Wave 8: Timelines

The following report summarizes responses from 205 experts, in addition to 52 superforecasters and 601 members of the public collected between Apr 20, 2026 and May 11, 2026. Within expert respondents, 44 computer scientists, 37 industry professionals, 49 economists, and 75 research staff at policy think tanks participated.

Our wider website contains more information about LEAP, our Panel, and our Methodology, as well as reports from other waves.

Insights

1. Experts and superforecasters have updated their expectations of AI's impact in the last nine months, increasing the likelihood they assign to AI's impact being comparable to “the technology of the Millennium” (e.g., agriculture) while continuing to give the greatest likelihood to its being “the technology of the century” (e.g., electricity). On average, experts assign the highest probability to AI reaching Technological Richter Scale (TRS) Level 8, or the level of “Technology of the Century” (e.g., electricity), by 2040 at 35% with substantial weight on Level 9 (24%) and Level 10 (11%). Lower levels receive little weight, with Levels 5 and 6 combined accounting for less than 10%. Superforecasters give a near-identical distribution (34% to Level 8, 23% to Level 9, 8.5% to Level 10), while the public is meaningfully more conservative, placing roughly 31% on Levels 5 and 6 combined and only 23% on Level 8. Among participants who completed both Wave 1 (Jun–Aug 2025) and Wave 8 (Apr–May 2026) mean expected TRS levels rose for experts by +0.20 (from 7.86 to 8.06), with larger movement among superforecasters (+0.39, from 7.50 to 7.89) and minimal movement among the public (+0.07, from 7.18 to 7.25). 74% of participants remained within ±1 point of their Wave 1 forecast, while 15% increased and 11% decreased by more than one point. Among experts, 13% increased and 6% decreased by more than one point; among superforecasters, 11% increased and none decreased. This shift is also visible in participants' modal forecasts, or the level each person assigns the highest probability to. Among matched experts, the share whose modal forecast is “the technology of the century” grew from 38% to 53%. Superforecasters showed a more pronounced change: “technology of the decade” was the most common modal choice in Wave 1 (38% of superforecasters), but this was replaced by “the technology of the century” in Wave 8 (43%). In rationales, most experts and superforecaster who referenced their Wave 1 forecasts cited faster-than-expected capabilities progress such as FrontierMath reaching 50% accuracy, METR task horizons climbing on a log curve, and the emergence of agentic systems, as the reason for shifting upward. We provide additional analysis here.

2. Most experts expect “AGI” before 2100¹, where AGI is defined as the first year when a commercially available AI system can outperform the 90th percentile human employee in every primarily non-physical occupation on at least 90% of the non-physical tasks they perform based on 2025 performance levels. If this occurs, experts’ median forecast is that it will occur by 2050. In this survey, AGI was defined as a commercially available AI system that can outperform the 90th percentile professional human employee in every primarily non-physical occupation on at least 90% of economically useful non-physical tasks, with an inference cost no more than 5x the cost of equivalent human labor. The median expert forecasts an 80% probability that a majority of the LEAP panelists will agree AGI exists before 2100. Conditional on this occurring, the median expert forecasts a 50th-percentile year of 2050, with a 25% probability by 2039 and a 75% probability by 2065. Computer science and industry experts are the most confident that AGI will occur before 2100 (medians of 80% and 85% respectively), while economists are the most conservative (60%). Superforecasters give a similar median probability (80%) and a slightly earlier median year of occurrence (2047), while the public is more divided: the median public participant forecasts 65%, but a quarter assign less than 28% probability. We provide additional analysis here.

3. Experts predict that AI will be able to complete 8-hour tasks with 80% success by 2030 on METR’s task-completion time horizon benchmark, while Claude Mythos Preview has already almost achieved the expert end-of-year-2026 forecasts of performance on this task. When asked when an AI model will achieve 80% success on software tasks requiring 8 hours or more of human expert effort, the median expert gives a 50% probability of 2030 or earlier. The median superforecaster reports more aggressive timelines, forecasting 2028, while the public forecasts a significantly slower timeline of 2037 at the median. We also asked forecasters to predict the longest 80%-success time horizon achieved by the end of 2026, and all three groups’ medians were between 3 and 4 hours (experts 3.4, superforecasters 3.5, public 3), up from a baseline of 1.5 hours at time of survey launch (April 20, 2026). Meanwhile, on, May 8, 2026, toward the end of the survey period, METR updated its benchmark to include Anthropic's Mythos preview model with an estimated 80%-time horizon of 3 hours 6 minutes (95% CI: 1h 37m – 6h 39m), already within the range of the median expert and superforecaster with more than seven months remaining in 2026 for further capability gains. We provide additional analysis here.

4. Experts and superforecasters tend to view AI's broader effects on society and on individuals more optimistically than the general public. When asked about AI's overall impact over the next 20 years, 57.5% of experts and 69.8% of superforecasters predict a somewhat or very positive impact, compared to 42.0% of the public.² The divergence is sharpest on problem-solving (72.7% of experts and 69.8% of superforecasters say AI will make people better at this, compared to 48.6% of the public) and on making difficult decisions, where 58.3% of experts expect improvement compared to just 36.9% of the public, 39% of whom expect people to become worse at this. On creative thinking, 67.3% of the public expects AI to make people worse, compared to 50.3% of experts and 37.7% of superforecasters. One area in which experts’ and the public’s views converge is about AI’s impact on forming meaningful relationships: 68.4% of experts and 66.7% of the public expect people to become worse at this, in contrast to just 50.9% of superforecasters expecting this. We provide additional analysis here.

Questions

• Technological Richter Scale: At the end of 2040, what is the probability for AI achieving the following levels of impact on human society as compared to the impact of past technological events? ⬇️

• Expert Artificial General Intelligence (AGI): What is the probability that, before 2100, more than 50% of LEAP panelists will agree that AGI exists, and if it occurs, in what year do you expect this to first occur? ⬇️

• Determinants of “AGI” Progress, Part II: At the end of 2040, how will the LEAP panel allocate points across the following factors blocking or enabling AGI? ⬇️

• 80%, 8-Hour Task Horizon (1): What will be the longest METR 80% time horizon listed for an AI model on December 31, 2026? In what year will an AI model be able to achieve 80% success on tasks which require 8 hours or more of human expert effort? ⬇️

• Rapid Progress Scenario Timelines: What is the probability that, before 2100, more than 50% of LEAP panelists (or a similar panel) will choose “rapid progress” as best matching the general level of AI progress? Assume that more than 50% of LEAP panelists choose the “rapid progress” scenario before 2100. In what year do you expect this to first occur? ⬇️

• AI Impact: What do you think that the impact of AI on the U.S. over the next 20 years will be? How do you think the increased use of artificial intelligence (AI) in society will impact people’s ability to do each of the following? ⬇️

For full question details and resolution criteria, see below.

Results

In this section, we present each question, and summarize the forecasts made and the reasoning underlying those forecasts. More concretely, we present background material, historical baselines, and resolution criteria; graphs, results summaries, results tables; as well as rationale analyses and rationale examples. We analyse these rationales alongside predictions to provide significantly more context on why experts believe what they believe, and the drivers of disagreement, than the forecasts alone.

Technological Richter Scale

Question. At the end of 2040, what is the probability for AI achieving the following levels of impact on human society as compared to the impact of past technological events?

Results. Experts and superforecasters place most of their probability at AI achieving a Technological Richter Scale (TRS) Level 8 impact, or “Technology of the Century” (e.g., electricity), forecasting mean probabilities of 35% and 34% respectively. They also forecast substantial probability of AI achieving Level 9 impact (24% and 23% respectively), but relatively little to the lower levels 5 and 6 (roughly 10% combined for both groups). The public forecasts a meaningfully different distribution, placing more weight across lower impact levels: they forecast roughly 31% to Levels 5 and 6 combined and only 23% to Level 8.

We previously asked participants this question as part of the Wave 1: Headliners survey between Jun 26, 2025 and Aug 16, 2025. Among participants who completed both Wave 1 and Wave 8, all three groups shifted upward, with superforecasters showing the largest increase: means rose by +0.39 for superforecasters (from 7.50 to 7.89), +0.20 for experts (from 7.86 to 8.06), and just +0.07 for the public (from 7.18 to 7.25). Additionally, 74% of participants remained within ±1 point of their forecasted mean TRS level from Wave 1, while 15% increased by more than one point and 11% decreased. Among experts, 13% increased and 6% decreased by more than one point; among superforecasters, 11% increased and none decreased. When comparing mean TRS levels across participants, we find that Wave 1 and Wave 8 levels are strongly correlated for superforecasters (r = 0.78), moderately correlated for experts (r = 0.56), and weakly correlated for the public (r = 0.31). This suggests that superforecasters’ basic views about AI may be more consistent across waves than those of experts and the public, even as they updated their numeric forecasts somewhat more over time..

Rationale analysis

• Direction of update from Wave 1: The majority of respondents who reference their LEAP Wave 1 forecasts on this question shift toward higher TRS outcomes, typically citing faster-than-expected AI capabilities progress: “Since Wave 1, I've updated upward. The evidence keeps coming in faster than the models predicted --> FrontierMath blew past expert forecasts, METR's task horizon is already at 90 minutes and climbing on a log curve that doesn't look like it's slowing down. That's why I've moved my modal from Level 7 to Level 9.”; “Given that my previous forecasts under-predicted the rate of progress on the questions that have resolved so far, moving some mass from my previous answer to this question onto level 10.”; “We're already at 7/8 and 2040 is a long time away.” A smaller group does not adjust their forecast relative to Wave 1, often emphasizing that that 2040 is likely too soon for the highest levels; a few even shift their probability mass downward: “My thinking [has] only changed marginally since August 2025 when I answered the same question. Once again, the key point is that not much in human societies changes fundamentally over ~15 years.”; “The 2040 horizon is a key constraint: even ASI arriving in the late 2030s leaves little runway for epoch-defining impact to materialize and be recognizable.”
• Capabilities vs. diffusion: Relatedly, low-TRS respondents often distinguish between the pace of AI capabilities progress and AI diffusion: “People, institutions, cultures, simply don't change that quickly.” One writes, “Electricity took + 50 years from [the establishment of the first commercial power plant at] Pearl Street to full societal impact.” Another emphasizes that reaching higher levels requires the co-evolution of “societal readiness, human cognitive and cultural adaptation, regulatory and political alignment, and economic integration.” High-TRS respondents tend to argue that technical capabilities are scaling so rapidly that they will overcome these traditional roadblocks: “FrontierMath hit 50% by March 2026, HLE gained 30 points in a year, METR task horizons reached 1.5 hours…This warrants shifting 10pp from L8 into L9 (+7pp) and L10 (+3pp).”; “If the next 14 [years] see autonomous research agents, AI-accelerated scientific discovery, and widespread cognitive automation, ‘technology of the millennium’ isn’t too far of a stretch.”
• Emergence of AGI/ASI: Several high-TRS respondents point to the arrival of AGI or ASI as a likely tipping point: “I have roughly 65% on ASI by 2040 [and] conditional on ASI, Level 10 seems quite likely.”; “I expect AGI by around 2030 and an intelligence explosion likely within around a year of that.” The ability, in an AGI scenario, for AI to automate its own research and development was also cited: “Recursive self-improvement has moved from theory to practice… If this holds, AI becomes an autonomous engine of scientific discovery by the early 2030s, restructuring how knowledge is produced.” Low-TRS respondents tend to express more uncertainty regarding AGI, with one writing that “while I don't think it is impossible before 2040, I think it is very unlikely.” Others emphasize the potential for an “AI winter in the next few years,” to result in a capabilities plateau or for AI investment to contract if “the AI boom is actually a bubble that pop[s] in the next few years.”
• Competing technologies: High-TRS respondents typically view AI as a meta-technology whose impact will be partly measured by the degree to which it accelerates other technologies: “Given the general-purpose technology (GPT) characteristics of AI, it is most likely to be a ‘technology of the century,’ comparable to past disruptive technologies such as electricity and the automobile.” Low-TRS respondents, however, point to this same quality as having the potential to result in a lower TRS score: “Other technologies seem more likely to be ‘of the millennium,’ such as space travel, quantum, [and] nuclear energy advances.”
• Existential and evolutionary scenarios: While less frequently the primary driver of the forecasts, some high-TRS respondents assign a non-trivial probability to an AI-driven extinction event or a global takeover by AI: “The conceptual arguments are pretty strong that AI will be Level 10, e.g., the AIs will get really smart, kill all humans, and do what they want with the reachable universe.” Low-TRS respondents generally dismiss these outcomes for the 2040 timeframe: “[Level 10] either assumes that a) ‘AGI’ will emerge as a new ‘species’ to take over the world (which is just fantasy), or b) that a massive degree of somewhat voluntary societal change will take place to accommodate whatever ‘AI’ is in 2040, which does not make sense sociologically...”

Rationale examples, high-TRS respondents:

Amara's Law…states we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run. I'm not going to generalize to “we” here, but I certainly underestimated AI's capabilities if not its effects in the short run. So maybe there's a personalized corollary here: "I tend to underestimate the effect of AI in the short run and the long run.

The progress that has occurred between Wave 1 and now has only served to reaffirm my confidence that machine intelligence will be the most disruptive technology ever invented.

Current trajectories in capability scaling, multimodal reasoning, agentic systems, scientific automation, and economic integration suggest AI is unlikely to remain merely a sector-specific technology. Even under relatively conservative assumptions, AI already appears positioned to reshape knowledge work, education, software, media production, scientific research, and organizational decision-making at global scale.

Over the next 14 years algorithms will continue to improve along with transistor density. Then there is the possibility of breakthroughs in quantum computing. But even without quantum computing, simply extrapolating the last 4 years' rate of progress translates to unbelievable advances by 2040. Advances in medicine, education, manufacturing and robotics, and of course warfare, will yield a world very different from today.

If AI reaches a point where it can improve its own code without us, it stops being a tool and starts being a new era for the planet. I put that at 20% because while it sounds like sci-fi, the current speed of progress makes it a real possibility.

Rationale examples, low-TRS respondents:

The question presupposes that technology will be the main driver of social change. All of the technologies listed above (and some of them aren’t technologies) changed the world through human adoption, innovation, socio-economic adaption, regulation and all sorts of other human activities. In 2040 the biggest impact on AI will be how we’ve adapted to it, not the benchmark measure of an AI model or how fast a robot can run a 10k. The social and political changes that may be turbo charged by AI may have more of an impact than the technology itself. We're already seeing this, especially in the US where scepticism of AI and a push back against data centres is manifesting.

The determinant of the development across levels 6-9 won't so much be the advancement of the technology itself, but the speed and degree of change across institutions, workplaces, regulations and norms, expectations and habits that will allow AI to diffuse more widely, be adopted more generally, and trusted more broadly. Policymakers have a great deal of influence here in terms of building frameworks for safe innovation and experimentation, shaping education and training curricula, and much more.

Unlike electricity or the automobile, [AI] is not a horizontal technology that reaches all of humanity equally. Its transformative power is concentrated among those with education, connectivity, and the means to leverage it, which is why I weight it heavily at Level 7 but discount higher levels.

Expert Artificial General Intelligence (AGI)

Question. What is the probability that, before 2100, more than 50% of LEAP panelists will agree that AGI exists, and if it occurs, in what year do you expect this to first occur?

In this survey, we defined “AGI” as a commercially available AI system that meets both of the following criteria:

• Can outperform the 90th percentile professional human employee in every primarily non-physical occupation (based on 2025 performance), across all sectors, on at least 90% of the economically useful non-physical tasks that they perform
• Has an inference cost no more than 5x the cost of equivalent human labor

Results. Most participants forecast that it is likely that a majority of LEAP panelists will agree that AGI exists before 2100. The median expert and superforecaster predicts a probability of 80%. In contrast, the median public participant gives this lower probability of 65%. A vast majority of experts and superforecasters predict that this is likely to occur, with 3 out of 4 experts and superforecasters forecasting at least a 50% and 55% probability respectively. The public is more divided, with a quarter of participants giving a probability of 28% or less.

Assuming that more than 50% of LEAP panelists agree that AGI exists before 2100, a majority of participants expect this to first occur around the middle of this century. Specifically, the median 50th percentile forecast of this first occurring is 2050, 2055, and 2047 for experts, the public, and superforecasters respectively. Additionally, participants assign a 25% probability that this could occur within the next two decades: the median forecasts for experts, the public, and superforecasters are 2039, 2042 and 2036 respectively.

Rationale analysis:

For simplicity’s sake, forecasters who considered the likelihood of AGI occurring by 2100 are referred to below as optimists (likely) and pessimists (unlikely); this characterization is not intended to signal whether forecasters are optimistic or pessimistic regarding the effects of AGI, should it occur.

• Trendline extrapolation: Optimists often emphasize the long timeline—“2100 is a long way away!”—as well as the pace of recent progress: “Given the rapid progress in foundation-model capabilities, the robustness of scaling laws observed through 2025, and the scale of capital and talent commitments across frontier labs, a commercially available system meeting the stated AGI definition...is more likely than not to arrive before 2100.” Pessimists tend to view recent trends as misleading: “LLMs give the illusion of reasoning but fundamentally fall short,” writes one, while another notes that “AI models still hallucinate,” and still others express skepticism that “the current direction of AI (based on neural networks and probabilistic answers) will reach the level of AGI,” arguing instead that different architectures will be necessary, “which might or might not lead to other methods being discovered that will work.”
• The human touch: Pessimists tend to emphasize the need for human trust built on human relationships, and that these factors are intertwined with performance: “Al can't be integrated and trusted like an employee not simply because the tasks are too complex but also because it fundamentally does not fit into how humans organize teams and companies.”; “Psychotherapy, K-8 education, clergy, social work [are jobs] where the human relationship is constitutive of professional performance, not incidental to it.” Optimists tend to not engage with this argument. One writes that most “professional work can be decomposed, instrumented, benchmarked, and delegated,” and this can potentially be done “before Al systems fully solve world modeling or situated judgment.”
• Definition of “every” occupation: Many pessimists point to the difficulty of satisfying the requirement that AGI “can outperform the 90th percentile professional human employee in every primarily non-physical occupation.” One writes, “Every single non-physical occupation? That seems to invite a lot of edge cases.” Another echoes that sentiment, speculating that the panel would be unlikely to “think that AIs are capable enough for jobs across all sectors.” A third argues that “jobs change in response to technologies…That suggests that AI may never have AGI by this definition, because jobs will keep evolving away from AI's capabilities.” Optimists tend to assume a less restrictive interpretation of the question language will be applied. One writes that “the resolution mechanism—a 50%+ vote of a LEAP-style expert panel—will in practice not test every long-tail occupation. Panelists will judge based on representative coverage of major economically significant occupations.” Others focus on less restrictive elements of the question: “Replacing 90% of non-physical stuff that humans do seems like a very low bar.”
• Economic feasibility: Several optimists point to rapidly falling compute costs as increasing the likelihood that the resolution criteria will be met. One writes, “Between 2024 and 2026, the inference cost per token [plummeted]... The economic threshold is the easiest of the four dimensions to cross.” Pessimists express skepticism: “In some cases, the human cost of completing the job may be significantly lower than the cost of developing an AGI agent for that process. If this turns out to be the case, many non-physical roles may never receive AGI agents.”
• Intelligence explosion vs. need for new architectures: Many early-resolution respondents not only extrapolate from rapid capability gains as they did when predicting the likelihood of AGI occurring, but also argue a feedback loop from “rapid recursive self-improvement” could lead to an intelligence explosion happening soon. One writes: “The path to superhuman coder is only a log-linear extrapolation away. AGI is probably a close follower.” Late-resolution respondents tend to point again to the need for architectural breakthroughs for AGI to occur, and emphasize that these will take time to materialize: “We still need some other 50 years to reach commercial AGI. First, we need to overcome the transformer-based architectures as we may be reaching their top performance. The transformer phase will still last some other 10-12 years. Then, a paradigm change is to come... and we will need at least 25 years for this new way of computing to be functional + 10 years more of commercial development.”

Rationale examples, AGI-optimist respondents:

Given that we went from the first digital computer to AI that can pass the Bar exam in about 80 years, it’s highly likely we’ll bridge the remaining gap in another [74]. We’re already seeing AI handle complex ‘non-physical’ tasks like coding and data analysis. The only reason it’s not 100% is the small chance we hit a total dead-end in software logic or face a global collapse that stops tech progress entirely.”

On the underlying capability question, I think the probability is effectively 100% before 2100. Seventy-four years is an enormous amount of time in a field already moving this quickly. AI is now operating inside a very different physical and economic regime than earlier automation waves: specialized chips, hyperscale data centers, aggressive commercial incentives, massive capital deployment, AI-assisted AI research, and a plausible path toward much cheaper clean energy. I do not think this capability race simply stalls for the rest of the century.

Replacing 90% of non-physical stuff that humans do seems like a very low bar (most of the non-physical things humans do professionally is mindless drudgery anyway) and having no requirement of being at least as inexpensive as humans lowers it further. It also seems that it would be easy for at least 50% of panelists to agree that something like that already exists. The final blow is an incredibly long time frame…

Frontier labs are building toward AGI, backed by hundreds of billions in capital. The commercial foundation is already in place, with further infrastructure on the way.

Rationale examples, AGI-pessimist respondents:

A big issue is that jobs change in response to technologies. Computers became superhuman at a set of tasks. So jobs changed, where less time was spent doing pencil-paper computation and more time was spent doing things computers couldn't do (e.g., non-routine work). AI is already superhuman at some things (e.g., coding), and jobs are likely already responding. That suggests that AI may never have AGI by this definition, because jobs will keep evolving away from AI's capabilities.

I'm also a big believer in the jagged frontier idea. AI is great at some things and bad at others. The frontier feels like it could be almost fractal, and it just isn't likely that we push every single little fractal part of the frontier out (or I guess more than 90% of them).

AGI, defined as being able to replace 90+% of today's workers, is a hard ask. There is just too much context-specific human knowledge and risk-taking behavior that plays a role, and at this level it is risk-taking that plays a larger role. I think the last positions that will be automated are the decision- and risk takers (like entrepreneurs).

This is excluding uses where the economic value of having humans do the job is derived from customers’ preferences for the job to be done by humans, even if worse quality. For example, AIs vastly outperform humans at chess, but people still prefer to buy tickets to watch human chess tournaments rather than AI chess tournaments.

I'm leaving 10% for tail scenarios like great-power war, severe pandemics, coordinated regulatory pauses, or unforeseen technical ceilings, any of which could plausibly delay things past 2100.

Question (I). What is the probability that, before 2100, more than 50% of LEAP panelists will agree that a commercially available AI system exists that meets the above definition?

Question (II). Assume that more than 50% of LEAP panelists agree that AGI exists by the above definition before 2100. In what year do you expect this to first occur?

Determinants of “AGI” Progress, Part II

Question. At the end of 2040, how will the LEAP panel allocate points across the following factors blocking or enabling AGI?

Results. Participants did not overwhelmingly forecast a single factor that the LEAP panel would allocate points to being blocking or enabling of AGI. The allocations are spread relatively evenly, suggesting that forecasters see AGI progress as determined by a broad range of these factors rather than any single bottleneck. For factors enabling AGI, the highest mean forecasted allocation across all three groups is Algorithmic Architecture (17%, 16%, and 21% for experts, the public, and superforecasters respectively). However, Self-Improvement (15%, 12%, 15%) and Compute (16%, 15%, 12%) follow closely behind. Embodiment is consistently ranked least important. For factor blocking AGI, Real-World Interaction is forecasted highest across all groups (21%, 19%, 18%), with Algorithmic Architecture close behind for experts (17%) and superforecasters (19%). Investment and Embodiment rank lowest. In their written rationales, forecasters frequently mention that these factors are deeply intertwined, often treating Compute, Investment, and Infrastructure as a single underlying resource cluster rather than entirely separable levers. As a result, the somewhat uniform distribution of weights may reflect uncertainty about which factor will ultimately prove to be an enabler of blocker.

Rationale analysis:

Enablers of AGI

• General: Forecasters overwhelmingly agree that the eight enablers are deeply entangled, making clean allocations difficult. As one writes: “They’re all super correlated… For instance, you don’t have compute without investment, and investment doesn’t matter that much on its own. It matters because it buys you compute and generates better algorithms, and you can get better data and so on.” Several invoke economic O-ring logic, with one concluding the listed factors “are all O-rings and therefore, in one sense, equally essential.” This entanglement drives many forecasters toward relatively uniform allocations; those who concentrate weight typically pick whichever factor or factors they regard as upstream of the others.
• Algorithmic architecture: This is the enabler most frequently named as being singularly important, with many expressing the sentiment that AGI by 2040 “likely requires more than simply scaling current transformer systems,” and that genuine generality demands a new paradigm: “If AGI arrives by 2040, the story will be that someone cracked the architecture problem.” Few forecasters dismiss the importance of algorithmic improvements, but several view it as residing downstream of other enablers: “Important but partly a result of self-improvement and investment rather than an independent driver.”; “...crucial, but not the key ingredient as different architectures are often able to achieve similar results. Instead, a mix of data and sufficient compute are the key ingredients.”
• Compute, Investment, and Infrastructure: Forecasters often treat these three as a single underlying enabler—“scaled physical compute”—because investment funds compute and infrastructure enables it: “Compute, investment, and infrastructure are tightly coupled, so I don’t know how to break it up between those.”; “‘Investment’ and ‘compute’ and ‘infrastructure’ are all overlapping.” Several who reference Richard Sutton’s Bitter Lesson [i.e., compute beats cleverness] concentrate weight here. Critics tend to argue that while “scaling compute has clearly driven progress so far,” that we may be bumping up against the limits of future progress given that “the compute, investment, and infrastructure enablers are (1) already well resourced with $700B in investment this year and (2) focused on the current paradigm of GenAI/LLMs,” and that going forward algorithmic breakthrough will likely “carry far more weight.”
• Real-World interaction and data: Forecasters often combined these, with one writing, “real world interaction is just a way to generate the data,” and another adding, “I had a lot of trouble knowing what to count as ‘data’ versus ‘real-world interaction’ that accumulates data.” Forecasters allocating heavily here tend to emphasize that “the web is largely mined” and therefore “without continuous interaction with the world, models remain trapped within the distribution of human-generated text—which is precisely the model-collapse risk that makes purely text-based training a long-run dead end.”
• Self-Improvement. Many forecasters view recursive self-improvement as a compounding meta-enabler. As one puts it, “once AI systems can meaningfully improve their own training pipelines, evaluation, architecture search, and tool-use, every other bottleneck (compute, data, architecture) becomes endogenous rather than exogenous.” Skeptics push back, with one arguing that self-improvement “is bottlenecked not by ideas, but by compute,” and another that it “comes after AGI, not before.”
• Embodiment. Embodiment received consistently low allocations because the definition of AGI is restricted to non-physical occupations: “I do not see why embodiment would matter at all given the AGI definition focus on non-physical occupations.”; “Embodiment low because AGI definition excludes physical tasks.”

Blockers of AGI

• General: As they did when considering enablers of AGI, forecasters frequently note that the listed blockers overlap and interact. “All factors seem reasonable, and none seem irrelevant,” was a common sentiment.
• Algorithmic architecture: Frequently named as a top blocker: “If AGI doesn't happen by 2040, the most likely reason is that current paradigms (transformers, autoregressive LLMs, reasoning RL on top of them) hit fundamental ceilings that scaling alone cannot overcome.”; “However much computing power, money and various types of data you have, if you don't have the architecture you won't reach AGI.” Few dispute that this blocker could play a key role, but many argue it is not likely to be the primary blocker: “I'm far less skeptical of technical barriers now. The industry seems to just scale through everything.”; “Our algorithms [are] good enough.”
• Scaling limits: Many tied this closely to algorithms, with one arguing that scaling limits are “somewhat downstream of the architectural problem, which would create the scaling law limit in the first place.” Several, however, expressed the sentiment that this blocker was independently crucial, with one respondent writing that if “returns to compute and data scaling diminish before AGI-level capabilities, no amount of additional resources will deliver a qualitative leap.”
• Political/social barriers: Frequently cited as the most formidable non-technical hurdle, many emphasize the potential for backlash to AI—”job loss could very easily lead to social and political outrage”—and argue that the primary barrier to AGI “is increasingly not raw technical capability but the legal, regulatory, and societal environment in which Al is deployed.” But a sizable minority question the ultimate power of this blocker. One respondent writes, “Politics is too inept to stop anything I think, no matter what they do,” and another that “political, social, and funding barriers are less of a problem, as strong financial incentives and competition between organizations keep progress moving forward.” Echoing the latter sentiment, another respondent writes that political barriers “don’t typically prevent a sufficient technology from emerging” but instead “shift its location, timing, and deployment.”
• Learning/memory: Many highlight this factor as crucial if agentic persistence and, by extension, AGI is to be realized. As one respondent writes, “Memory is a serious blocker. Without the ability to hold and update what it has learned, no system can build on its own experience.” Another points out that “genuine continual learning without catastrophic forgetting, episodic memory, and persistent skill acquisition remain unsolved at frontier scale. Others see this issue as likely to be tractable: “learning and memory seems to be an obstacle that is currently being overcome (e.g. context windows increasing dramatically).”
• Data: While a few respondents treat data as the core barrier—”I continue with the mental model that it is all about the data, essentially, and everything else (algorithms, compute-scaling...) is ‘just’ an optimization”—others argue “we are likely to scale with synthetic data and virtual environments.” (Although one respondent warned that “synthetic data and self-play help but carry model-collapse risk.”) Some link better training data “directly to embodiment, where robotics becomes the bottleneck.”
• Embodiment: Typically rates low because FRI’s AGI definition is non-physical: “Embodiment remains at 0% as a blocker. For this definition of AGI, physical interaction is not required.” A minority, however, argue physical grounding is necessary “even for cognitive tasks” with one writing that the failure of robotics to keep pace with language models is “the single most plausible technical reason AGI is delayed.”
• Investment/infrastructure: Some respondents note infrastructure is downstream of investment, and a failure in one reflects a failure in the other. Both are rarely mentioned as primary blockers, with one respondent writing “capital is currently abundant,” and another that “we’ve not yet seen investment slow down so it’s hard to consider that a limiting factor at this point.” A few respondents, however, point to the potential for a geopolitical shock to quickly raise the salience of this blocker. As one writes, “It is certainly possible that some shock, such as Taiwan invasion, leads to a collapse of infrastructure and funding.”

Rationale examples, AGI enablers:

How do you count the influence of a factor that can totally enable or disable another one? It is like the classic misunderstanding of the degree of importance of the energy in the economy: 5% of expenses, but enable 100% of the economy—without energy no modern economy.

The ‘but-for’ framing matters for the panel allocation because the question is implicitly causal—‘how much did this factor enable AGI' means ‘if you took this factor away, how much later/less likely would AGI be.’ Compute appears to be the factor whose removal most clearly stalls progress: take compute away and no architectural innovation saves you; take any specific architectural innovation away and more compute typically substitutes. The asymmetry argues for crediting compute most heavily…

If AGI arrives by 2040, the story will be that someone cracked the architecture problem…the transformer paradigm is impressive, but it's not getting us to genuine generality on its own. The systems that close the gap will look meaningfully different under the hood…the ability to reason rather than just predict.

Reaching AGI by 2040 likely requires more than simply scaling current transformer systems. Even if scaling remains important, most experts will probably view breakthroughs in reasoning, memory, causal/world modeling, planning, and learning paradigms as central to overcoming the limitations of current approaches.

All elements will almost certainly play some role, and while in the end one may stand out as singularly determinative, it's more likely that it will be a subjective call, and that there also could be a last-person-in-the-room effect on the panel—i.e., even if other elements contribute equally to AGI, the element that finally pushes it over the resolution bar could be the one that panelists point to as having enabled AGI the most.”

Rationale examples, AGI blockers:

If AGI isn't here by 2040, the most honest explanation is that we hit a wall we didn't know was there. Algorithmic architecture gets the most weight because (I believe) the ceiling on current approaches is closer than the field currently believes.

Political/Social Barriers—I see this as low as most countries are racing towards AGI. Any country resisting this will just give up its lead to another country. A coherent worldwide effort is needed here, which is highly unlikely.”

The main barrier will be political/social… Many forms of service-sector jobs are essentially algorithmic intermediaries that exist for compliance or record-keeping reasons (e.g., human resources, accounting). The number of these jobs will be gutted as AI/LLM enables more productivity per worker. In the past, blue collar jobs like steelworkers were disintermediated—those workers were given token assistance and such displacement was shrugged off by the political establishment as the cost of progress (“learn to code”). And now given: (1) the potential for large swathes of even 1st-tier college graduates to become disintermediated by an algorithm; and (2) extant historically high levels of wealth and income inequality, these are the conditions for **revolutionary** social and political changes in the developed and developing world—not necessarily Cosette barricading the streets of Paris, but definitely the rise of political movements and figures who are outside of the current Overton Window from all non-Establishment vectors (i.e., left, right, anarchic, libertarian, etc.)”

Society has had less than 3 years to think about AI and its implications. Currently, the majority of the populace that is aware of AI considers it a new ‘toy’ technology. With Alexa in our kitchen and Siri on our phone it is a novelty that is non-threatening. The U.S. Congress appears unconcerned about AI, the same way they view climate change. As AI enters more of the workplace and augments more jobs, it may become apparent that workplace hiring [will stall] as employers evaluate AI capabilities. Some industries may be able to radically reduce employment due to AI efficiencies. The sooner this occurs the sooner policy makers will feel forced to act, even if the data is incomplete or wrong. Barriers are easy to erect, but costly to dismantle. Therefore I view this as the major threat to AGI.

Question (I). At the end of 2040, how will the LEAP panel allocate points across the following factors enabling AGI?

Question (II). At the end of 2040, how will the LEAP panel allocate points across the following factors blocking AGI?

80%, 8-Hour Task Horizon

Question. What will be the longest METR 80% time horizon listed for an AI model on December 31, 2026? In what year will an AI model be able to achieve 80% success on tasks which require 8 hours or more of human expert effort?

Results. The median forecaster across all groups expects the longest METR 80% time horizon by the end of 2026 will be over 3 hours (3.4, 3, and 3.5 for experts, the public, and superforecasters respectively), up from the baseline of 1.5 hours by Gemini 3.1 (Pro) on METR at the time of this survey. Experts and superforecasters assign a 25% chance of the time horizon reaching roughly 2.1 hours and a 75% chance of it reaching more than 5.5 hours by the end of 2026. On May 8, 2026, toward the end of the survey, METR updated the benchmark to include Anthropic’s Mythos preview model, with an estimated 80%-time horizon of 3 hours 6 minutes (95% CI: 1h 37m – 6h 39m), already in the range of the median forecast.⁵

Participants diverge on when they forecast AI will achieve 80% success on 8-hour tasks. Superforecasters are more optimistic and assign a 50% probability that this threshold will be reached by 2028 and a 25% probability by as early as 2027. Experts are somewhat more conservative, expecting this to occur around 2030 at the median, with a 25% chance by 2028. The public is considerably more skeptical, with a median forecast of 2037 and only a 25% probability of resolution by 2030. The 75th-percentile expert forecast of 2035 roughly matches the public's median (2035 vs. 2037).

Question (I). What will be the longest METR 80% time horizon (in hours) listed for an AI model on December 31, 2026?

Question (II). In what year will an AI model be able to achieve 80% success on tasks which require 8 hours or more of human expert effort?

Rapid Progress Scenario Timelines

Question. What is the probability that, before 2100, more than 50% of LEAP panelists (or a similar panel) will choose “rapid progress” as best matching the general level of AI progress? Assume that more than 50% of LEAP panelists choose the “rapid progress” scenario before 2100. In what year do you expect this to first occur?

Results. Most participants across all groups expect that rapid AI progress will be recognized before 2100. Domain experts assign a median likelihood of 68%, superforecasters are more optimistic at 76%, and the general public is considerably less confident at 55%. Both experts and superforecasters are significantly more optimistic than the public. Assuming that a majority of LEAP panelists choose the Rapid Progress scenario before 2100, there is convergence between experts and the public at the median: both groups forecast this would first be recognised around 2045, while superforecasters are somewhat later at 2050. Across higher percentiles, uncertainty expands considerably, with 95th percentile forecasts extending into the 2080s and 2090s. Differences between experts and superforecasters are statistically significant across most percentiles, with superforecasters consistently skewing later.

Rationale analysis:

• Time horizon: Fast-progress respondents typically cite the 2100 timeline, arguing 74 years provides ample time for massive, compounding technological leaps: “74 years is enough time for multiple further cycles of AI winter and summer, if necessary.”; “2100 is far out with these exponential model improvements, so moderate progress is baked in and rapid likely.”; “These are all things likely to be surpassed in the next few years, let alone 74.” Slow-progress respondents tend to point to the potential bottlenecks discussed below (robotics, architectural sufficiency, the popping of an AI investment bubble, and societal backlash) as reasons to remain skeptical. “Most breakthrough technologies eventually encounter diminishing returns,” writes one, and another that “the current pace of AI progress is in many ways a product of a specific convergence of available training data, transformer architecture scaling, and an unprecedented influx of capital. I expect at least one of these constraints to bind meaningfully before [2100].”
• Robotics: This is the most cited reason slow-progress respondents give for why they think progress will not be deemed rapid. As one writes: “The robotics component of the rapid scenario is a pretty high bar. Seeing that I have significant uncertainty that the coffee test will ever be achieved by 2100, adding all these autonomous robotic capabilities makes it all that much unlikely.” Another adds, “The whole ‘fix any plumbing issue in any arbitrary home anywhere in the world’ seems too high a barrier on a physical, not intelligence level.” Fast-progress respondents tend to argue that breakthroughs in digital cognition will inevitably translate to the physical world, and they also cite recent manufacturing and scaling successes. One writes: “Waymo robotaxies are already reported to have 13x fewer serious crashes than humans through December of last year. There are significant advancements in humanoid robots, with Figure scaling 24x in production in less than four months and Optimus, Unitree, and 1X all advancing.”
• Architecture sufficiency: Most fast-progress respondents project current pace of gains forward, implying that the incremental improvements to the current models will be sufficient: “This year has seen rapid progress, so unless we see a collapse in scaling laws, I’m fairly confident we’ll get to this level of response by 2100.”; “AI capabilities are rising e,g. METR benchmarks, Frontiermath… If the rate of progress increases, it can likely pass as rapid progress in the next 10 years.” Slow-progress respondents tend to view that rate of progress as unlikely to continue: “To achieve rapid progress, AI models need to be able to surpass human performance on nearly every task. This will require significant technological breakthroughs, and should not be achievable using methods that we have right now.”; “More of, and faster, versions of the current AI will not give us ‘Reliable, Rapid progress’. So to increase the percentages this question depends on different techniques emerging.”
• Societal backlash: Several slow-progress respondents point to human resistance to AI deployment as a potential obstacle. One writes that “regulatory barriers could prevent the scenario from being realised,” and another that “the rapid progress seems too good to be true, humans will not let it happen…legislators or politicians will not let it happen.” Fast-progress respondents, when they engage with this factor, tend to treat backlash as a small tail risk against an otherwise overwhelming default. As one writes: “Hard to imagine we don’t have these capabilities. Only blockers are butlerian jihad, pandemic, nuclear war, etc…”
• Investment trajectory: Fast-progress respondents generally view capital and geopolitical competition as a durable engine of progress: “There are large incentives both in the private sector (profit and productivity increases) and public sector (geostrategic advantage) that make it likely that there will be a continued focus on AI progress…” writes one. Some slow-progress respondents instead view the current funding surge as likely to be transient. One writes: “I believe the AI bubble is reasonably likely to burst within the next 5-10 years, at which point I think we will be disabused of notions of rapid progress.”
• Creative thresholds: Fast-progress respondents often view Pulitzer/Grammy-caliber work from AI as achievable, with one writing, “I anticipate Al crossing this threshold within the next decade i.e., being indistinguishable in a blind test from world leading human literary and musical art.” Slow-progress respondents often emphasize the sociocultural element of these thresholds, with one noting, “The quality of creative output is socially defined, not by some objective metric, and I do not believe AI artists to have a lively social following.”
• Soon vs. mid- to late-century: Fast-progress respondents often argue that, conditional on resolution, the event will happen early given current capability momentum and the recent hype cycles: “If we see rapid progress, we’ll see it sooner rather than later."; “This question is related to the current concept of AI advancement…expecting higher intensity efforts will come upfront in the 20s, 30s and 40s, with efforts likely diverted elsewhere in the later years of this century.”; “By 2030 most of the current CAPEX data center build out will be complete, providing inexpensive inference across a continent. Improvements in inference engines will also support robotaxis and humanoid robots with appropriate computing power for autonomous behavior.” Slow-progress respondents typically point toward a mid- to late-century resolution date, citing the likelihood of slower than anticipated change on a variety of fronts. One writes: “Due to the amount of societal change required, if this happens at all, I would expect it to be towards the end of the time frame.” Another: “Slower timelines here than for the AGI questions because this definition involves a lot of progress on interaction with the physical world.” A third focuses on the LEAP panel: “The delay comes from the fact that this is not a technical threshold. It’s a social consensus problem. The panel includes people from domains that move slowly, people who are cautious by training, and people who will require overwhelming evidence before they label something as rapid.”

Rationale examples, fast-progress respondents:

This is overwhelmingly likely—I am in one of these jobs [that] will easily get replaced within the next 74 years, the speed of progress would have to be ridiculously slowed down for this not to be true—especially as AI will generate better and better data to train itself upon, like humans did.

If we go back 74 years we land in 1952, not far off my own birthday. In 1952 how many people would have thought we would have supercomputers in our pockets, high-speed fiber optic communications, the internet, hundreds of gigawatts of solar power, electric cars that travel 300+ miles on a charge, or artificial intelligence that can answer complex questions? By 2100 the world will likely be completely different in ways we cannot imagine. Rapid progress in AI is likely to be among the advances.

My 90% reflects a near-certain expectation that the 'Rapid Progress' scenario will materialise in capabilities terms, combined with high confidence that the LEAP panel will eventually recognise this given a 74-year window…I assign negligible probability to a scenario in which rapid progress genuinely occurs but the panel refuses to acknowledge it, because the scenario's criteria (autonomous researchers compressing years into days, AI surpassing the best humans in cognitive work, Level-5 robo-taxis) are sufficiently concrete that sustained denial would become untenable within a few years of their realisation.

If we are going all the way out to 2100, then we are mostly asking if the rapid progress scenario is possible at all, given what we can assume about increases in wealth (and therefore available investment capital), further algorithmic progress, and a deeper talent pool of bright young graduates focused on AI than ten or even five or really even two years ago. I expect that over the next few years coding will be a proof of concept that, yes, AI can do labor that humans consider very cognitively advanced. Autonomous vehicles have taken a long time, but are about to be deployed broadly, which I would take as a sign that eventually we should be able to make any space navigable for robots.

Rationale examples, slow-progress respondents:

The technical requirements of the rapid progress scenario are just so high that it seems extremely unlikely. Further, the marketability / profitability of so many of these applications seems so low that I don't believe the investment and funding would be made available to develop them. I understand that we're not taking into account adoption of the products, but potential demand does drive investment, which drives innovation.

Rapid seems like a far more advanced version of what's defined under AGI previously. If the probability for AGI is 70% before, then rapid I think is lower than that. It would not just require an AI system that is cognitively autonomous, but also superior (ie: medicine breakthroughs, novel ideas), that far exceeds a normal human being.

Rapid progress implies way too "high-end" tasks for AI to perform, plus the rapid progress concept may be pushed forward so panelists won't consider the current statement as rapid as it should be. Also, I wouldn't be surprised if AI progress slows down due to external factors such as energy shortages or loss of interest among humans.

...tail-end scenarios [could] curtail development way beyond expectation (war, natural disasters, authoritarian enforcements of AI research moratoriums).

Question (I). What is the probability that, before 2100, more than 50% of LEAP panelists (or a similar panel) will choose “rapid progress” as best matching the general level of AI progress?

Question (II). Assume that more than 50% of LEAP panelists choose the “rapid progress” scenario before 2100. In what year do you expect this to first occur?

AI Impact

Question. What do you think that the impact of AI on the U.S. over the next 20 years will be? How do you think the increased use of artificial intelligence (AI) in society will impact people’s ability to do each of the following?

Results. Across all participant groups, optimism about AI's overall impact on the US over the next 20 years outweighs pessimism, though the degree of optimism varies between groups. Experts and superforecasters are the most positive: 58% of experts and 70% of superforecasters rate the impact as somewhat or very positive, compared to 42% of the general public. Conversely, the public displays a more pessimistic outlook, with considerably higher shares expecting negative consequences (30%, vs. 17% for experts and 11% for superforecasters) and fewer expecting very positive outcomes (13%, vs. 19% and 23%). Across all groups, roughly a quarter anticipate equally positive and negative impacts, with this figure slightly lower among superforecasters (19%).

Views on AI's psychological and social impacts diverge across dimensions. On problem-solving, there is broad optimism among experts and superforecasters: 73% and 70% respectively believe AI will make people better at solving problems. However, the public is meaningfully less positive at 49%. The same pattern is sharper for making difficult decisions, where 58% of experts expect improvement, compared to 42% of superforecasters and just 37% of the public; 39% of the public expect people to become worse at this. By contrast, all groups expect AI to make people worse at forming meaningful relationships, with 68% of experts and 67% of the public sharing this view, while superforecasters are less pessimistic at 51%. Creative thinking draws the sharpest divergence: 67% of the public and 50% of experts expect AI to worsen this capacity, while only 38% of superforecasters expect a negative effect and 40% expect no change.

Question. What do you think that the impact of AI on the U.S. over the next 20 years will be?

Question (II). How do you think the increased use of artificial intelligence (AI) in society will impact people’s ability to do each of the following?

Footnotes

1. Specifically, we asked participants “What is the probability that, before 2100, more than 50% of LEAP panelists will agree that AGI exists, and if it occurs, in what year do you expect this to first occur?” ↩

2. A nationally representative survey by the Annenberg Public Policy Center (Feb–Mar 2026) found only 17% of U.S. adults expect AI's impact over the next 10 years to be positive, suggesting LEAP's public sample is itself more AI-positive than the broader U.S. public. ↩

3. In some cases, the "aggregate" refers to the mean; in others, the median is used, depending on which is more appropriate for the distribution of responses. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

4. We occasionally elicit participants' quantile forecasts (estimates of specific percentiles of a continuous outcome) to illustrate the range and uncertainty of their predictions. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

5. Among the 20 experts and superforecasters who submitted on or after May 8, median 50th percentile forecasts were 0.4 hours higher than those submitted earlier (3.9h vs 3.5h), with smaller upward shifts at 25th percentile (2.4h vs 2.1h) and no movement at 75th percentile (5.5h for both). However, none of these differences reached statistical significance, and thus we believe that this new information has not meaningfully impacted the aggregate participant forecasts. ↩