AI Timelines and How They Inform FLF’s Plans and Work

Our Assessment

The Future of Life Foundation believes that AI systems with transformative impacts could be months to years away, rather than years to decades. This is an importantly contingent claim: we are convinced that the trends and potential are there, but not that they are inevitable. And the impacts of the resulting technology on society look concerning or even catastrophic by default — but could be steered onto better paths.

The Evidence

Steady, broad-based progress

AI capabilities are advancing by nearly every assessed metric, and there is no evidence of AI “hitting a wall.” While particular approaches may reach diminishing returns, the enormous financial and intellectual effort being poured into AI development has consistently opened new directions that sustain progress — in close analogy to Moore’s law, but arguably faster.

AI development and diffusion rest on computing capacity (compute). Global AI compute is doubling approximately every seven months — a trend driven by underlying efficiency growth in compute production (Moore’s Law) and compounded by AI developers’ increasing willingness to spend on compute stock. Contributions from algorithmic and data improvements further compound AI progress (by an uncertain amount). Hence, the frontier of AI capability continues to quickly advance, albeit unevenly, across a highly general spectrum of tasks.

Autonomous capability is scaling rapidly

One of the most striking trends is the growth in the length and complexity of tasks AI systems can complete autonomously. METR, a nonprofit that evaluates AI agent capabilities, tracks the effective hours of human-expert-equivalent work that AI agents can reliably complete end-to-end.

In software tasks, this metric has been growing exponentially (or perhaps even faster), repeatedly doubling on a months-level timescale. Concretely: in early 2024, frontier AI could, more often than not, handle software tasks taking a skilled human about four minutes. By early 2026, that figure is closer to twelve hours — a full workday’s-worth of autonomous software engineering. On this trend, week-long and then month-long autonomous software tasks could arrive within the next few years.

Near-superhuman performance in verifiable domains

Progress is particularly fast in domains where correct answers can be objectively and automatically verified, because this enables rapid improvement through reinforcement learning and valid synthetic training data. AI systems are now routinely solving real-world software engineering problems that would take professionals hours. In mathematics, problems that no model could touch in late 2024 are now being solved at rates approaching or matching human expert teams, with substantially superhuman performance likely on the near horizon. Cyber offense/defense and formal verification show similar dynamics.

This verifiability advantage is an important pattern: where correctness can be checked cheaply and automatically, AI self-improvement loops gain traction fastest. Epoch AI notes explicitly that the recent acceleration in capabilities is concentrated in these domains — an important caveat, but also a pointer to where transformative capability is likely to arrive first.

Known weaknesses are real, but may not be lasting

Current AI retains several weaknesses relative to the frontier of human cognition, including:

Effective long-term memory and retention of newly learned skills and information across contexts
Fully general problem-solving — especially where physical interaction with the world or accurate visual-spacial awareness is key
Learning and knowledge in domains with limited or slow feedback (such as medical innovation), because experience and reinforcement learning are more difficult to accrue
Honesty and transparency, which in AI can be beset by confabulation/hallucination and sycophancy, in significant part because they are insufficiently prioritised by training processes and developer incentives

These are real limitations which hold back both some beneficial and some dangerous applications of AI. But AI can be transformative with limited progress on these weaknesses — for example overhauling existing patterns in cybercrime or journalism (for better or worse) — and huge research resources are being directed at overcoming these barriers. We expect them to mostly recede in a small number of years, though some — particularly embodied problem-solving and learning in limited-feedback domains — may inherently proceed more slowly.

The AI improvement loop

We are now firmly in a period where AI can do some actual productive work. Today, this means AI tools and agents, ‘supervised’ by human workers, with the AI ‘share’ of work steadily (though unevenly) increasing. This is especially notable when the work involves production of successive AI systems (a readily-available application of the latest AI by the developers themselves):

Training environments and example data — curricula for learning and environments for training can increasingly be developed with AI assistance, or incorporate AI in self-play-like settings
Machine learning techniques — neural network architectures, training regimes, and ML optimisation continue to see innovation, which may be augmented by AI (and the actual experiments can increasingly be run and overseen by AI)
Scaffolding and (multi-)agent frameworks — developments in tool integrations, reasoning workflows, external memory use, and multi-agent (or multi-persona) communication and orchestration can now be designed with and by AI
‘Skills’ and ‘memories’ — AI systems working in situ encounter learnings (contextual or practical) which can be recorded and recalled, in principle by any successor (not only by the system in question)

Across all of these, writing the needed software and scripts is in many cases already accelerated by AI. As the AI contribution to each stage of this cycle grows, the cycle time itself may compress: more AI involvement means faster iteration, bringing each next improvement sooner than the last. These feedback loops already proceed without resolving all of the weaknesses of frontier AI — in fact they may hasten their resolution.

The culmination of this ‘self improvement’ loop is difficult to predict, hotly debated, and likely depends substantially on which parts of the process are ‘handed over’ to AI systems and under what conditions.

In summary

Taken together, these observations imply that in a matter of a few years we might see AI capable enough to rapidly learn and carry out any activity, limited only by compute availability and access. By combining access to software, cyber, robotics, and perhaps human assistants, this could extend into any domain, including physical activities. Regardless of whether and when these transformative thresholds are reached, highly- but inconsistently-capable AI systems are already available and will be applied in a variety of beneficial and harmful ways.

This emphasizes the importance now of grasping the unfolding situation and steering it toward widely preferred prospects rather than sitting back to see what happens.

What This Means for FLF’s Work

These timeline beliefs have direct consequences for how we choose and structure our work.

Fast moonshots and concentrated bets

If the current window for shaping AI’s trajectory is measured in months-to-years rather than decades, fast interventions have outsized effects — either by shaping key junctures now or by equipping people to effectively handle imminent subsequent windows of opportunity.

Rather than spreading effort thinly, we aim to develop a small number of major initiatives that can scale to very large impact if they succeed, or act as multipliers and inspiration for wider fields. On the way, we don’t want to be deterred by things looking difficult or improbable: we’re willing to make bets with meaningful chances of not working out. And we and our collaborators are eager to be agile, looking for learnings, adapting to what we discover as we go, and taking multiple shots on target as needed.

Second-order influence on AI development

FLF is very concerned about the rapid pace of AI advancement and the risks it poses. However, we think society will face many key questions about technological direction — some difficult to predict exactly — so we are substantially focused on building the infrastructure around AI decision-making. This work enables better overall governance and steering of AI development and broader societal-technological direction. This includes:

Coordination — helping governments, labs, and civil society communicate, commit, and cooperate more effectively around risks and priorities, especially in AI
Epistemic tools — improving the quality of information, analysis, and forecasting available to key decision-makers and the broader public
Institutional capacity — strengthening the organizations and governance bodies that will need to respond to rapidly changing AI and other tech disruption

Additionally, this work should keep more humans more informed and able to participate (‘in the loop’) for longer… which should be broadly palliative to concerns around gradual disempowerment — hopefully for long enough to navigate to a stable and acceptable trajectory.

Keeping pace with AI

Finally, we are committed to using AI tools in our own operations and understanding how they can serve broader human benefit. If AI capabilities are advancing as fast as we believe, organizations that fail to integrate these tools will fall behind in their ability to understand and respond to the technology. We aim to stay at the frontier of AI adoption — both to be more effective, and to develop practical knowledge about how AI tools can be deployed well for general human uplift.

Last updated: June 2026