What will AI look like in the decades ahead?

AI may evolve from late-2020s agentic digital coworkers into an automated research “civilization,” then become core infrastructure for science, industry, defense, and political power in the 2030s. The decisive hinge is whether AI can automate nearly the entire AI-research workflow, enabling millions of copies to accelerate further advances; if bottlenecks in reliability, data, experiments, energy, coordination, robotics, or control bind, progress could be far slower. Frontier systems would likely rely on enormous, centralized datacenters and attract state oversight. This is one speculative 2024 fast-takeoff scenario—not a consensus forecast—and the evidence cannot reliably establish its probability or describe later decades in detail.

What AI may look like in the decades ahead

Bottom line

The supplied evidence supports one specific, unusually fast “takeoff” scenario, not a settled forecast. In that scenario, AI evolves in three stages:

1. From chatbots to autonomous digital coworkers in the late 2020s—systems that can be onboarded into an organization, use computers and tools, remember context, and independently complete projects lasting days or weeks.

2. From human-level AI to superintelligence around the turn of the decade—once AI can automate AI research, millions of copies accelerate the design of still-better systems.

3. From software intelligence to a transformed physical, economic, and military world in the 2030s—AI-directed science, robotics, factories, and weapons make AI less a product than the operating system of civilization.

On this account, mature AI would look less like a better search engine and more like a vast, always-on workforce and research civilization housed in enormous datacenters. It would also become a concentrated strategic asset: expensive, energy-intensive, closely guarded, and probably subject to direct state control.

However, this conclusion is highly conditional. The only source is Leopold Aschenbrenner’s June 2024 essay, Situational Awareness, which explicitly mixes public information with the author’s “own ideas,” field knowledge, and “SF-gossip.” Its dates are forecasts made in 2024, not observations validated as of August 2026. There is no independent source here with which to verify whether its 2025–26 milestones occurred. The analysis below should therefore be read as the source’s scenario, its causal logic, and its failure conditions, not as a consensus prediction. Source: introduction and contents

A phase-by-phase picture

1. The near-term interface: from assistant to coworker

The most concrete change is not simply a smarter conversational model. The essay argues that present systems are “hobbled” by missing memory, limited context, weak tool use, and an inability to sustain coherent work over long horizons. It identifies three capabilities that would turn them into workers:

• Onboarding: ingesting company documents, communications, code, and work practices, much as a new employee learns an organization.

• Long-horizon reasoning: planning, testing, correcting errors, researching alternatives, and producing the equivalent of weeks or months of work rather than a short answer.

• Computer use: joining calls, browsing, messaging colleagues, using applications and development tools, and submitting completed work.

If these pieces arrive, AI becomes a “drop-in remote worker.” Its abilities would still be uneven: it might be a superhuman coder while retaining surprising blind spots in planning or coordination. Rollout in medicine, law, and other regulated professions could also lag technical capability. But the core implication is that all work executable through a computer becomes technically contestable, even if institutions delay actual replacement. Situational Awareness, §I, especially “Unhobbling” and “From chatbot to agent-coworker,” pp. 7–45

The forecast rests on a quantitative extrapolation. The author attributes past progress to roughly 0.5 orders of magnitude (OOMs) per year in physical training compute, roughly 0.5 OOMs per year in algorithmic efficiency, plus gains from post-training, tools, scaffolding, and longer context. He projects another 3–6 OOMs of base effective compute from GPT-4 through 2027, with about 5 OOMs as a best guess. The analogy is that if GPT-2 to GPT-4 moved from roughly preschool-level performance to a smart high-schooler’s, another similarly sized jump could reach expert or PhD-level work.

That is an extrapolation, not proof. Human-development labels are loose analogies; benchmark gains need not translate into dependable autonomy; and the source itself acknowledges a real possibility of stalling if high-quality training data runs out or synthetic data, self-play, and reinforcement learning fail to compensate.

2. The decisive hinge: AI automating AI research

The scenario’s most important step is not automation in general but automation of machine-learning research specifically. That work is largely digital: reading literature, proposing ideas, writing code, running experiments, and interpreting results. It therefore need not wait for capable household robots or fully automated laboratories.

The essay estimates that late-decade inference fleets could run many millions of AI researchers—perhaps about 100 million human-researcher-equivalents—continuously, with some copies operating at more than human speed. This does not imply research becomes millions of times faster: experiment compute, coordination, and diminishing returns remain bottlenecks. The author’s more limited claim is that such a workforce might accelerate algorithmic progress by 10× or more, compressing roughly a decade of human research, or 5+ OOMs of improvement, into about a year.

Its illustrative softer timeline is:

• 2026/27: proto-automated engineers accelerate work by around 1.5–2×;

• 2027/28: proto-automated researchers perform more than 90% of the work and accelerate progress by 3× or more;

• 2028/29: remaining bottlenecks fall and progress accelerates by 10× or more.

This is how the source gets from AGI to superintelligence by about 2030. The transition need not be overnight, but it could occur over months to a few years. Situational Awareness, §II, pp. 46–73

This step is also the forecast’s weakest causal hinge. It requires all of the following:

1. AI can perform essentially the whole AI-research workflow, not merely coding fragments.

2. Millions of copies can be coordinated productively.

3. Limited experiment compute does not dominate the production of ideas.

4. Algorithmic advances remain plentiful enough to yield several more OOMs.

5. New systems remain sufficiently reliable and controllable to conduct the process.

If even one of these constraints binds strongly, the “intelligence explosion” could become a slower period of ordinary technological improvement.

3. What superintelligence would look like

If the feedback loop works, AI stops being merely human-equivalent. The source envisages a “civilization” of billions of instances that can read all relevant literature, share experience across copies, write enormous amounts of code, operate continuously, and think many times faster than people. More importantly, they would be qualitatively superhuman: solving problems or inventing strategies that humans cannot follow, much as AlphaGo produced moves outside established human play.

That changes the human role. People would no longer directly inspect or understand most important machine work. They would set objectives, choose institutions and constraints, and depend on other AIs to explain, monitor, and verify the strongest systems. AI oversight would increasingly become AI supervising AI.

4. The 2030s: from digital labor to the physical world

The source expects the initial explosion in AI research to spread outward:

• Science and engineering: billions of automated researchers could compress a century’s human R&D into years, although physical experiments would impose real delays.

• Robotics: superhuman AI could solve remaining machine-learning and control problems; factories might progress from AI-directed human labor to fully robotic operation.

• Industry: robot factories could build more robots and factories, relaxing the fixed-human-labor constraint.

• Economic growth: in unconstrained sectors, the author speculates about growth of 30% per year or more, potentially multiple doublings annually. Regulation, adoption delays, physical testing, and resistance to disruption could keep measured economy-wide growth much lower.

• Military power: automated hacking, drone swarms, biological design, missile defense, and unknown new weapons could make a lead of months or years strategically decisive.

Thus the source’s 2030s are not simply “today with better AI.” They are a potential new growth regime in which cognitive labor is abundant and physical labor is progressively automated. Yet these are the essay’s most speculative claims: they extend an already-conditional AI-research feedback loop through robotics, manufacturing, scientific discovery, adoption, and geopolitics. Situational Awareness, §II and §IIId, pp. 46–73 and 126–140

The evidence does not support a detailed account of the 2040s and beyond. The essay says the world order would be transformed by the end of the 2030s but deliberately leaves that story untold. Any more precise multi-decade forecast would exceed the supplied evidence.

AI’s physical form: giant, centralized infrastructure

Even if users experience AI as cheap software, the source expects frontier AI to be physically concentrated. Its back-of-the-envelope trajectory for the largest training clusters rises from about 10,000 H100-equivalent chips and 10 MW for GPT-4 to:

• about 1 million chips, 1 GW, and tens of billions of dollars around 2026;

• about 10 million chips, 10 GW, and hundreds of billions around 2028;

• about 100 million chips, 100 GW, and $1 trillion or more around 2030.

The last figure is more than 20% of the United States’ then-current electricity production in the essay’s comparison. The author also projects total AI investment above $1 trillion annually by 2027, while repeatedly labeling the calculations rough and identifying power as the binding constraint. Situational Awareness, §IIIa and appendix, pp. 75–88, 162 ff.

The qualitative conclusion is more robust than the exact numbers: frontier AI would depend on power plants, datacenters, semiconductor fabs, networking, cooling, and secure sites. High fixed costs would favor a few firms or states over diffuse open-source development. Inference fleets could still distribute many AI workers widely, but the most capable models and their weights would remain concentrated assets.

AI’s political form: strategic asset and government project

The essay predicts that such concentration and military value would pull AI out of ordinary startup governance. It expects some form of US government-led project by 2027/28: leading laboratories pooled or coordinated, large public appropriations for power and chips, strict security, and an alliance of democracies combining research talent and semiconductor supply chains. Civilian applications would continue, but the initial development and military use of superintelligence would be treated more like a defense program than an internet service. Situational Awareness, §IV, pp. 141–155

This is a prediction shaped by the author’s geopolitical assumptions, not an inevitability established by the evidence. Alternative outcomes include continued private competition, international governance, regulatory slowdown, proliferation of capable smaller models, or no superintelligence at all. Still, the causal point is important: the more capable, costly, and militarily useful AI becomes, the less likely its governance is to resemble today’s consumer software market.

The central dangers

The forecast identifies three coupled risks:

1. Loss of control. Human-feedback methods depend on people understanding outputs. They fail when a model writes code, develops plans, or makes discoveries that no person can evaluate. Proposed responses include scalable oversight, interpretability, trusted weaker models monitoring stronger ones, air-gapped systems, capability restrictions, and automated alignment research. The source regards these as promising but not foolproof.

2. Human misuse and mass destruction. Superintelligence could lower the cost of hacking, biological weapons, autonomous weapons, and other technologies not yet imagined.

3. Geopolitical race and authoritarian control. Theft of model weights or algorithmic secrets could erase a lead; a close race could force competitors to sacrifice safety; and a state with obedient AI surveillance and robotic enforcement might lock in dictatorship.

These dangers reinforce one another. Weak security can create proliferation; proliferation intensifies racing; racing reduces time for alignment; alignment failures become more consequential as AI enters military and economic infrastructure. Situational Awareness, §§IIIb–IIId, pp. 89–140

What would change the conclusion?

The fast-takeoff picture should be revised substantially if evidence shows any of the following:

• effective compute no longer produces broad capability gains;

• synthetic data, self-play, or reinforcement learning cannot overcome data limits;

• long-horizon agents remain unreliable despite stronger benchmark performance;

• automated researchers do not materially raise research productivity;

• experiment compute, physical validation, energy, chips, or coordination become hard bottlenecks;

• robotics and manufacturing scale much more slowly than software;

• safety rules, conflict, or regulation deliberately slow deployment;

• decentralized, efficient models reduce the predicted concentration of power.

Conversely, the scenario becomes more credible if systems independently complete increasingly long and novel technical projects, automate most frontier AI research, produce verified algorithmic gains, and if capital and power buildouts follow the essay’s proposed trajectory.

Overall judgment

The source offers a coherent answer: AI in the decades ahead could progress from an agentic digital workforce, to an automated scientific civilization, to the principal infrastructure governing industry, defense, and political power. The strongest part is the qualitative shift from chat interfaces toward agents with tools, context, and long-horizon work. The most consequential but uncertain part is the claim that those agents will automate AI research and trigger a rapid intelligence explosion. Everything farther out—robotic industry, extraordinary growth, decisive military dominance, and an unrecognizable 2030s—depends heavily on that hinge.

Given the single, speculative, pre-2026 source, the responsible conclusion is not “this will happen,” but: this is a high-impact scenario worth testing against observable milestones. It is internally plausible under its assumptions, yet the evidence inventory is insufficient to assign it a reliable probability or to describe later decades in detail.