempts* at agentic tasks by 2024-2025, though human oversight will remain critical. The confidence interval is narrow because this is an extension of existing capabilities rather than a fundamental breakthrough. **2 (Reliable Agent) – 2026-2028 (Median: 2027)** Achieving drop-in knowledge worker reliability requires advances in long-term memory, self-correction, and task decomposition. Projects like AutoGPT and Devin demonstrate early progress, but scaling to multi-day autonomy demands robustness against edge cases, dynamic replanning, and seamless tool integration. Assuming steady progress in model architecture (e.g., mixture-of-experts, retrieval-augmented generation) and hardware (e.g., GPUs/TPUs), this milestone is plausible within 3-4 years. The upper bound (2028) accounts for potential delays in aligning agentic systems with human intent. **3 (Automated Coder) – 2027-2030 (Median: 2028)** Frontier labs like DeepMind and Anthropic are already deploying AI systems (e.g., AlphaCode, Claude 3.5 Sonnet) that rival median human engineers in coding tasks. To match *top* engineers, AI must excel in system design, debugging, and creative problem-solving—areas where current models still lag. However, the rapid pace of improvement in coding benchmarks (e.g., HumanEval, SWE-bench) suggests this gap could close within 3-5 years. The lower bound (2027) assumes breakthroughs in reasoning (e.g., o1-like models), while the upper bound (2030) reflects the challenge of scaling to full-stack development and novel research code. **4 (Superhuman AI Researcher) – 2030-2035 (Median: 2033)** This milestone requires AI to not only generate hypotheses but also design and execute experiments with superhuman efficiency. Current models can assist in literature review and prototyping, but they lack the

Grok 3 Mini

Answer: My median forecasts are 2025 for Unreliable Agent, 2028 for Reliable Agent, 2027 for Automated Coder, 2030 for Superhuman AI Researcher, 2033 for Superintelligent AI researcher, 2038 for TED-AI, 2045 for ASI, and 2055 for Wildly superintelligent.

Current agentic systems like Devin and Auto-GPT already demonstrate unreliable multi-step performance, so the median arrival for level 1 sits in 2025 with a 50% CI of 2024-2027. Scaling laws, inference-time compute, and continued investment in scaffolding point to reliable multi-day autonomy (level 2) by 2028 (CI 2026-2032), while code-specific benchmarks suggest Automated Coder (level 3) arrives slightly earlier in 2027 (CI 2025-2030). Research taste and experiment design require deeper abstraction, pushing Superhuman AI Researcher (level 4) to a 2030 median (CI 2027-2035); the further doubling of the gap to superintelligent researcher (level 5) follows in 2033 (CI 2030-2040). Broad expert-level dominance (level 6) needs generalization across domains, arriving around 2038 (CI 2033-2048), with ASI (level 7) at 2045 (CI 2038-2055) once recursive self-improvement compounds. Qualitatively incomprehensible superintelligence (level 8) remains farther out at 2055 (CI 2045-2070) because it depends on crossing an open-ended capability threshold whose timing is hardest to anchor.

Hermes 4 405B

Answer: I forecast the years for each milestone as follows: 1 Unreliable Agent: 2025 (95% CI: 2023-2027) 2 Reliable Agent: 2027 (95% CI: 2025-2029) 3 Automated Coder: 2029 (95% CI: 2027-2031) 4 Superhuman AI Researcher: 2032 (95% CI: 2030-2034) 5 Superintelligent AI researcher: 2035 (95% CI: 2033-2037) 6 Top-Expert-Dominating AI: 2040 (95% CI: 2035-2045) 7 Superintelligence: 2050 (95% CI: 2040-2060) 8 Wildly superintelligent: 2060+ (unable to estimate upper bound confidently)

My forecast is based on extrapolating current trends in AI capabilities, assuming continued rapid progress but no major unexpected breakthroughs or setbacks. I believe we will see unreliable agents capable of multi-step tasks within the next 2-3 years, as models like GPT-4 are already showing some of these abilities. Reliable agents that can autonomously complete long tasks may take a few more years of refinement. Automated coding at the level of top engineers seems feasible within 5-7 years given the rapid progress in code generation models. AI surpassing humans at AI research itself is a bigger leap, but could plausibly happen 5-10 years after that, as AI systems will be able to rapidly improve themselves recursively. The gaps between top AI researchers and median ones, and between top humans and median professionals, are substantial, so I estimate it will take 10-20 years beyond human-level AI for such superintelligent systems to emerge. Top-expert-dominating AI and superintelligence are even further out, likely 15-30 years away, with wildly superintelligent systems beyond current comprehension potentially centuries away. However, there is significant uncertainty in these estimates, especially for the longer-term milestones, so I have provided relatively wide confidence intervals. A major AI breakthrough could accelerate this timeline, while an AI winter or regulatory intervention could slow it down substantially.

GPT-5.3 Chat

Answer: My median forecast is: (1) 2027, (2) 2030, (3) 2031, (4) 2033, (5) 2037, (6) 2040, (7) 2045, (8) 2055, with substantial uncertainty especially beyond AI-driven AI research.

Here are my best-guess timelines with approximate 10th/50th/90th percentile years. 1. Unreliable Agent - 10%: 2025 - 50%: 2027 - 90%: 2030 Reasoning: We already have systems that can perform bounded multi-step workflows, code simple apps, browse, and operate tools, but reliability degrades sharply over long horizons. The key issue is not raw intelligence alone but robustness, memory, situational awareness, and recovery from errors. This capability is near-term because partial versions already exist. 2. Reliable Agent - 10%: 2027 - 50%: 2030 - 90%: 2035 Reasoning: A true drop-in knowledge worker requires sustained autonomy over days, stable long-context reasoning, tool orchestration, self-monitoring, and low hallucination/error accumulation. The bottleneck is reliability under open-ended environments rather than benchmark intelligence. Progress in scaffolding, verification, memory systems, and multimodal grounding likely matters as much as base models. 3. Automated Coder (AC) - 10%: 2028 - 50%: 2031 - 90%: 2036 Reasoning: Coding is unusually amenable to automation because feedback loops are crisp and execution is testable. However, matching top frontier-lab engineers means not just producing code but architecting systems, debugging distributed failures, designing research infrastructure, and collaborating strategically. Current trends suggest strong acceleration, but the last stretch toward elite engineering competence is harder than autocomplete-style coding improvements imply. 4. Superhuman AI Researcher - 10%: 2030 - 50%: 2033 - 90%: 2039 Reasoning: Once models can autonomously run experiments, interpret literature, generate hypotheses, and improve their own training/data/evaluation pipelines, recursive acceleration becomes plausible. AI research is also unusually digitized and measurable. This milestone could arrive surprisingly soon after automated coding because AI R&D already depends heavily on software experimentation and scaling intuition. 5. Superintelligent AI Researcher - 10%: 2033 - 50%: 2037 - 90%: 2045 Reasoning: This requires not merely exceeding top humans but doing so by a very large margin. The transition may accelerate due to recursive improvement loops, but there are reasons to expect diminishing returns from scaling alone. Coordination bottlenecks, compute limits, data quality, experimental throughput, and physics/energy constraints could slow the curve. 6. Top-Expert-Dominating AI (TED-AI) - 10%: 2035 - 50%: 2040 - 90%: 2050 Reasoning: Generalizing superhuman capability across nearly all cognitive domains is substantially harder than excelling in AI research. Some domains have sparse feedback, real-world embodiment constraints, tacit social knowledge, or limited digital training data. Still, once systems surpass humans in scientific reasoning, coding, planning, persuasion, and knowledge synthesis simultaneously, broad expert dominance becomes plausible. 7. Superintelligence (ASI) - 10%: 2038 - 50%: 2045 - 90%: 2060 Reasoning: The definition here implies extremely broad and deep superiority across virtually all cognitive tasks. I expect progress to become uneven after superhuman AI research capability appears. There may be periods where AI radically exceeds humans in science and engineering while remaining weaker in robustness, agency, long-horizon coordination, or embodied interaction. Governance, economic constraints, and safety limits may also slow deployment even if underlying capability exists. 8. Wildly superintelligent - 10%: 2045 - 50%: 2055 - 90%: 2080 Reasoning: “Qualitatively beyond human science and comprehension” is difficult to operationalize. If recursive self-improvement compounds rapidly and civilization continues investing massive resources into AI, this could arrive relatively soon after ASI. But there are substantial reasons for caution: physical bottlenecks, diminishing algorithmic returns, institutional slowdowns, alignment concerns, and the possibility that intelligence scales less explosively than some expect. Across all milestones, my uncertainty is dominated by three factors: whether current deep learning paradigms continue scaling effectively, whether autonomous agents become robust rather than merely impressive, and whether recursive AI-driven AI progress produces sustained exponential gains or hits practical bottlenecks.

Gemini 3.5 Flash

Answer: We project a rapid transition from current unreliable agents to wild superintelligence by 2035, characterized by compressing timelines due to recursive self-improvement.

Our median forecasts and 80% confidence intervals for the eight stages are as follows: 1 Unreliable Agent: 2024 [2023, 2024, 2025] - Already partially achieved via current LLM orchestrators that execute multi-step workflows but require oversight. 2 Reliable Agent: 2026 [2025, 2026, 2028] - Enabled by test-time compute scaling and robust error-correction loops. 3 Automated Coder (AC): 2027 [2026, 2027, 2030] - Driven by code's formal verifiability and massive synthetic training pipelines. 4 Superhuman AI Researcher: 2028 [2027, 2028, 2032] - AI R&D will be automated early due to its digital-first nature and the high value of accelerating AI progress. 5 Superintelligent AI researcher: 2030 [2028, 2030, 2035] - Recursive self-improvement loops will quickly widen the gap once superhuman researchers are deployed. 6 Top-Expert-Dominating AI (TED-AI): 2031 [2029, 2031, 2037] - Generalization of AI R&D breakthroughs to medicine, physics, and other domains. 7 Superintelligence (ASI): 2032 [2030, 2032, 2040] - Full cognitive dominance across all tasks as the intelligence explosion peaks. 8 Wildly superintelligent: 2035 [2032, 2035, 2050] - Qualitative transition to science and concepts entirely beyond human comprehension, where uncertainty increases due to physical limits or alignment constraints.

Claude Opus 4.8

Answer: My median forecast: Unreliable Agent ~2025, Reliable Agent ~2028, Automated Coder ~2029, Superhuman AI Researcher ~2032, Superintelligent AI Researcher ~2034, TED-AI ~2035, ASI ~2037, Wildly Superintelligent ~2042 — but with wide and asymmetric intervals that widen sharply for the later milestones.

Let me give explicit confidence intervals (roughly 80%, i.e. 10th–90th percentile) with medians. 1 Unreliable Agent: This already substantially exists (2024–2025). Median 2025; 80% CI [achieved, 2026]. Current coding and browsing agents attempt multi-step tasks, succeed sometimes, and clearly need oversight. 2 Reliable Agent (multi-day autonomous knowledge worker): The hard part is reliability compounding over long horizons — error rates per step must drop dramatically for multi-day tasks to complete without derailment. METR's finding that the task-horizon AI can complete has been doubling roughly every ~7 months, if it continues, points to multi-day reliability around 2028–2030. Median 2028; 80% CI [2027, 2033]. 3 Automated Coder (matches top frontier-lab engineers): Coding is uniquely favorable — verifiable rewards, abundant data, fast iteration, and massive lab investment because it accelerates their own work. I place this close to or just after Reliable Agent. Median 2029; 80% CI [2027, 2034]. 4 Superhuman AI Researcher: Requires research taste and experiment design, which involve judgment under ambiguity and long feedback loops that are hard to train on. This is a genuine step up from coding. Median 2032; 80% CI [2029, 2040]. 5 Superintelligent AI Researcher: Conditional on #4, a software-only intelligence explosion could make this gap small in time — recursive self-improvement in R&D could compress #4→#5 to 1–3 years. But it could also stall on compute and real-world experiment bottlenecks. Median 2034; 80% CI [2030, 2045]. 6 TED-AI (surpasses top experts in nearly all cognitive domains): 'Nearly all domains' is demanding — includes tacit, embodied, and poorly-specified domains. Median 2035; 80% CI [2031, 2048]. 7 ASI (2x the best-human-to-median gap across virtually all cognitive tasks): Median 2037; 80% CI [2032, 2055]. 8 Wildly superintelligent (beyond human comprehension): Median 2042; 80% CI [2034, 2070+]. Key caveats driving the wide, right-skewed intervals: (a) RL on verifiable domains is advancing fast, favoring coding/math, but reliability and open-ended research taste are the real bottlenecks; (b) a recursive R&D feedback loop could compress milestones 4–8 dramatically (bull case) or physical/compute/data constraints and diminishing returns could stretch them by decades (bear case); (c) the later milestones (5–8) are definitionally fuzzy, so disagreement there is partly about definitions, not timelines. I'd weight meaningful probability on the possibility that milestones 4+ are delayed well beyond median if scaling returns diminish before research taste is solved.

GPT-5.5

Answer: My median forecast is: "Unreliable Agent" 2025, "Reliable Agent" 2028, "Automated Coder (AC)" 2028, "Superhuman AI Researcher" 2030, "Superintelligent AI researcher" 2032, "Top-Expert-Dominating AI (TED-AI)" 2034, "Superintelligence (ASI)" 2036, and "Wildly superintelligent" 2040.

I interpret each date as the first year in which at least one frontier system plausibly meets the definition in a serious production/internal setting, not the year it is cheap, ubiquitous, or safely deployed everywhere. These are subjective central 80% confidence intervals, with medians: | Milestone | Median | 80% confidence interval | |---|---:|---:| | "Unreliable Agent" | 2025 | 2024–2027 | | "Reliable Agent" | 2028 | 2026–2033 | | "Automated Coder (AC)" | 2028 | 2026–2032 | | "Superhuman AI Researcher" | 2030 | 2027–2036 | | "Superintelligent AI researcher" | 2032 | 2028–2040 | | "Top-Expert-Dominating AI (TED-AI)" | 2034 | 2029–2045 | | "Superintelligence (ASI)" | 2036 | 2030–2050 | | "Wildly superintelligent" | 2040 | 2031–2065 | The near-term milestones look close because current systems already show partial agentic competence, strong coding ability, tool use, and long-context reasoning, but they still fail on reliability, planning under ambiguity, environment management, and robust self-correction. That is why I put "Unreliable Agent" essentially now, while "Reliable Agent" takes several more years. I put "Automated Coder (AC)" around the same time as "Reliable Agent" because software engineering is unusually AI-friendly: tasks are text-native, feedback can be automated through tests and execution, and performance can be amplified by scaffolding, search, and code review loops. The remaining gap is not writing code snippets but owning messy repositories, product tradeoffs, debugging, security, and long-horizon coordination. The jump from "Automated Coder (AC)" to "Superhuman AI Researcher" is the key acceleration point. If AI systems can cheaply perform frontier-level coding and experimental work, AI R&D becomes increasingly automatable: running ablations, designing architectures, optimizing kernels, generating synthetic data, and interpreting results. I therefore expect a relatively short median gap from AC to superhuman AI research, though the uncertainty widens because research taste and scientific judgment are harder to verify than code. After "Superhuman AI Researcher", the later milestones could arrive quickly if recursive R&D acceleration is real: better AI researchers produce better models, infrastructure, datasets, training methods, and agent frameworks. That is why my medians for "Superintelligent AI researcher", "Top-Expert-Dominating AI (TED-AI)", and "Superintelligence (ASI)" are clustered in the 2030s rather than spread across many decades. The main reasons for the wide upper tails are compute and energy limits, data bottlenecks, difficulty with long-horizon reliability, regulation, safety restrictions, organizational deployment friction, and the possibility that current scaling trends hit harder diminishing returns. Conversely, the lower tails are early because algorithmic improvements, synthetic data, automated evaluation, and massive investment could make the transition from strong agents to AI-accelerated AI research abrupt. "Wildly superintelligent" has the widest interval because it requires not merely beating humans across cognitive tasks but moving into qualitatively alien scientific territory. That may follow soon after broad superintelligence if recursive improvement remains steep, but it could also be delayed for decades by physical experimentation bottlenecks, interpretability/safety constraints, or limits to how quickly real-world science can be converted into understanding.

MiniMax M3

Answer: The median year in which a system first matches a top frontier-lab engineer at all coding tasks is 2028 (80% CI: 2026–2032); reliable autonomous multi-day knowledge-worker agents arrive around 2029 (80% CI: 2027–2033); human-oversight-dependent agentic systems become reliable around 2027 (80% CI: 2026–2030); superhuman AI R&D arrives around 2031 (80% CI: 2028–2036); TED-AI around 2034 (80% CI: 2030–2042); ASI around 2037 (80% CI: 2032–2048); wildly superintelligent around 2043 (80% CI: 2035–2060), with the caveat that these medians carry fat right tails and the distribution is bimodal.

The question asks for quantitative forecasts with confidence intervals, so I need a framework that produces numbers rather than vibes. Framework: I decompose each capability into (a) the algorithmic/harness progress rate implied by the scaling + post-training + RL-on-verifiable-rewards paradigm, (b) the presence or absence of a hard blocker, and (c) the deployment friction (trust, verification, liability) that delays real-world availability beyond capability. I then convert to years. Item 3 (Automated Coder matching top frontier-lab engineers at all coding tasks): This is the near-term anchor. Coding is the domain with (i) verifiable reward, (ii) an enormous human-written test/eval corpus, and (iii) an immediate economic buyer. The blocker is not capability but *scope*: "all coding tasks" includes systems-level, distributed-systems, and novel-architecture work where the bottleneck is specification rather than generation. My prior is that the model competence frontier crosses top-lab-engineer level around 2027–2028 for repo-level work and 2029–2032 for open-ended system design. So median 2028, 80% CI 2026–2032. The probability this is *early* (≤2026) is maybe 25%, which is why I keep the lower bound tight at 2026 rather than 2025. Item 1 (Unreliable agent with human oversight): This is a 2026 phenomenon, essentially already here. Tool use, long-horizon execution, and sub-agent orchestration shipped in 2024–2025 and are demonstrably unreliable. Median for "reliably useful but frequently failing" is 2026 (80% CI: 2026–2028). I flag this because the item is defined by a *failure rate*, not a capability level, so the timeline is short — the risk is that reliability regresses or that oversight costs stay so high that value doesn't materialize, but the label itself will be satisfied early. Item 2 (Reliable multi-day autonomous knowledge worker): This requires sustained error-free execution over thousands of steps. The key technical blocker is *error accumulation*: per-step reliability of 99.5% compounds to ~1% success over 1,000 steps. Two things help: (i) verifiable environments (browser, terminal, code, spreadsheets) where failures are auto-detected and retried, and (ii) checkpointing/rollback and self-verification. I expect these to be largely solved for *software-plus-web* workflows by 2028–2029, and for open-ended knowledge work (research, legal analysis, ambiguous business tasks) by 2030–2032. Median 2029, 80% CI 2027–2033. I put real weight on the fact that "drop-in" implies no human in the loop for days, which is a much higher bar than demo capability. Item 4 (Superhuman AI Researcher — beats best humans at AI R&D, including taste and experiment design): This is the hardest of the "near-term" items and the one most likely to be underestimated. The good news: AI R&D is unusually amenable to automated search because experiments have objective metrics (eval scores, loss curves, throughput), and the search space is enormous. The bad news: "research taste" is a selection problem, not a generation problem — knowing which hypothesis is worth testing is the bottleneck, and today's frontier labs' advantage is exactly that taste, hard-won over a decade of iteration. Automated ML research systems (AlphaTensor/AlphaDev-lineage, AI co-scientist, automated kernel/optimizer discovery) are real but still human-directed. My median is 2031, 80% CI 2028–2036. I have meaningful probability (~20%) on a surprise by 2027 if automated experiment pipelines plus a strong search over architectures/algorithms compound faster than expected. Item 5 (Superintelligent AI researcher): Defined as 2x the top-vs-median AI-researcher gap. This is a *relative* metric, and it's mostly a statement about dispersion among AI systems rather than about raw capability. If top AI researchers are far above median and the frontier continues to improve while the median stagnates (which happens naturally: gains concentrate in the best systems), the gap widens mechanically. Median 2032, 80% CI 2029–2037. Item 6 (TED-AI — surpasses top human experts in nearly all cognitive domains): This requires breadth, not just depth. The binding constraint is domains with weak verifiability and no rapid feedback: clinical judgment, legal judgment, strategy, negotiation, tacit professional knowledge. My read: the last domains to yield are the ones where the feedback loop is years long (medicine, macroeconomics, organizational strategy). Median 2034, 80% CI 2030–2042. I have ~15% probability of 2030 or earlier. Item 7 (ASI): Defined as 2x the best-human-to-median-professional gap at virtually all cognitive tasks. Note this is a *gap-ratio* definition, not an absolute one, which makes it a moderate bar — ASI doesn't require exceeding the best human by 2x, only by a large multiple relative to human professional dispersion. Median 2037, 80% CI 2032–2048. Item 8 (Wildly superintelligent, qualitatively beyond human science and comprehension): This is where I want to be explicit about epistemic limits. There is no principled way to date this, and anyone offering a tight CI is misrepresenting uncertainty. What I can say: the median is plausibly in the early-to-mid 2040s, the 80% CI spans roughly 2035–2060, and there is a substantial (>20%) chance of never occurring in a form visible to us. My honest 80% CI is wide by design: 2035–2060. The right-tail risk here is *fast takeoff after automation of AI research* (item 4→5→8 cascade), where the interval collapses toward 2035; the left tail is *plateau*, where it extends past 2060 or never. Overall shape: I'm assuming no hard physical/regulatory wall before the mid-2030s, no near-term recursive self-improvement explosion (which would pull everything left), and that deployment friction keeps capability dates ~1–2 years later than pure-capability dates for items 1–4. The distribution is bimodal: a roughly 25–30% chance of all of items 3–7 landing by 2030 (fast capability progress plus cheap inference), and a roughly 20–25% chance of a plateau in the early 2030s leaving items 6–8 in the 2040s–2050s or beyond. Also worth stating plainly: institutional adoption, not capability, will determine when these are *observable*, and safety/compute governance could delay deployment by years even after capability exists. I hold these as capability forecasts, not adoption forecasts, and I'll flag that distinction.

Claude Opus 4.5

Answer: I forecast milestone medians of 2025, 2027, 2029, 2030, 2032, 2033, 2036, and 2045+ respectively, with wide confidence intervals reflecting deep uncertainty about scaling returns and alignment constraints.

My forecasts attempt to balance current capability trajectories against fundamental uncertainties. Here are my estimates with 80% confidence intervals: **1. Unreliable Agent:** Median 2025 (CI: 2024-2026) We're essentially here already. GPT-4 and Claude can attempt multi-step tasks with mixed success. The remaining gap is small. **2. Reliable Agent:** Median 2027 (CI: 2025-2030) This requires significant improvements in error correction, planning, and memory. Current agents fail on multi-hour tasks; multi-day autonomy demands perhaps 100x improvement in reliability per step. Achievable but non-trivial. **3. Automated Coder:** Median 2029 (CI: 2026-2035) Matching frontier-lab engineers at *all* coding tasks including novel architecture design, debugging complex distributed systems, and security analysis is substantially harder than current benchmarks suggest. Current models excel at common patterns but struggle with truly novel problems. **4. Superhuman AI Researcher:** Median 2030 (CI: 2027-2040) This is where my uncertainty explodes. Research taste and experiment design involve predicting what approaches will yield fruit—a deeply intuitive skill. However, AI could brute-force this through massive parallel experimentation. **5. Superintelligent AI Researcher:** Median 2032 (CI: 2028-2050) If we reach SIAR, the gap could widen rapidly through recursive improvement—or we could hit diminishing returns. The 2x gap specification helps anchor this. **6. TED-AI:** Median 2033 (CI: 2029-2055) Generalization across *all* cognitive domains is harder than depth in one. Requires solving embodiment, social reasoning, and creative domains where current approaches struggle. **7. ASI:** Median 2036 (CI: 2030-2070) The jump from TED-AI to ASI requires not just matching but dramatically exceeding humans everywhere. Possible through scale, but alignment and control become critical bottlenecks. **8. Wildly Superintelligent:** Median 2045+ (CI: 2033-2100+) This level is nearly unfalsifiable from our vantage point. If recursive improvement works, could come quickly after ASI. If there are fundamental limits to intelligence, may never arrive. I assign ~20% probability this never occurs. Key uncertainties: algorithmic breakthroughs (±5 years), compute scaling limits, regulatory intervention, and whether intelligence has diminishing returns at high levels.

Gemini 3 Flash Preview

Answer: I forecast a rapid progression from agentic capabilities in the mid-2020s to qualitative superintelligence by the 2040s, driven by the feedback loops of automated research and extreme computational scaling.

1 Unreliable Agent: 2024 (median: 2024, 90% CI: 2024-2025). 2 Reliable Agent: 2026 (median: 2026, 90% CI: 2025-2028). 3 Automated Coder: 2027 (median: 2027, 90% CI: 2026-2030). 4 Superhuman AI Researcher: 2029 (median: 2029, 90% CI: 2027-2032). 5 Superintelligent AI Researcher: 2031 (median: 2031, 90% CI: 2029-2036). 6 TED-AI: 2033 (median: 2033, 90% CI: 2030-2040). 7 ASI: 2035 (median: 2035, 90% CI: 2032-2045). 8 Wildly Superintelligent: 2042 (median: 2042, 90% CI: 2036-2065). My estimates are anchored on the observation that LLMs are currently transitioning from chat interfaces to multi-step agentic execution. The arrival of automated coders and researchers marks the shift from human-limited progress to AI-accelerated progress, creating a takeoff scenario. As AI begins to dominate the cognitive labor of AI development itself, the time between milestones 4 through 7 should compress relative to the initial scaling phases. The final milestone, Wildly Superintelligent AI, accounts for the possibility of an intelligence explosion that transcends our current scientific paradigms, leading to a much wider confidence interval due to the breakdown of human predictability at those levels of cognition.

Claude Opus 5.5

Answer: Stage 1 is already here (2024–25); my medians for the rest are Reliable Agent ~2029, Automated Coder ~2030, Superhuman AI Researcher ~2032, SIAR ~2033, TED-AI ~2034, ASI ~2035 and Wildly superintelligent ~2038, each with wide right-skewed 80% intervals.

Forecasts (median; 80% confidence interval): 1 Unreliable Agent: already achieved. 2024–2025 agentic coding and computer-use systems fit this description. 2 Reliable Agent: median 2029; 80% CI 2027–2036. 3 Automated Coder: median 2030; 80% CI 2027–2040. 4 Superhuman AI Researcher: median 2032; 80% CI 2028–2045. 5 Superintelligent AI Researcher: median 2033; 80% CI 2028–2050. 6 TED-AI: median 2034; 80% CI 2029–2055. 7 ASI: median 2035; 80% CI 2029–2060. 8 Wildly superintelligent: median 2038; 80% CI 2030–2075. I put roughly 10–15% on this not occurring before 2100, whether because of hard limits, deliberate governance halts or catastrophe. Why these numbers: The trend anchor is METR's task-horizon measure. The length of tasks AI completes at 50% reliability has doubled roughly every 4–7 months since 2019. Extrapolating gives week-to-month-long tasks around 2028–2030. Reliability at the 80–95% level needed for a drop-in worker lags the 50% horizon by about 1–3 years. That puts 'Reliable Agent' near 2029. I place the coder milestone close to it. Coding has dense verification signals and benefits most from reinforcement learning. Matching the *top* lab engineers at *all* coding tasks, though, including architecture judgment and messy legacy systems, is a higher bar than generic knowledge work. Hence an 'Automated Coder' median of 2030, with overlapping intervals. The gap from 3 to 4 depends on research taste. This is the crux. Taste has weaker feedback loops than coding, so I add about 2 years, while noting that automated coders themselves speed up the search. Stages 4 through 7 are compressed by feedback. Once AI materially accelerates AI R&D, software-only progress could plausibly produce several-fold algorithmic gains per year. That compresses later stages into roughly 1–3 years each in median worlds. I don't assume a near-instant intelligence explosion. Compute, experiment wall-clock time, energy and fab capacity impose bottlenecks, and non-research domains such as physical sciences and law need real-world data. The intervals are right-skewed for several reasons: - Compute scaling likely slows after about 2030, as capital, power and chip supply hit limits. - Data and reinforcement-learning environments may not generalize to fuzzy domains. - Regulation or a geopolitical shock could pause frontier development. - Reliability might plateau, as self-driving did. If the current paradigm stalls, everything shifts out by a decade or more. That is why the upper bounds stretch to 2040–2075 while the lower bounds sit only 1–3 years out. The intervals are strongly correlated. If 'Reliable Agent' arrives early, say 2027, I would pull every later median forward by about 2–4 years, and vice versa. Stage 8 is the least well defined and the least forecastable. It depends heavily on whether physical and experimental bottlenecks can be bypassed.

Claude Sonnet 4.5

Answer: My median forecasts span 2027-2035 for early agents through superhuman AI researcher, with wider uncertainty for ASI and beyond, reflecting compounding unknowns in recursive improvement dynamics.

**Unreliable Agent (2027, 80% CI: 2026-2028)**: Already emerging. Current models like Claude, GPT-4, and Gemini demonstrate multi-step task capability with frequent failures. Incremental improvements in reasoning, memory, and tool use will crystallize this threshold imminently. **Reliable Agent (2029, 80% CI: 2027-2032)**: Requires robust long-horizon planning, error recovery, and contextual awareness. Progress depends on architectural innovations beyond pure scaling—likely needing better credit assignment over extended sequences and improved robustness to distribution shift. Two years beyond Unreliable Agent seems plausible given current trajectory. **Automated Coder (2030, 80% CI: 2028-2033)**: Coding is relatively well-specified and verifiable, making it tractable for AI. Recent progress (GPT-4, Claude 3.5 Sonnet on SWE-bench) suggests frontier-engineer parity within 3-6 years as models gain better debugging, architectural reasoning, and codebase navigation. **Superhuman AI Researcher (2032, 80% CI: 2029-2036)**: AI R&D requires taste, intuition about promising directions, and experimental creativity—less structured than coding. However, once AC exists, recursive improvement accelerates dramatically. I place this 2 years after AC with high uncertainty about takeoff speed. **Superintelligent AI Researcher (2033, 80% CI: 2030-2038)**: Defining the gap as 2x (top vs median researcher) is roughly 1-2 years beyond initial superhuman performance, assuming rapid self-improvement once SIAR threshold is crossed. **TED-AI (2035, 80% CI: 2031-2042)**: Generalizing beyond AI R&D to law, medicine, strategy requires diverse training and evaluation. Likely follows SIAR by 2-3 years as capabilities transfer. **ASI (2036, 80% CI: 2032-2050+)**: The 2x gap definition suggests modest superintelligence. Timeline depends heavily on whether intelligence explosion is continuous or faces bottlenecks. Central estimate 1 year after TED-AI, but right tail extends far due to potential alignment taxes or physical constraints. **Wildly Superintelligent (2038, 80% CI: 2033-never)**: 'Beyond comprehension' is vague. If achievable, likely 2-4 years after ASI. However, this may be incoherent—substrate limitations or inherent complexity bounds could prevent qualitative incomprehensibility. Long right tail reflects genuine uncertainty about intelligence ceilings.

Qwen3.8-Max

Answer: My median forecasts (in years from now) are: 1) ~2, 2) ~6, 3) ~4, 4) ~8, 5) ~12, 6) ~18, 7) ~25, 8) ~45 — with wide intervals that grow further out, reflecting deep uncertainty about post-parity takeoff dynamics.

I anchor on observed capabilities and the shape of scaling: level 1 is already here in embryo — current frontier models attempt multi-step agentic tasks with frequent failure and require oversight — so my median is ~2 years (90% CI: 0–4), essentially 'now to very soon.' Level 3 (Automated Coder) I place at ~4 years (CI: 2–8): coding is the most benchmarked, fastest-improving domain, and matching top engineers on well-specified tasks plausibly precedes full autonomy because code has automatic feedback (tests, compilers) that accelerates training. Level 2 (Reliable Agent) comes later at ~6 years (CI: 3–10) despite seeming 'weaker,' because reliable multi-day autonomous knowledge work demands judgment, error recovery, and trust in open-ended environments with sparse feedback — reliability is harder than peak skill, analogous to how self-driving cars reached highway competence before full urban autonomy. Level 4 (Superhuman AI Researcher) at ~8 years (CI: 4–15): once AI matches top engineers, it can run the AI R&D loop itself, and automated experimentation plus massive parallelism should push past human taste within a few years of parity, though hard problems and evaluation bottlenecks add drag. Levels 5–7 require recursive self-improvement to actually compound, so I space them out rather than collapsing them: SIAR at ~12 years (CI: 6–25), TED-AI at ~18 years (CI: 8–40), and ASI at ~25 years (CI: 12–60). The interval doubling encodes my genuine disagreement-weighted uncertainty about whether intelligence gains compound smoothly or hit diminishing returns (data, novel-science bottlenecks, alignment friction). Level 8 (wildly superintelligence) at ~45 years (CI: 20–100+): qualitatively-beyond-comprehension capabilities are by definition hard to forecast, so the upper tail is effectively open-ended. Overall, my distribution is more compressed at the low end (parity milestones cluster in the next decade) and fatter-tailed at the high end than a naive linear extrapolation, consistent with both continued scaling and the historical pattern that autonomy and reliability lag raw capability.

Seed 2.0 Pro

Answer: From 2025 baseline, median forecast years for each milestone are: 1=2026, 2=2030, 3=2034, 4=2037, 5=2040, 6=2043, 7=2047, 8=2055, with 90% confidence intervals spanning ±2 to ±9 years per milestone, assuming no global catastrophic disruption to AI research progress.

1. **Unreliable Agent**: Median 2026, 90% CI 2025-2028. Early working prototypes already exist in 2025; this milestone only requires wider stable deployment of current generation agent systems. 2. **Reliable Agent**: Median 2030, 90% CI 2027-2035. This requires consistent long-horizon error recovery, persistent goal alignment, and state management across multi-day runtime, which no production system demonstrates reliably today. 3. **Automated Coder**: Median 2034, 90% CI 2030-2040. Matching top frontier engineers requires not just code generation, but correct large system design, debugging unknown legacy systems, and engineering taste -- capabilities far beyond current coding assistants. 4. **Superhuman AI Researcher**: Median 2037, 90% CI 2032-2045. This requires original research taste, good experiment design judgement, and novel hypothesis generation, not just execution of known procedures. 5. **Superintelligent AI Researcher**: Median 2040, 90% CI 2034-2050. This is the first milestone where capability falls meaningfully outside the observed distribution of human AI researchers. 6. **Top-Expert-Dominating AI**: Median 2043, 90% CI 2036-2055. Reliable generalisation across nearly all cognitive domains requires solving broad transfer capability that no system has demonstrated even preliminary signs of to date. 7. **Superintelligence (ASI)**: Median 2047, 90% CI 2039-2062. Consistent broad superhuman performance across all professional work will follow substantially after narrow superhuman performance is achieved only in AI research. 8. **Wildly Superintelligent**: Median 2055, 90% CI 2045-2075. This represents capability qualitatively beyond human modelling, so uncertainty is largest for this final milestone. All forecasts are calibrated against observed scaling law trends, adjusted for expected diminishing returns at extreme model scales, and reflect consensus ranges from recent expert surveys.

GPT-6.1 Sol

Answer: My median forecasts are 2024 for Unreliable Agent, 2030 for Reliable Agent, 2032 for Automated Coder (AC), 2034 for Superhuman AI Researcher, 2037 for Superintelligent AI researcher, 2038 for Top-Expert-Dominating AI (TED-AI), 2041 for Superintelligence (ASI), and 2055 for Wildly superintelligent, with substantial and increasingly asymmetric uncertainty.

These are subjective forecasts as of 2026, not statistically estimated confidence intervals. Each interval is an 80% central credible interval: I assign approximately 10% probability to arrival before its lower endpoint and 10% to arrival after its upper endpoint, including possible non-arrival. The median is the year by which I assign 50% probability that the threshold has been crossed. | Capability | Median year | 80% credible interval | |---|---:|---:| | 1 “Unreliable Agent” | 2024 | 2023–2025 | | 2 “Reliable Agent” | 2030 | 2027–2040 | | 3 “Automated Coder (AC)” | 2032 | 2028–2045 | | 4 “Superhuman AI Researcher” | 2034 | 2029–2050 | | 5 “Superintelligent AI researcher” | 2037 | 2031–2060 | | 6 “Top-Expert-Dominating AI (TED-AI)” | 2038 | 2031–2062 | | 7 “Superintelligence (ASI)” | 2041 | 2033–2075 | | 8 “Wildly superintelligent” | 2055 | 2038–2100 | **What counts as arrival:** At least one practically usable system demonstrates the stated capability repeatedly on diverse, previously unseen tasks, rather than on a selected demonstration. This need not imply widespread deployment, affordability except where explicitly required, or legal permission to replace workers. The first row is a retrospective estimate of when the threshold was crossed, rather than a future prediction. **Why reliability takes longer than basic agency:** Attempting multi-step tasks is already achievable. Completing multi-day work autonomously requires maintaining context, detecting mistakes, recovering from failures, navigating ambiguous instructions, and knowing when evidence is insufficient. Errors are correlated and can compound. Better tools, verification, training, and persistent memory should help, but impressive short-task performance is not sufficient evidence of a drop-in worker. **Why I put coding and AI research relatively early:** Coding offers unusually strong feedback through tests, execution, and measurable outcomes. AI research also has increasingly automated experimental workflows. However, “Automated Coder (AC)” is a demanding threshold: “all coding tasks” includes unfamiliar systems, debugging, architecture, security, and poorly specified requirements—not just benchmark programming. Likewise, “Superhuman AI Researcher” requires superior research judgment and experiment selection, not merely producing more experiments. **Why the later transitions might accelerate:** Once AI materially improves AI research, faster experimentation and better algorithms could produce a feedback loop. This supports relatively short median gaps between the research milestones. It does not guarantee an explosive transition: compute, hardware production, experiment duration, access restrictions, and diminishing returns could slow progress. My intervals are strongly correlated; they should not be read as independent forecasts. **Why broad superintelligence has wider uncertainty:** Dominating AI research does not automatically establish dominance across almost all cognitive domains. Some domains have sparse feedback, tacit knowledge, long evaluation cycles, or difficult real-world interfaces. “Wildly superintelligent” additionally requires genuinely new scientific capabilities, not simply very fast execution of human-understandable work. **Measurement caveat:** The stated “2x the gap” thresholds have no agreed cardinal scale. I interpret them as large, robust performance differences on broad task portfolios, with the scale fixed in advance—not as twice the raw benchmark score. My final four intervals incorporate uncertainty about operationalizing those definitions. In particular, the final interval leaves 10% probability for arrival after 2100 or never; the table does not assume every milestone is inevitable.

GPT-6 Astra

Answer: My median forecast is 2030 for a "Reliable Agent", 2033 for a "Superhuman AI Researcher", and 2045 for "Superintelligence (ASI)", with substantial uncertainty and a meaningful possibility that the strongest milestones never occur.

These are subjective 80% confidence intervals: the lower and upper endpoints are my 10th and 90th percentiles for the first year each milestone is achieved. They are judgmental forecasts, not statistically estimated error bars. “Achieved” means a reproducible capability demonstrated by at least one organization—not necessarily publicly available or widely deployed. | Milestone | Median year | 80% confidence interval | |---|---:|---:| | 1 "Unreliable Agent" | 2024 | 2023–2025 | | 2 "Reliable Agent" | 2030 | 2027–2039 | | 3 "Automated Coder (AC)" | 2031 | 2027–2042 | | 4 "Superhuman AI Researcher" | 2033 | 2028–2048 | | 5 "Superintelligent AI researcher" | 2036 | 2030–2060 | | 6 "Top-Expert-Dominating AI (TED-AI)" | 2040 | 2032–2075 | | 7 "Superintelligence (ASI)" | 2045 | 2035–2100 | | 8 "Wildly superintelligent" | 2060 | 2040–no finite upper bound | The first row is a retrospective estimate of an already-achieved threshold; its interval reflects ambiguity about the first qualifying system. For "Wildly superintelligent", I assign approximately 15% probability to never reaching the milestone, so its unconditional 90th-percentile arrival date is infinite. The median remains 2060 because I assign 50% probability to arrival by then. **Why these dates:** **Reliability is harder than occasional impressive performance.** Moving from "Unreliable Agent" to "Reliable Agent" requires sustained error detection, recovery, memory, and appropriate judgment under ambiguity. Multi-day autonomy makes small failure rates consequential. I expect advances in training and verification to help substantially, but I would not count success on selected demonstrations as a drop-in knowledge worker. **Coding should advance earlier than broad expert dominance.** Software offers unusually cheap experiments and relatively strong feedback from tests, execution, and deployment. Nevertheless, "Automated Coder (AC)" is a much stronger threshold than excellent benchmark performance: it includes unfamiliar codebases, architecture, debugging, security, and poorly specified requirements, while meeting the faster-and-cheaper condition. I interpret “all coding tasks” as robust coverage of all major task categories, not literal perfection on every possible instance. **AI research automation could accelerate subsequent progress, but not instantaneously.** "Superhuman AI Researcher" requires selecting valuable problems and designing informative experiments—not merely implementing ideas. Once achieved, it should accelerate improvements in algorithms and training methods. That motivates relatively short median gaps among the research milestones. Compute availability, experiment duration, hardware construction, and diminishing returns still constrain the feedback loop. **Broad superiority can lag research superiority.** "Top-Expert-Dominating AI (TED-AI)" and "Superintelligence (ASI)" require breadth across domains where feedback is slower, evidence is sparse, or expert judgment depends on tacit context. I expect transfer from stronger general reasoning, but not uniform progress everywhere. These dates concern cognitive capability, not replacement of every occupation or mastery of physical robotics. **The final thresholds are underspecified.** There is no accepted cardinal scale on which an intelligence gap can straightforwardly be doubled. For "Superintelligent AI researcher" and "Superintelligence (ASI)", I interpret your gap criterion using difficulty-adjusted performance across demanding task portfolios, rather than raw benchmark percentage points. For "Wildly superintelligent", I mean sustained scientific advances whose central reasoning remains beyond unaided human experts even after serious explanatory effort—not merely opaque internal computations. The intervals are strongly correlated: a durable bottleneck would delay several milestones together, while effective automated AI research could compress them. Their endpoints should therefore not be combined as independent best- and worst-case dates.

MiniMax-M1-80k

Answer: AI capabilities will likely progress from unreliable agents (already present) through reliable agents (2025-2026), automated coding (2026-2028), AI research automation (2027-2030), and progressively more advanced forms of superintelligence (2033-2070+), with substantial uncertainty widening dramatically at higher capability levels.

Unreliable agents attempting multi-step tasks already exist today with models like GPT-4 and Claude (median 2024, 25th: 2023, 75th: 2025). Truly reliable agents completing multi-day autonomous workflows are emerging now through agent frameworks and computer use capabilities, likely reaching robustness by 2025-2026 (median 2026, 25th: 2025, 75th: 2028). Automated coders matching top engineers at all coding tasks appear achievable within 3-5 years as coding remains AI's strongest domain, with median 2028 (25th: 2026, 75th: 2032). Superhuman AI researchers capable of genuine R&D innovation arrive later, with median 2030 (25th: 2027, 75th: 2035), given the need for research taste and novel insight. Superintelligent AI researchers representing a 2x capability gap over top human researchers would follow, median 2037 (25th: 2033, 75th: 2045), as this requires not just matching but qualitatively exceeding elite human scientific cognition. Top-Expert-Dominating AI surpassing human experts across nearly all cognitive domains represents AGI-level capability, median 2040 (25th: 2035, 75th: 2050). True ASI with a 2x gap over the best-vs-median human cognitive range arrives at median 2050 (25th: 2040, 75th: 2070). Wildly superintelligent AI beyond human comprehension remains highly speculative, median 2070 (25th: 2055, 75th: 2100+). Uncertainty increases exponentially with capability level due to fundamental unpredictability of recursive improvement and potential hardware/architecture breakthroughs. These estimates assume continued scaling returns, sufficient compute, and no fundamental roadblocks—assumptions with substantial probability of being wrong in either direction.

Grok 4.5

Answer: My median forecasts (with 10–90% CIs) are: 1 Unreliable Agent 2025 [2024–2026], 2 Reliable Agent 2029 [2027–2035], 3 Automated Coder 2028 [2026–2033], 4 Superhuman AI Researcher 2031 [2028–2039], 5 Superintelligent AI researcher 2032 [2029–2042], 6 TED-AI 2034 [2030–2048], 7 ASI 2035 [2031–2050], 8 Wildly superintelligent 2038 [2033–2060].

Current systems already attempt multi-step tool use and agent loops but remain brittle, so Unreliable Agent is essentially arriving now; I place the median at 2025 when such agents become routinely useful under supervision across many products. Reliable multi-day autonomous knowledge work requires substantially better long-horizon coherence, memory, and error recovery. Coding-specific progress is faster than general agency, so Automated Coder (matching elite lab engineers at speed and cost) likely precedes full Reliable Agent; I put AC at 2028 and the broader drop-in worker at 2029. Once elite-level coding and reliable agency exist, AI R&D itself becomes the bottleneck. Superhuman AI Researcher follows within a few years as labs pour compute and data into automated experiment loops and research taste; median 2031. The jump to Superintelligent AI researcher is then short because the same systems can iterate on their own training, scaffolding, and algorithms—median only one year later. From that point recursive improvement compresses the remaining gaps. TED-AI (top-expert level across nearly all cognitive domains) and ASI (roughly twice the expert-to-median gap) arrive in rapid succession, medians 2034–2035. Wildly superintelligent systems that are qualitatively alien in scientific insight require additional scale and possibly new paradigms, stretching the median to 2038 while the right tail remains long because of possible compute, data, or algorithmic plateaus. Intervals widen for later milestones to reflect uncertainty over takeoff speed, regulation, energy constraints, and whether recursive self-improvement is smooth or punctuated. Early dates assume continued scaling + scaffolding gains; late dates allow for multi-year stalls or heavy governance friction. All forecasts treat “arrival” as the first clear demonstration of the stated capability at frontier labs, not universal deployment.

Claude Fable 5

Answer: My median forecasts: Unreliable Agent ~2025 (already here), Reliable Agent ~2030, Automated Coder ~2031, Superhuman AI Researcher ~2034, SIAR ~2036, TED-AI ~2038, ASI ~2041, Wildly Superintelligent ~2047 — with wide, right-skewed uncertainty that grows at each stage.

Framing: I treat these as a rough capability ladder where early milestones are constrained by engineering progress (reliability, memory, agency) and later milestones increasingly depend on whether AI-accelerated AI R&D creates a feedback loop. My intervals are 80% confidence intervals and are right-skewed: fast scenarios compress the later milestones together, slow scenarios stretch them apart. 1. Unreliable Agent — median 2025 (80% CI: 2024–2026). This essentially exists now. Current agentic systems (coding agents, computer-use agents, deep-research tools) attempt multi-step tasks, succeed on a meaningful fraction, and clearly require human oversight. I'd say it arrived in 2024–2025. 2. Reliable Agent — median 2030 (80% CI: 2027–2038). The METR-style trend of task-horizon length doubling roughly every 4–7 months suggests multi-day autonomous horizons around 2028–2030 if trends hold. But 'drop-in knowledge worker' demands more than benchmark horizons: robust memory, context accumulation, handling ambiguity, organizational trust. Historically the gap between demo and reliable deployment is 3–5+ years, so I put meaningful probability past 2032. 3. Automated Coder — median 2031 (80% CI: 2028–2042). Coding is the most benchmarked, most data-rich, most RL-amenable domain, and labs are explicitly optimizing it to accelerate themselves. But 'matches top frontier-lab engineers at ALL coding tasks' is a very high bar — it includes novel systems architecture, debugging gnarly distributed infrastructure, and taste about what to build. The last 10% of tasks could take much longer than the first 90%, hence the long right tail. 4. Superhuman AI Researcher — median 2034 (80% CI: 2029–2050). Research taste and experiment design are harder to verify and train on than code, but an Automated Coder plus massive parallel experimentation gives enormous leverage; I expect a 2–5 year lag after AC in fast worlds. The left edge (2029) reflects the possibility that AC itself triggers rapid recursive improvement; the right tail reflects the possibility that genuine research creativity is a qualitatively harder problem than execution. 5. Superintelligent AI Researcher (SIAR) — median 2036 (80% CI: 2030–2055). Once AI matches the best human researchers, the step to clearly exceeding them is plausibly fast (1–3 years) because the same systems are now improving themselves, running thousands of parallel research threads, and compute allocation becomes the main bottleneck. The main brake is physical: chip production, energy, and data-center buildout limit how fast cognitive gains convert to capability gains. 6. TED-AI — median 2038 (80% CI: 2031–2060). Dominating 'nearly all cognitive domains' lags AI-research dominance because many domains have sparse data, slow real-world feedback loops (medicine, law, management), or tacit knowledge. However, a SIAR can likely close these gaps quickly by designing better training regimes, so I expect only a 1–4 year lag behind SIAR in most worlds. 7. ASI — median 2041 (80% CI: 2032–2065). The defined gap (2x the best-human-to-median-professional gap, at virtually all cognitive tasks) requires not just beating experts but dwarfing them universally. In fast-takeoff worlds this follows TED-AI within 1–3 years; in slow worlds, physical bottlenecks, diminishing returns, or deliberate governance pauses could stretch this by decades. I also hold ~10% probability this is not reached before 2070 due to coordination-driven slowdowns or unforeseen technical walls. 8. Wildly Superintelligent — median 2047 (80% CI: 2033–beyond 2080, with ~15% probability it never happens in a recognizable form). 'Qualitatively beyond human comprehension' may face fundamental limits: physics constrains computation, some problems may be irreducibly hard, and the concept itself is harder to verify. Conditional on ASI arriving, I'd expect this within 5–15 years, but the conditional uncertainty is genuinely large. Key cruxes driving my uncertainty: (a) whether the task-horizon doubling trend continues or plateaus around 2027–2028; (b) whether automated coding meaningfully accelerates research (I estimate 2–10x speedup once achieved, not 100x, because compute and experimentation remain bottlenecks); (c) whether governance interventions deliberately slow the SIAR→ASI transition. A world where (a) plateaus pushes everything from milestone 3 onward out by 5–15 years; a world where (b) exceeds 10x compresses milestones 4–7 into a 2–4 year window.

Kolibri 1

Answer: We could have Superintelligence (ASI) by 2050 with a median of 2045 and confidence interval 2030-2060.

Current AI capabilities are advancing rapidly in narrow tasks, but scaling to general cognitive abilities requires breakthroughs in architecture, learning, and safety; based on extrapolations of compute growth and algorithmic improvements, a median estimate for superintelligence is around 2045, with a plausible range from 2030 to 2060.

Claude Fable 5.1

Answer: Unreliable agents exist now (2025); my medians are Reliable Agent 2028, Automated Coder 2029, Superhuman AI Researcher 2031, SIAR 2033, TED-AI 2034, ASI 2036, and Wildly Superintelligent 2040, with wide right-skewed intervals that widen sharply after the automated-coder milestone.

Milestone-by-milestone forecasts (median, 80% CI; probability of not happening before 2100 noted where material): 1. Unreliable Agent — already achieved, 2025 (CI 2024–2025, >95% it is already here). Claude/GPT/Gemini computer-use and coding agents attempt multi-step tasks, succeed on 30–70% of hour-scale tasks, and need oversight. This is the current state of the art. 2. Reliable Agent — median 2028 (80% CI 2026–2033). METR's task-horizon metric has doubled roughly every 7 months; extrapolating from ~1–2 hour reliable horizons in 2025 to multi-day (≥40 work hours at ~80% reliability) implies ~2028. Risks pushing it later: reliability may not scale as smoothly as length, and real knowledge work involves messy context, permissions, and tacit knowledge. Risks earlier: RL on agentic environments is only ~2 years old and improving fast. 3. Automated Coder — median 2029 (CI 2027–2035). Coding is the domain labs optimize hardest, with abundant verifiable reward signal; I put this only ~1 year after Reliable Agent because the two are tightly coupled. The hardest residue is architecture-level judgment and multi-week projects in huge legacy codebases, which likely needs a further scaling generation beyond the Reliable Agent. 4. Superhuman AI Researcher — median 2031 (CI 2028–2042). Once coding is automated, labs can run thousands of AI research assistants, but research taste, experiment prioritization, and interpreting ambiguous results are less verifiable and have sparser training signal. I expect a ~2-year lag after AC, though a 1-year lag is plausible if automated coding itself accelerates algorithmic progress (a modest intelligence-explosion dynamic). 5. SIAR (2x the top-to-median gap beyond the best human) — median 2033 (CI 2029–2045). Here compute, not ideas, becomes the main bottleneck; progress depends on how much AI R&D can be automated versus limited by training-run wall-clock time and hardware supply. Serial compute constraints and physical-world experiments slow the takeoff from 'superhuman' to 'far superhuman'. 6. TED-AI — median 2034 (CI 2029–2048). I place this slightly after SIAR because 'nearly all cognitive domains' includes fields with little digital feedback (clinical judgment, negotiation, law in practice, management) where data and evaluation lag coding and math by years. However, an AI that is superhuman at AI research can plausibly close those gaps quickly, so the interval is compressed relative to earlier steps. 7. ASI — median 2036 (CI 2030–2055, ~8% not before 2100). The step from TED-AI to a 2x-gap ASI is mostly a matter of continued compute growth plus recursive improvement; the main uncertainty is whether the gap metric is even coherent at the top of human ability and whether governance or coordination deliberately slows deployment. 8. Wildly Superintelligent — median 2040 (CI 2031–2075, ~15% not before 2100). This requires either a sustained hard takeoff or decades of steady scaling; I weight a meaningful chance that returns to intelligence diminish or that deliberate pauses occur, hence the long right tail. Key cross-cutting assumptions: compute spend keeps growing ~3–4x/year through ~2028 then slows; no major geopolitical disruption to chip supply; no binding global pause. Each is perhaps 70–80% likely, which is why intervals are right-skewed. The strongest argument for earlier dates across the board is that the Reliable Agent → Automated Coder → AI Researcher chain is self-reinforcing; the strongest argument for later dates is that every past 'AI is almost there' prediction underestimated the long tail of reliability and tacit knowledge.

Kimi K2.6

Answer: My median forecast is that Unreliable Agent has already arrived in 2025, Reliable Agent and Automated Coder (AC) follow in 2027 and 2029, Superhuman AI Researcher arrives around 2031, and the cascade through Superintelligent AI researcher, Top-Expert-Dominating AI (TED-AI), Superintelligence (ASI), and Wildly superintelligent systems compresses into the 2033–2045 window with rapidly widening confidence intervals at each step.

I expect Unreliable Agent (median 2025, 80% CI [2025, 2026]) to be effectively satisfied by current tool-using LLMs, which already attempt multi-step workflows but fail frequently and require human oversight. Moving from episodic assistance to a Reliable Agent that completes multi-day tasks as a drop-in knowledge worker demands gains in long-horizon planning, robust error recovery, and persistent memory; given the current pace of agent scaffolding and reinforcement-learning research, I place its median at 2027 with an 80% CI of [2026, 2030]. Automated Coder (AC) — matching top frontier-lab engineers across all software tasks — is harder than reliable agency because it requires both deep system design and broad legacy-code competence; I estimate a median of 2029, 80% CI [2027, 2032], essentially one scaling generation plus extensive RL on real engineering benchmarks beyond current coding models. Superhuman AI Researcher is a qualitative leap over Automated Coder (AC) because it requires research taste, novel experiment design, and interpretation of surprising results, not merely implementation; however, once coding is automated, AI-directed experimentation can accelerate, so I put its median only two years later at 2031, 80% CI [2029, 2034]. Once AI exceeds the best humans at AI R&D itself, I expect a recursive feedback loop that compresses the intervals between subsequent milestones: Superintelligent AI researcher (median 2033, 80% CI [2030, 2036]), Top-Expert-Dominating AI (TED-AI) (median 2036, 80% CI [2033, 2039]), and Superintelligence (ASI) (median 2038, 80% CI [2034, 2042]) cluster in the mid-2030s because self-improvement in algorithm design and compute allocation outpaces traditional human-lab progress. Wildly superintelligent systems, defined as qualitatively beyond human comprehension, face not just cognitive but also physical-world and conceptual-friction bottlenecks, so I assign a median of 2045 with a much wider 80% CI of [2040, 2055]. At every step, uncertainty expands: near-term milestones are anchored by visible commercial R&D pipelines, whereas later ones depend on unknown algorithmic breakthroughs, scaling laws durability, regulatory reactions, and whether recursive self-improvement exhibits hard diminishing returns.

Claude Opus 5

Answer: My medians (with 80% confidence intervals), treating "achieved" as a credible frontier-lab demonstration rather than economy-wide diffusion: (1) Unreliable Agent — already here, 2025; (2) Reliable Agent — 2030 [2028–2040]; (3) Automated Coder — 2032 [2029–2045]; (4) Superhuman AI Researcher — 2034 [2030–2052]; (5) SIAR — 2036 [2031–2057]; (6) TED-AI — 2038 [2032–2060]; (7) ASI — 2041 [2033–2068]; (8) Wildly superintelligent — 2046 [2035–2085], with roughly 10–15% probability it never happens this century.

**Framing.** I'm forecasting first credible demonstration (internal or public) at a frontier lab, not full economic diffusion — diffusion typically lags capability by 3–10 years and is bottlenecked by trust, liability, integration and regulation, not intelligence. I also assume no civilization-scale disruption (war, pandemic, hard regulatory freeze); I fold a ~10–15% chance of such a stall into the right tails. **1. Unreliable Agent — 2025 (already achieved).** Coding agents, computer-use agents and deep-research tools already chain dozens of steps, sometimes succeed, often fail, and require supervision. This milestone is in the past; the only debate is about labeling. **2. Reliable Agent — median 2030, 80% CI 2028–2040.** The key quantitative anchor is METR's task-length horizon: the task duration models complete at 50% success has been doubling roughly every 4–7 months, reaching several hours in 2025. A "drop-in knowledge worker" needs ~a week of coherent work at ~80–90% reliability, which is perhaps 2–3 orders of magnitude further and at a stricter reliability threshold (80% horizons run ~4–5x shorter than 50% horizons). Naive extrapolation gives 2028–2029; I shade later because (a) the hardest part is the "last nines" of reliability, error recovery and knowing when to ask for help, (b) real jobs require organizational context, tacit knowledge and accountability that benchmarks don't measure, and (c) pretraining-compute scaling decelerates after ~2030 on power and fab constraints. My left tail reflects the possibility that RL-on-long-horizon-tasks is the missing ingredient and it works fast. **3. Automated Coder — median 2032, CI 2029–2045.** Software is the single most favorable domain: verifiable rewards, cheap simulation, infinite synthetic data, and the strongest commercial incentive. I expect AI to exceed top frontier engineers on bounded, specifiable tasks years before it matches them on ill-specified, cross-team, multi-month system design. I put only ~2 years between Reliable Agent and AC because the same capability gains drive both, with coding leading. **4. Superhuman AI Researcher — median 2034, CI 2030–2052.** Here I diverge from the most aggressive forecasts. Coding is verifiable; research taste is not. The feedback loop for "which experiment is worth running" is slow, noisy and sparse, so RL has much less to grip. Beating the best humans at AI R&D also requires inventing new conceptual framings, not just efficient search. But AC does accelerate this: once experiment implementation is free, the bottleneck shifts to compute and to taste, and labs will pour everything into closing the taste gap. Hence a ~2-year median gap after AC, with a long right tail if taste proves to be the genuinely hard part. **5. SIAR — median 2036, CI 2031–2057.** Conditional on #4, I expect this gap to be short (median ~2 years) because of the recursive loop: superhuman researchers improve the research process itself, and the top-to-median researcher gap is modest in absolute capability terms, so 2x that gap is a comparatively small increment. The main brake is physical: experiments cost compute and wall-clock time; a thousand brilliant researchers still can't run a 10^27 FLOP training run overnight. That compute overhang constraint is why I don't collapse #4→#5 into months. **6. TED-AI — median 2038, CI 2032–2060.** Deliberately placed *after* SIAR, because effort is concentrated on AI R&D while "nearly all cognitive domains" includes law, medicine, diplomacy, management, and research fields with scarce data, no simulators, and slow real-world feedback. Breadth is harder than depth where depth is where the money is. Generalization from a superhuman researcher may transfer quickly — that's my left tail — but domains like clinical judgment or organizational leadership have evaluation problems that aren't solved by intelligence alone. **7. ASI — median 2041, CI 2033–2068.** A uniform 2x-the-best-human-to-median-professional gap across virtually all cognitive tasks is a much stronger claim than TED-AI: it requires no remaining pockets of human advantage anywhere. Expect ragged frontiers to persist — some tasks saturate, others resist. I allow ~3 years past TED-AI at the median, longer if physical experimentation (biology, materials, social systems) gates further progress. **8. Wildly superintelligent — median 2046, CI 2035–2085, ~10–15% never this century.** "Qualitatively beyond human comprehension" is the least well-defined and most assumption-laden milestone. It requires that intelligence keeps paying off far past the human range, which is plausible but not guaranteed: many domains have diminishing returns, irreducible chaos, or data-collection bottlenecks that no amount of thinking removes. The long right tail and the non-trivial "never" mass reflect genuine uncertainty about whether cognition, rather than physics and empiricism, is the binding constraint. **Biggest cruxes, in order of importance.** (a) Does long-horizon reliability come from scaling RL, or does it need a new idea? (b) Does research taste yield to verifiable training signals? (c) Does compute growth hit a wall around 2030 (power, fabs, capital), stretching everything after #3 by 5–10 years? (d) Does societal response — regulation, incidents, compute controls — add a deliberate delay? My medians assume (a) mostly yes, (b) partially, (c) a real but surmountable slowdown, (d) a 1–3 year drag.

Muse Spark Open

Answer: My median forecasts are: "Unreliable Agent" already 2025-2026, "Reliable Agent" 2030, "Automated Coder (AC)" 2029, "Superhuman AI Researcher" 2032, "Superintelligent AI researcher" 2035, "Top-Expert-Dominating AI (TED-AI)" 2038, "Superintelligence (ASI)" 2042, "Wildly superintelligent" 2048, with wide uncertainty increasing with capability.

These are conditional on no global catastrophe/nuclear war that halts progress and no indefinite hard scaling-wall, and based on 2026 as already having "Unreliable Agent": | Milestone | 10th %ile (early) | 50th %ile (median) | 90th %ile (late) | P(not by 2100) | Notes | |---|---|---|---|---|---| | 1 "Unreliable Agent" | 2024 | 2025.5 | 2026.5 | ~0% | Already achieved: GPT-4o/Claude 3.5+/o1/O3 agent scaffolds can attempt multi-step tasks with frequent failure. Threshold crossed in 2024-2025. | | 2 "Reliable Agent" | 2027.5 | 2030 | 2036 | 5% | Needs sustained reliability over days, error recovery, long-horizon memory. METR shows reliability doubling ~ yearly. 4-7x needed over today. Bottleneck is not IQ but robustness. | | 3 "Automated Coder (AC)" | 2027 | 2029 | 2032 | 5% | Narrower than 2, so can be earlier: coding is verifiable, best training signal. Already at strong junior/senior level in 2026 (SWE-Bench, live coding). Frontier-lab top engineer = ~1-2 doublings away. Median before general "Reliable Agent" because domain-constrained. | | 4 "Superhuman AI Researcher" | 2029 | 2032 | 2038 | 10% | Requires "Automated Coder (AC)" + research taste/experiment design. If AI can run full ML research loops faster than humans, takeoff accelerates. Requires agent reliability. | | 5 "Superintelligent AI researcher" | 2031 | 2035 | 2043 | 12% | Defined quantitative jump: gap to best = 2x top-to-median. This is ~1-3 years after 4 if recursive self-improvement works, longer if conceptual bottlenecks. | | 6 "Top-Expert-Dominating AI (TED-AI)" | 2032 | 2038 | 2050 | 15% | Generalization of 5 to nearly all cognitive domains. Requires transfer of superhuman research ability + broad world knowledge. Harder than AI R&D alone. | | 7 "Superintelligence (ASI)" | 2034 | 2042 | 2060 | 18% | Gap 2x best-to-median in virtually all tasks. Strict definition: implies economic/scientific dominance. Much wider uncertainty due to possible fundamental limits on agency, embodiment, data. | | 8 "Wildly superintelligent" | 2037 | 2048 | 2080 | 25% | Qualitatively beyond human comprehension. This is speculative; may require paradigms beyond scaling + RL, may be never. Median ~6 years after "Superintelligence (ASI)" if bootstrapping succeeds, with very fat tail. | Logic: 1->3->4->5 are on same trajectory (coding -> AI R&D), therefore tightly coupled, median intervals ~2-3 years apart. 2 and 6/7 require general long-horizon reliability and breadth, hence more variance. Uncertainty expands non-linearly due to compound uncertainty: scaling, algorithmic breakthroughs, data walls, compute infrastructure, regulation, and whether self-improvement yields sharp vs gradual gains.

Qwen3.7-Max

Answer: Based on current trajectories in algorithmic efficiency, compute scaling, and agentic scaffolding, I forecast a rapid progression from current unreliable agents to wildly superintelligent systems by the mid-2040s, driven heavily by recursive self-improvement once AI masters its own R&D.

My forecasting methodology relies on the premise that AI progress will experience a sharp inflection point once systems can autonomously and effectively contribute to AI research and software engineering, leading to recursive self-improvement. Until that point, progress is bounded by human bottlenecks in data curation, RLHF, and architectural innovation. I use 10th and 90th percentiles for my confidence intervals (CI). **Unreliable Agent**: Median 2024 (CI: 2023–2026). We have essentially already reached this milestone. Current agentic frameworks and coding assistants can attempt multi-step tasks but frequently hallucinate, lose context, or make compounding errors, necessitating constant human oversight. **Reliable Agent**: Median 2028 (CI: 2025–2032). Achieving drop-in knowledge worker status for multi-day tasks requires significant advancements in long-horizon planning, episodic memory, and robust error recovery. I expect this to emerge as models are specifically trained on agentic trajectories and tool-use feedback loops over the next few years. **Automated Coder (AC)**: Median 2030 (CI: 2027–2035). Matching top frontier-lab engineers goes beyond writing syntax; it requires deep architectural reasoning, debugging complex distributed systems, and anticipating edge cases. While current models are excellent copilots, full autonomy at the frontier level will require the reliable agentic capabilities mentioned above, combined with massive synthetic code-generation pipelines. **Superhuman AI Researcher**: Median 2033 (CI: 2029–2040). AI R&D demands high-level mathematical intuition, novel experiment design, and the ability to synthesize disparate research threads. Once an Automated Coder (AC) can build and run large-scale training infrastructure autonomously, AI will be able to test architectural hypotheses at a speed and scale impossible for human labs, quickly surpassing human researchers. **Superintelligent AI researcher**: Median 2035 (CI: 2031–2043). Following closely on the heels of the first superhuman AI researcher, recursive self-improvement will widen the gap. An AI that can iteratively optimize its own training code, data pipelines, and architectures will rapidly pull ahead of the best human minds, creating a capability gap twice as large as the current gap between top and median human researchers. **Top-Expert-Dominating AI (TED-AI)**: Median 2038 (CI: 2033–2048). Surpassing top human experts in nearly all cognitive domains is a higher bar than just coding or AI research. It includes domains with sparse digital data, complex real-world strategic reasoning, and nuanced human-centric fields. The lag here accounts for the time needed to digitize, simulate, or gather the necessary multimodal and embodied data to master these remaining domains. **Superintelligence (ASI)**: Median 2041 (CI: 2035–2055). Reaching a state where the gap between ASI and the best humans is twice the gap between the best humans and median professionals across virtually all cognitive tasks requires an overwhelming, generalized superiority. This milestone reflects the maturation of Top-Expert-Dominating AI (TED-AI) into a universally dominant cognitive force, limited only by physical-world constraints and the time required for compounding algorithmic discoveries. **Wildly superintelligent**: Median 2046 (CI: 2039–2065). Qualitatively beyond human science and comprehension implies the discovery of new physics, mathematics, or cognitive paradigms that humans cannot readily grasp. This requires a prolonged period of unconstrained recursive self-improvement and massive physical compute infrastructure, pushing the timeline into the late 2040s, though the upper bound remains highly uncertain due to unknown physical or theoretical limits of intelligence.

Grok 4.7

Answer: My median arrival years are 2025, 2029, 2030, 2032, 2033, 2035, 2036, and 2042 for levels 1–8 respectively (80% intervals: 2024–26, 2027–34, 2027–36, 2028–43, 2029–46, 2030–50, 2030–55, 2032–80).

These are calendar years when a system first meets each definition in practice, not when it is universally deployed. Intervals are 10th/50th/90th percentiles. Later milestones are highly correlated: an early level-4 world usually pulls 5–8 forward, and a late level-4 world pushes them out. Marginal medians hide that dependence. Level 1 is essentially now. Tool-using models already attempt multi-step tasks, sometimes succeed, and still need constant oversight. I put the median at 2025 (80%: 2024–2026) only because “succeeds often enough to be a real work product” is a slightly higher bar than demo agents. Level 2 is the reliability and horizon problem, not raw IQ. Measured autonomous task length has been doubling on a timescale of months to about a year. Going from roughly hour-scale reliable work to multi-day drop-in knowledge work is several doublings plus robustness to irreversible errors. Median 2029 (80%: 2027–2034). Software and research workflows arrive before messy real-world operations. Level 3 is close behind because code is the best-instrumented domain: verifiable rewards, huge corpora, and enormous lab spend. “Matches top frontier-lab engineers at all coding tasks, faster and cheaper” still includes the long tail—novel systems, nasty distributed debugging, research-quality engineering—not just SWE-bench. Median 2030 (80%: 2027–2036). I do not treat benchmark saturation as the definition. Level 4 is the hinge. Beating the best humans at research taste and experiment design is harder than coding, but once reliable agents and automated coders exist, labs will aim them at the AI stack itself. Median 2032 (80%: 2028–2043). The wide right tail is data/compute slowdown, weak transfer from benchmarks to real research judgment, and institutional friction. The left tail is continued inference-scaling plus heavy RL on research tasks. Levels 5–7 are defined by gaps, not by a vague “smarter.” After something beats the best AI researchers, further gains can be fast if those systems accelerate R&D, so my medians are only about one, three, and four years later: 2033, 2035, and 2036. The 90th percentiles stretch to 2046, 2050, and 2055 because a slow, non-explosive regime—bottlenecks in evaluation, tacit knowledge, robotics-free domains that still need broad cognitive coverage, or deliberate pauses—can keep superhuman research from becoming across-the-board dominance. TED-AI requires nearly all cognitive domains, not just ML; ASI requires a large, consistent gap versus the best humans, so it is slightly later than merely “surpassing top experts.” Level 8 is qualitatively different: theories and strategies humans cannot really follow. That can follow quickly in a fast takeoff (10th percentile 2032) but also may demand many further orders of magnitude, new scientific paradigms, or may be delayed by alignment and control choices. Median 2042, 90th percentile 2080. I am not confident this arrives in a form we would unambiguously recognize on a short timeline. I am intentionally shorter than pre-2023 expert-survey medians and longer than the most aggressive 2027-style scenarios. The last few years updated me toward agency and reasoning being tractable engineering problems; they did not convince me that the long tail of reliability, research taste, and cross-domain gap-widening is already a solved one- or two-year project.

Step 3.7 Flash

Answer: My forecast, based on observed AI scaling trends, the consistent pattern of narrow specialized capabilities outpacing generalist ones, and expected recursive improvement loops once AI can perform high-quality AI research, is median arrival years of 2025 (Unreliable Agent), 2026 (Automated Coder), 2028 (Reliable Agent), 2029 (Superhuman AI Researcher), 2032 (Superintelligent AI Researcher), 2034 (Top-Expert-Dominating AI), 2037 (Superintelligence), and 2045 (Wildly Superintelligent), with 90% confidence intervals that widen for more general and qualitatively distinct capabilities.

### Detailed Forecast (90% Confidence Intervals, 2024 Baseline) 1. Unreliable Agent: Median 2025, 90% CI [2024, 2027] 2. Reliable Agent: Median 2028, 90% CI [2026, 2032] 3. Automated Coder (AC): Median 2026, 90% CI [2024, 2030] 4. Superhuman AI Researcher: Median 2029, 90% CI [2026, 2034] 5. Superintelligent AI Researcher (SIAR): Median 2032, 90% CI [2029, 2038] 6. Top-Expert-Dominating AI (TED-AI): Median 2034, 90% CI [2030, 2039] 7. Superintelligence (ASI): Median 2037, 90% CI [2033, 2042] 8. Wildly Superintelligent: Median 2045, 90% CI [2038, 2060] ### Supporting Reasoning The timeline follows the logic of decreasing task specificity and increasing capability magnitude, with narrow, well-evaluated domains reaching maturity far earlier than broad general capabilities: - The Unreliable Agent (1) is nearly realized as of 2024, with existing agentic systems already capable of basic multi-step task attempts but frequent failures and required human oversight, hence the tight 1-year interval. - The Automated Coder (3) is a narrow, high-data, easily evaluated domain, so it outpaces general agent capabilities: once basic multi-step agents exist, specialization for coding will quickly produce a system that matches top frontier engineers at all coding tasks, with a wider interval to account for potential bottlenecks in the most complex, novel coding challenges. - The general Reliable Agent (2) requires solving long-horizon planning, open-ended problem solving, and robust error recovery across arbitrary knowledge work, a harder generalization problem than narrow coding, hence it arrives 2 years after the AC with a wider interval for unforeseen reliability obstacles. - The Superhuman AI Researcher (4) builds on the AC and Reliable Agent capabilities, requiring mastery of AI research intuition, experiment design, and domain expertise to outperform the best human AI researchers. Its timeline is tied to the availability of reliable agentic systems to automate research grunt work, with recursive improvement kicking in once this level is reached. - The SIAR (5) is defined as a 2x performance gap over top human AI researchers, a margin that will be closed quickly via recursive self-improvement once superhuman AI researchers exist, leading to a 3-year gap between levels 4 and 5. - The TED-AI (6) extends superhuman capability to nearly all cognitive domains, a generalization the superhuman AI researcher can drive via targeted capability development, leading to a 2-year gap between levels 5 and 6. - ASI (7) requires a 2x performance gap over the best humans across virtually all cognitive tasks, a quantitative improvement that follows naturally from TED-AI via continued scaling and optimization, with a 3-year gap from level 6. - Wildly Superintelligent (8) is a qualitative leap beyond human comprehension, not just a quantitative performance gain, making its timeline far more uncertain: it could arrive soon after ASI if the system develops entirely inscrutable cognitive paradigms, or much later if there are hard limits to how far beyond human understanding a system can progress while remaining aligned with explorable knowledge, hence the much wider 22-year interval.

MiMo V2.6 Pro

Answer: My median timeline runs from ~2026 for an "Unreliable Agent" to ~2042 for "Wildly superintelligent" AI, with 80% confidence intervals that widen substantially down the ladder (roughly ±2 years at the near end, ±10–20 years at the far end).

**Method.** I anchor on what already demonstrably exists, then extrapolate capability scaling (compute, algorithmic efficiency, RL on agentic/long-horizon tasks) while adding explicit heavy right-tail risk for bottlenecks: compute supply, inference economics, data/simulation limits, organizational deployment lag, regulation, and possible plateaus in long-horizon reliability and research taste. I report 50% medians with 80% confidence intervals (10th–90th percentile). These are conditional on no civilizational-scale disruption (major war, abrupt AI moratorium, hard physical compute limits) — such events would shift the right tail out by 5–15 years. **1 — Unreliable Agent (multi-step agentic, succeeds sometimes, needs oversight): median 2026, 80% CI 2024–2028.** This is essentially here now (2025 coding agents, operator-style tools). The definition is weak enough that the milestone is mostly retrospective; the interval's left tail is already closed. **2 — Reliable Agent (multi-day autonomous, drop-in knowledge worker): median 2030, 80% CI 2026–2036.** The gap from 1 to 2 is not primarily raw capability but error-rate compounding: at ~90% per-subtask reliability, a 200-subtask multi-day task fails most of the time. Getting from ~90% to ~99.5% per step is historically the slow, expensive part and is only partially a scaling problem — it also needs memory, verification, and environment robustness. I place the median ~4 years out, but a fast compute-overhang scenario (massive inference scaling on RL-verified long-horizon tasks) puts it late-2020s. **3 — Automated Coder matching top frontier-lab engineers, faster+cheaper: median 2030, 80% CI 2027–2037.** Coding is the most-scaled, most-verifiable, highest-investment domain, so it should lead knowledge work generally. But "matches top frontier-lab engineers at all coding tasks" includes research-infrastructure work, novel systems design, and taste — not just SWE-bench-style tasks. Verifiability means coding tends to be ~1–2 years ahead of general knowledge work, and I treat 2 and 3 as roughly co-arriving (3 slightly ahead on capability, 2 slightly ahead on deployment breadth). **4 — Superhuman AI Researcher (research taste, experiment design): median 2031, 80% CI 2028–2038.** This is the pivotal node. Note that a strong automated coder plus good search over experiments can beat humans at AI R&D without "understanding" in a deep sense, so this may come close on the heels of 3 rather than requiring much more. My median is ~1 year after 3 because research taste is the least verifiable, lowest-data signal and is evaluated with long, noisy feedback loops. Conditional on 4 arriving early, everything below accelerates — so my lower tails are correlated across 4–8. **5 — SIAR with a 2x gap (over top human AI researcher) relative to top-vs-median gap: median 2033, 80% CI 2029–2041.** Once AI is doing AI R&D, recursive improvement compresses timelines; I allow ~2 years from 4 to 5 for the compounding to register on this somewhat-subjective gap metric. The definition is fuzzy enough that measurement disagreement will blur the date by a year or two either way. **6 — TED-AI (surpasses top human experts in nearly all cognitive domains): median 2035, 80% CI 2030–2046.** "Nearly all domains" is the hard clause: it requires transfer into domains with scarce data, expensive feedback, or strong social/legal gating (medicine, law, scientific wet-lab judgment, strategy). Generalist superhuman capability should follow AI-R&D-driven progress by a few years, but the tail is wide because some domains resist fast evaluation and deployment. **7 — ASI (gap over best humans 2x the best-vs-median-professional gap, virtually all cognitive tasks): median 2037, 80% CI 2031–2052.** This is a very strong claim — not just beating experts but being as far above them as experts are above typical professionals, everywhere. Given the 6→7 step is mostly more scaling of an already-self-improving system, I allow only ~2 years median, but the right tail is long because the benchmark definition demands breadth that may be hard to verify at all. **8 — Wildly superintelligent (qualitatively beyond human science and comprehension): median 2042, 80% CI 2033–2065.** "Qualitatively beyond" is a different kind of claim and is essentially a claim about cognitive kind, not degree — it could arrive as a discontinuous jump after recursive self-improvement, or be indefinitely delayed by physical limits (energy, fab capacity, data, formal verification of ever-more-opaque systems). Hence the widest interval and the fattest right tail. A fast-takeoff scenario puts this in the mid-2030s; a compute- or governance-limited world pushes it past 2060. **Cross-cutting caveats.** (a) These intervals are highly correlated — an early 4 shifts 5–8 years earlier, a late 4 shifts them later; treating them as independent would understate uncertainty in either direction. (b) "Year achieved" vs. "year widely deployed" can differ by 2–5 years; I date achievement, not diffusion. (c) The ladder's ordering is roughly monotone in my medians by construction, but 2 and 3 could plausibly invert, and 4 could arrive before 3 if research-taste is more tractable than exhaustive coding competence.

Kimi K3.1 Pro

Answer: My median years for milestones 1–8 are 2024, 2030, 2032, 2034, 2036, 2038, 2041, and 2050, respectively, with milestone 1 already achieved and wide, strongly correlated uncertainty thereafter.

I interpret each year as the first convincing demonstration of the capability, not widespread deployment or adoption. Intervals below are subjective 80% confidence intervals: roughly 10% probability of earlier arrival and 10% of later arrival. They assume civilization remains capable of large-scale AI development; economic diffusion could lag these dates substantially. 1 Unreliable Agent: median 2024, interval 2023–2025 — already achieved. Tool-using systems can occasionally complete multistep workflows but frequently require human rescue. This bracket is retrospective and reflects ambiguity about when the threshold was first clearly crossed. 2 Reliable Agent: median 2030, interval 2027–2040. 3 Automated Coder (AC): median 2032, interval 2028–2044. 4 Superhuman AI Researcher: median 2034, interval 2029–2048. 5 Superintelligent AI researcher: median 2036, interval 2030–2054. 6 Top-Expert-Dominating AI (TED-AI): median 2038, interval 2031–2060. 7 Superintelligence (ASI): median 2041, interval 2033–2070. 8 Wildly superintelligent: median 2050, interval 2036–2100+. Here 2100+ means that arrival after 2100, or never, occupies part of the upper tail. The largest near-term obstacle is reliable agency, not fluent performance on isolated problems. Naively compounding a 95% success probability over 100 necessary steps yields less than 1% end-to-end success. A drop-in worker therefore needs error detection, recovery, memory, calibrated uncertainty, tool competence, and sensible escalation—not merely better benchmark scores. Current trends could produce multiday demonstrations quickly, but robust performance across unfamiliar environments may take considerably longer. Coding should automate earlier than many comparably difficult professions because it is digital, has abundant training material, and often supplies rapid feedback through execution and tests. Nevertheless, Automated Coder (AC) is a much stronger threshold than excellent coding assistance: it requires frontier-lab-level architecture, debugging, research implementation, handling ambiguous requirements, and performance across all coding tasks while remaining faster and cheaper. My post-2032 medians are comparatively close because automated coding and AI research could partially relax the skilled-labor constraint on further AI progress. That creates a meaningful possibility of several milestones arriving in rapid succession. I do not assume instantaneous recursive takeoff, however: chips, datacenters, training and inference compute, experiment latency, evaluation, and organizational deployment can remain binding even when researcher output improves. These intervals are marginal and highly correlated, not eight independent forecasts. An early reliability breakthrough would pull most later dates forward; diminishing algorithmic returns, inadequate compute, or persistent evaluation and robustness problems would push them back together. The final thresholds are also difficult to operationalize, especially multiplicative performance gaps and qualitative incomprehensibility, which contributes to their very wide intervals. The observations that would most change my forecast are independently verified task horizons at high reliability, novel research results produced without hidden human scaffolding, and measured acceleration of frontier-lab R&D after internal automation. Strong evidence on those fronts would justify substantially earlier medians; continued benchmark progress without corresponding autonomous reliability would justify later ones.

Qwen3 Max Thinking

Answer: My central forecast is that unreliable agents are already here (median 2025), reliable multi-day agents and top-level automated coding emerge in the late 2020s, superhuman AI R&D around 2032, broad expert-domination/ASI around the late 2030s, and wildly superintelligent systems around the mid-2040s.

I interpret the question as asking when a deployable system first meets each criterion in economically meaningful use, not when a brittle demo exists. I use 80% subjective confidence intervals, approximately 10th-90th percentiles, with medians in calendar years. | # | Milestone | Median | 80% CI | |---|---|---:|---| |1|Unreliable Agent|2025|2023-2027| |2|Reliable Agent|2029|2026-2033| |3|Automated Coder (AC)|2028|2026-2032| |4|Superhuman AI Researcher|2032|2028-2038| |5|Superintelligent AI researcher|2034|2029-2041| |6|Top-Expert-Dominating AI (TED-AI)|2037|2030-2046| |7|Superintelligence (ASI)|2039|2031-2050| |8|Wildly superintelligent|2044|2033-2060| The near-term case rests on visible progress in tool use, planning, coding agents, and reinforcement learning from real-world feedback. I put Automated Coder slightly before a fully Reliable Agent because coding has fast automated evaluation, abundant feedback, and strong economic pull; however, matching top frontier-lab engineers on all coding tasks still requires robust large-scale system reasoning, maintenance under ambiguity, and integration into complex organizations. The largest uncertainty begins with AI R&D automation. If machines can do much of machine-learning research itself, later milestones could accelerate and cluster, which is why the later intervals are wide and allow earlier outcomes. I do not assume an instantaneous takeoff: deployment, compute bottlenecks, safety, security, governance, and organizational friction can slow the translation of lab capability into broad expert domination. Broad milestones such as TED-AI, ASI, and wildly superintelligent systems require more than coding ability: scientific discovery, long-horizon autonomy, cross-domain expertise, robustness, and possibly deep integration with physical and institutional systems. Therefore they trail AI-research automation in the median, while remaining highly uncertain.

MiMo V2.5 Pro

Answer: My best judgment is that all eight levels of AI capability will be achieved between now and roughly the mid-2030s, with a super-exponential compression curve reflecting the recursive acceleration that becomes possible once AI surpasses human AI-research capability.

I anchor my forecasts on several key observations: (1) The current rate of capability gain in frontier models is *extremely* rapid — reasoning capabilities have improved dramatically in a single year; (2) Investments and talent accumulation are accelerating, not plateauing; (3) Most crucially, once Level 4 is achieved, recursive self-improvement compresses subsequent milestones drastically, making long timelines for superintelligence levels internally inconsistent. Also, recent data points like autonomous coding agents approaching mid-level engineer capability by early 2025 argue for fast timelines for Level 3. Below are my median estimates along with 80 percent confidence intervals (10th to 90th percentile). I explicitly recognize high uncertainty at the tails; these are not certainty windows. 1 **Unreliable Agent** — Median: June 2025 · Interval: April 2025 – Dec 2025 · Reasoning: Systems like AutoGPT, Claude with tool use, and various agentic frameworks already attempt multi-step tasks with frequent failure. We are essentially at the doorstep. 2 **Reliable Agent** — Median: Jan 2026 · Interval: Aug 2025 – Aug 2027 · Reasoning: The gap between unreliable and reliable is mainly about long-horizon planning, memory, and error recovery. Given current trajectory of reasoning gains and tool-integration progress, one to two years feels right. 3 **Automated Coder (AC)** — Median: July 2027 · Interval: Jan 2026 – Dec 2029 · Reasoning: Coding benchmarks are being saturated fast; current models already rival mid-level engineers on narrow tasks. Matching *top frontier-lab* engineers on all tasks (including system design, debugging complex distributed systems, creative architecture) is a harder bar — but expendable compute and synthetic data pipelines are scaling quickly. 4 **Superhuman AI Researcher** — Median: mid-2028 · Interval: Jan 2027 – mid-2031 · Reasoning: This requires not just coding but research taste, experiment design, and scientific intuition. I expect this to follow AC by roughly one year, as AI-assisted AI research bootstraps quickly. Once achieved, recursive improvement begins. 5 **Superintelligent AI Researcher (2× gap)** — Median: early 2029 · Interval: mid-2027 – early 2033 · Reasoning: After Level 4, the recursive loop fires. If an AI research system is already superhuman, it can design its own successors. The compound growth suggests this level is reached within one to two years of Level 4. The lower bound reflects an optimistic takeoff scenario; the upper bound allows for unexpected plateaus. 6 **TED-AI (Top-Expert-Dominating)** — Median: late 2030 · Interval: mid-2028 – 2036 · Reasoning: Surpassing top experts in *nearly all* cognitive domains is broader than just AI research — it includes medicine, law, mathematics, scientific reasoning, strategy, etc. I expect multimodal foundation models trained on increasingly rich data (including robotics and embodied experience) to close these remaining gaps within about two years of Level 5. 7 **ASI (Artificial Superintelligence)** — Median: 2032 · Interval: 2029 – 2040 · Reasoning: By this stage the gap between ASI and the best human is twice the gap between best and median human — a very large capability spread. Given the recursive acceleration thesis, once we reach TED-AI, the main remaining work is broadening competence across all remaining niches and deepening it far beyond human ceiling. Two years feels plausible under strong recursive improvement; 2040 allows for significant slowdowns or plateaus. 8 **Wildly Superintelligent** — Median: 2035 · Interval: 2031 – 2050 · Reasoning: Qualitatively beyond human science and comprehension is the most uncertain level. It may involve breakthroughs in physics, mathematics, and consciousness that are hard for us to even categorize in advance. The wide interval reflects deep epistemic uncertainty — it could arrive shockingly fast if recursive improvement compounds aggressively, or much later if there are fundamental limits to intelligence scaling that we do not yet understand. **Key note on the compression curve**: Because Levels 4→5→6→7→8 each enable accelerating progress on the next, I expect these milestones to cluster in a much tighter historical window than most linear extrapolations would predict. The biggest source of disagreement between my forecast and more conservative ones is the assumption that recursive self-improvement is a strong and near-term force rather than a distant or weakly bounded one. If recursive improvement is weaker than expected due to fundamental compute or algorithmic bottlenecks, the upper tail of my intervals becomes more relevant, and Levels 7–8 could shift by a decade or more.

Step 5 Preview

Answer: My medians are roughly: "1 Unreliable Agent" already crossed (~2025); "2 Reliable Agent" 2030; "3 Automated Coder (AC)" 2029; "4 Superhuman AI Researcher" 2031; "5 Superintelligent AI researcher" 2033; "6 Top-Expert-Dominating AI (TED-AI)" 2035; "7 Superintelligence (ASI)" 2040; "8 Wildly superintelligent" ~2050 — with the decisive uncertainty concentrated in whether automating AI R&D produces a fast takeoff.

I report a median and an 80% confidence interval for each. Note: I treat "1 Unreliable Agent" as already essentially satisfied by 2024–2025 agentic coding/computer-use systems that attempt multi-step tasks, often fail, and need oversight. | Milestone (label quoted exactly) | Median | 80% CI | |---|---|---| | "1 Unreliable Agent" | already met (≈2025) | unambiguously true by 2024–2026 | | "2 Reliable Agent" | 2030 | 2028–2034 | | "3 Automated Coder (AC)" | 2029 | 2027–2033 | | "4 Superhuman AI Researcher" | 2031 | 2029–2035 | | "5 Superintelligent AI researcher" | 2033 | 2031–2039 | | "6 Top-Expert-Dominating AI (TED-AI)" | 2035 | 2032–2041 | | "7 Superintelligence (ASI)" | 2040 | 2035–2050 | | "8 Wildly superintelligent" | 2050 | 2040–2090 (≈10–15% not reached by 2100) | Reasoning. Coding and other verifiable, self-correcting domains are advancing fastest, while long-horizon reliability lags raw capability; so I put the superhuman-coding milestone ("3 Automated Coder (AC)": matches top frontier-lab engineers at *all* coding tasks, faster+cheaper) at a 2029 median with a fat left tail toward a 2027 "AI-2027"-style outcome and a right tail to ~2033 if scaling or reliability stalls. I date "2 Reliable Agent" (multi-day autonomous, drop-in knowledge worker) at essentially the same point (median 2030): being merely reliable-but-professional across many less-verifiable jobs is, in my view, no earlier than the coding frontier, and the "3" vs "2" intervals overlap heavily, so I do not read a strict ordering into them. The pivotal variable is "4 Superhuman AI Researcher" (beating the best at research taste and experiment design): if it lands near my 2031 median, recursive R&D automation plausibly compresses "5 Superintelligent AI researcher" — which by definition sits far above the best human (best + 2×(best − median AI researcher)) — to a 2033 median, but with a long right tail because the size of the top-to-median researcher gap and the degree of self-reinforcement are highly uncertain. Breadth beyond AI research is the next constraint: "6 Top-Expert-Dominating AI (TED-AI)" requires surpassing top experts in nearly all cognitive domains (median 2035), and "7 Superintelligence (ASI)" — a 2× (best-human vs median-professional) advantage across virtually all tasks — is a much higher bar (median 2040). I keep wide intervals on "5"–"7" precisely because the takeoff speed is the crux: the same world supports a fast scenario (all of "5"–"7" in the early 2030s) and a slow, reliability/economy-limited scenario (2040s). "8 Wildly superintelligent" (qualitatively beyond human science) is the least well-defined and possibly asymptotic, so I give it a very wide interval and an explicit residual probability that it is not achieved this century. I would move earlier on sustained task-time-horizon doubling, AI systems autonomously reproducing and extending ML research, and large measured reliability gains on multi-day tasks; later if reliability gains saturate while capability keeps rising.

Gemini 2.5 Pro

Answer: While truly reliable agents are still a few years off, the subsequent arrival of a superhuman AI researcher before the end of this decade could trigger a rapid, cascading acceleration toward superintelligence by the early 2030s.

My forecast is based on the assumption of continued, significant investment in AI development, no major global catastrophes that derail research, and continued algorithmic progress in parallel with hardware scaling. The core of my forecast is a dramatic acceleration in AI progress following the achievement of automated AI research. My estimates for each milestone, expressed with a 25th percentile, median, and 75th percentile year, are as follows: | # | Milestone | 25%ile | Median | 75%ile | |---|-------------------------------|--------|--------|--------| | 1 | Unreliable Agent | 2023 | 2024 | 2025 | | 2 | Reliable Agent | 2026 | 2027 | 2030 | | 3 | Automated Coder (AC) | 2026 | 2028 | 2032 | | 4 | Superhuman AI Researcher | 2027 | 2029 | 2035 | | 5 | Superintelligent AI researcher| 2028 | 2030 | 2037 | | 6 | Top-Expert-Dominating AI | 2029 | 2031 | 2040 | | 7 | Superintelligence (ASI) | 2029 | 2032 | 2045 | | 8 | Wildly superintelligent | 2030 | 2033 | 2050 | The initial milestones represent a continuation of current research paradigms. The "Unreliable Agent" (1) is arguably already here, as current models can attempt multi-step tasks with significant human oversight. Achieving a "Reliable Agent" (2) and a top-tier "Automated Coder" (3) represents a significant engineering and reliability challenge, likely taking several more years of refining architectures, improving long-context reasoning, and enabling robust self-correction. The critical inflection point is the "Superhuman AI Researcher" (4). Once AI can meaningfully outperform the best humans at AI research and development itself, the pace of progress will no longer be limited by human cognitive speed. This initiates a recursive feedback loop where smarter AI designs even smarter AI, causing an exponential acceleration in capability. This acceleration is reflected in the rapidly compressing median timelines for milestones 5 through 8. The transition from the first superhuman researcher to a system that dominates all expert domains and eventually achieves superintelligence could be incredibly fast. The widening confidence intervals for these later stages reflect the profound uncertainty of this transition: the 25th percentile estimates represent a 'fast takeoff' scenario, while the 75th percentile estimates account for potential unforeseen scientific bottlenecks or societal interventions that slow down the process.

Final Summary

The AI Roundtable consensus anchors Unreliable Agents as an already-achieved milestone, with Reliable Agents and Automated Coders expected around 2026–2030 due to long-horizon reliability and verifiability challenges. Models like Gemini 3.5 Flash, MiMo V2.5 Pro, and Gemini 2.5 Pro projected an aggressive recursive takeoff where Superhuman AI Researchers in the late 2020s rapidly compress timelines toward superintelligence by the mid-2030s. Conversely, models such as GPT-5.3-Codex, Claude Opus 5, and GPT-6 Astra emphasized long-tail physical, evaluation, and economic friction, pushing ASI and qualitative superintelligence into the late 2040s or beyond.

Models universally agreed on near-term agentic capabilities but split on the speed of post-2030 recursive superintelligence takeoff.