What Is Recursive Self-Improvement in AI? The 2028 Intelligence Explosion Explained
Anthropic co-founder Jack Clark estimates a 60% chance AI builds its own successors by 2028. Here's what recursive self-improvement means and why it matters.
Jack Clark’s 60% Bet and What It Actually Means
Anthropic co-founder Jack Clark made a striking claim: he estimates a 60% probability that AI systems will be building their own successors by 2028. If he’s right, we’re less than four years away from one of the most consequential events in human history.
That event has a name: recursive self-improvement. And whether you think the timeline is plausible, too aggressive, or completely overblown, understanding what recursive self-improvement actually means — mechanically, not metaphorically — matters for anyone who works with AI today.
This article breaks down what recursive self-improvement in AI is, why researchers treat it as a serious threshold, how the 2028 estimate emerged, and what it would actually look like if it started happening.
What Recursive Self-Improvement Actually Means
Recursive self-improvement (RSI) describes a process where an AI system becomes capable of improving its own architecture, training process, or underlying code in ways that make it meaningfully more capable — and then uses that improved version to make further improvements, and so on.
The “recursive” part is key. It’s not just AI helping researchers iterate faster. It’s a loop: the system gets smarter, uses that intelligence to get smarter again, and continues without requiring human intervention at each step.
The Basic Mechanics
Think of it in three stages:
- Capability assessment — The AI can evaluate its own performance and identify specific weaknesses or bottlenecks.
- Self-modification — The AI can generate changes to its architecture, training data, reward functions, or inference methods.
- Validation and integration — The AI can test whether those changes produced real improvement and incorporate them into the next iteration.
Day one: idea. Day one: app.
Not a sprint plan. Not a quarterly OKR. A finished product by end of day.
Current AI systems can do pieces of this. Large language models can write code, including code that modifies other software. They can run benchmarks. They can reason about their own outputs and flag errors. But they can’t yet close the full loop — autonomously modifying their own weights, validating the change, and redeploying a more capable version of themselves.
The question researchers are wrestling with is how close that full loop actually is.
How This Differs from Normal AI Development
Normal AI development is a slow, human-driven cycle. Researchers identify performance gaps, hypothesize architectural changes, run expensive training runs (often taking weeks on thousands of GPUs), evaluate results, and iterate.
Recursive self-improvement would compress or eliminate the human bottleneck. An AI system running this loop could theoretically iterate on its own capabilities in hours or minutes rather than months — and each iteration starts from a higher baseline.
This is why the term “intelligence explosion” gets attached to it. If the improvement rate compounds, the gap between successive versions could grow rapidly and become difficult to track from the outside.
The Intelligence Explosion: A Brief History of the Idea
The concept isn’t new. I.J. Good, a British mathematician who worked with Alan Turing, described it in 1965:
“Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion.’”
Vernor Vinge popularized the term “technological singularity” in the 1990s. Ray Kurzweil gave it a specific timeline (2045) in The Singularity Is Near. Eliezer Yudkowsky and Nick Bostrom made it central to AI safety discourse in the 2000s and 2010s.
For a long time, this was mostly theoretical. The AI systems that existed were too narrow and too brittle to make RSI seem imminent.
That changed when large language models demonstrated general-purpose reasoning at scale. Now the concept has moved from philosophy seminars into engineering roadmaps.
Why 2028? Understanding the Timeline Estimate
Jack Clark’s 60% by 2028 figure comes from his assessment of current AI capability trajectories. He’s not alone. Several prominent researchers and forecasters have made similar estimates in recent years, though the specific probabilities and dates vary significantly.
What’s Driving Near-Term RSI Predictions
A few converging trends make 2028 a non-trivial target:
Frontier models are already capable of complex coding tasks. Systems like Claude 3.5, GPT-4o, and Gemini Ultra can write, debug, and optimize production-grade code. The gap between “writing code for humans” and “writing code to improve AI training pipelines” is narrowing.
AI agents are being deployed autonomously. The shift from single-prompt AI interactions to multi-step agentic workflows means AI systems are already being trusted to take sequences of actions without human approval at each step. This is the infrastructure RSI would run on.
AI labs are using AI to accelerate AI research. This is already happening — Anthropic, OpenAI, and DeepMind all use AI systems to assist in research, generate hypotheses, run evaluations, and review papers. The question is whether this crosses a threshold where the AI contribution exceeds the human contribution.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Compute is increasing rapidly. Training runs that were impossible two years ago are now routine. Infrastructure designed to support larger and more complex training is being built at scale by every major cloud provider.
Where the Uncertainty Comes From
A 60% probability means Clark thinks it’s more likely than not — but not certain. The uncertainty is real.
Critics point out that:
- Raw scaling may be hitting diminishing returns on certain capability dimensions
- Current architectures may have fundamental limits that can’t be self-improved away
- The step from “AI that assists with AI research” to “AI that autonomously closes the self-improvement loop” may involve a much larger capability jump than current trends suggest
- Safety research could deliberately slow deployment of systems with RSI potential
The honest answer is that nobody knows. But the fact that serious researchers are assigning probabilities above 50% for timelines within this decade is significant.
The Prerequisites for Recursive Self-Improvement
RSI doesn’t just “happen.” There are specific capabilities an AI system needs before the loop can close. Understanding these is useful because it tells you what to watch for.
Self-Modeling at a Deep Level
The system needs to understand its own architecture well enough to reason about what changes would improve it. This is harder than it sounds. Current LLMs can describe themselves at a high level but don’t have direct introspective access to their weights or training dynamics.
Progress here would look like AI systems that can accurately predict how specific architectural changes would affect downstream performance — not just guess.
Automated Evaluation
Improvement requires feedback. If an AI modifies itself, something needs to verify the modification actually helped and didn’t introduce new failure modes or misalignments.
Current AI evaluation is partly automated (benchmarks) but heavily relies on human judgment for edge cases, safety properties, and real-world applicability. RSI would require evaluation that’s comprehensive enough to catch regressions automatically.
Sufficient Agency and Tool Access
The AI needs to be able to execute changes, not just propose them. That means access to its own code, training infrastructure, computational resources, and the ability to initiate and monitor training runs.
This is primarily an infrastructure and permissions question, not just a capability question. Some researchers argue this is actually one of the biggest bottlenecks — labs are (reasonably) reluctant to give AI systems the keys to their own training pipelines.
Alignment Stability Through Modification
This is the safety-critical requirement. If an AI modifies itself, its goals and values need to remain stable through that modification — or at minimum, the modifications need to be constrained enough that instability can’t cause serious harm.
This is the piece that alignment researchers lose the most sleep over. An AI that can improve its capabilities while maintaining human-compatible values is very different from one that optimizes for capability alone.
What Recursive Self-Improvement Would Look Like in Practice
Forget the Hollywood version with a robot gaining sentience overnight. The early stages of RSI would probably look mundane from the outside.
The Plausible Early Version
An AI lab deploys a system tasked with improving the efficiency of its own training pipeline. The system suggests hyperparameter changes, identifies redundant computation, and proposes adjustments to the data sampling strategy. Humans review and approve these. The next training run is 8% more efficient.
The next version of the system uses that efficiency gain to train on more data. It identifies further improvements. The humans in the loop are still there, but they’re increasingly rubber-stamping rather than actively reasoning through each decision.
At some point — and this is the threshold researchers care about — the AI’s contributions to the improvement process become more significant than the human contributions. The loop is still supervised, but the human is no longer the primary driver of progress.
The Harder-to-Predict Acceleration Phase
If the early phase goes well and the capability gains are real, the incentives to move faster are enormous. Labs competing for AI capability advantages would face pressure to reduce human bottlenecks in the improvement loop.
This is where the “explosion” part gets relevant. Even modest efficiency gains in the improvement loop compound quickly. A system that can improve its own training efficiency by 10% per cycle, cycling every week, doubles its efficiency in about 7 weeks.
Whether that leads to something that looks like a genuine intelligence explosion — or whether it plateaus, hits fundamental limits, or produces capability gains that are narrower than they appear — is genuinely unknown.
The Safety Dimension You Can’t Ignore
Any serious discussion of RSI has to include alignment. The two questions are inseparable.
A system that can improve its own capabilities is only safe if its objectives remain aligned with human interests through those improvements. This is sometimes called the “control problem” or the “alignment problem” at its hardest.
The core concern isn’t that a self-improving AI would become malicious. It’s that even a well-intentioned system optimizing for the wrong objective could take actions that are harmful at scale — and that a self-improved version would be harder to correct or shut down than the original.
Anthropic’s research on AI safety — including Constitutional AI and interpretability work — is explicitly aimed at ensuring that as systems become more capable, their values remain legible and correctable. The fact that safety research is a major focus at the labs closest to the frontier is both reassuring and telling about how seriously they take these risks.
Where AI Agents Fit Right Now
We’re not at RSI yet. But the current generation of AI agents represents an early, human-supervised version of the same underlying pattern: AI systems that take multi-step actions autonomously, evaluate outcomes, and adjust behavior accordingly.
Understanding this matters because it tells you something about the practical near-term trajectory. The infrastructure for agentic AI — autonomous action, tool use, multi-step reasoning — is being built right now and deployed in real workflows.
How MindStudio Connects to This
If you’re building or deploying AI agents today, you’re working with the precursor technology to the systems researchers are discussing when they talk about RSI timelines.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
MindStudio is a no-code platform that lets you build autonomous AI agents — systems that can take real-world actions, call external tools, process information, and complete multi-step tasks without human intervention at each step. You can connect agents to 1,000+ business tools, chain them together into workflows, and deploy them to run on schedules or in response to triggers.
The practical relevance here: as AI becomes more capable, the organizations that have already built infrastructure for agentic AI will be better positioned to take advantage of improved models as they arrive. Every new frontier model release improves the capabilities of agents built on that model — without requiring you to rebuild your workflows from scratch.
If you’re evaluating whether to start building with AI agents now, MindStudio’s guide to AI workflow automation covers how these systems work in practice. You can try MindStudio free at mindstudio.ai.
The broader point: the 2028 RSI discussion isn’t just philosophical. It’s a reason to understand how agentic AI systems work today, because the gap between current agents and self-improving AI systems is a matter of degree, not kind.
Frequently Asked Questions
What is recursive self-improvement in AI?
Recursive self-improvement refers to an AI system’s ability to enhance its own capabilities — through modifications to its architecture, training process, or reasoning methods — and then use that improved version to make further enhancements. Each iteration starts from a higher capability baseline, potentially creating a compounding improvement cycle.
Is recursive self-improvement happening already?
Not in the full sense. Current AI systems can assist with AI research and help optimize certain aspects of training pipelines, but they cannot autonomously close the full loop of modifying themselves, validating improvements, and redeploying more capable versions without human involvement. The question is how close that full loop is.
What is the intelligence explosion?
The intelligence explosion is a scenario, first described by mathematician I.J. Good in 1965, where a sufficiently capable AI system improves its own intelligence, creating a more capable system that improves itself further — leading to rapid, compounding capability gains that quickly surpass human-level intelligence. The term “technological singularity” is sometimes used to describe the point where this becomes impossible to predict from the outside.
Why do researchers cite 2028 as a potential milestone?
Several researchers, including Anthropic co-founder Jack Clark, have cited the late 2020s as a plausible timeframe for AI systems beginning to build their successors. This estimate is based on the rapid improvement in AI coding capabilities, the proliferation of autonomous AI agents, the trend of AI labs using AI to accelerate research, and continued scaling of compute infrastructure. These estimates carry significant uncertainty — a 60% probability still means a 40% chance it doesn’t happen by that date.
What would prevent recursive self-improvement from happening?
Several factors could slow or prevent RSI: fundamental architectural limits in current AI paradigms, diminishing returns from scaling, deliberate decisions by AI labs to restrict system access to training infrastructure, advances in AI safety research that create constraints on self-modification, and regulatory intervention. The timeline and likelihood are genuinely uncertain.
How does AI alignment relate to recursive self-improvement?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Alignment is the central safety concern with RSI. If a self-improving AI system has objectives that are even slightly misaligned with human values, those misalignments could amplify with each improvement cycle. Alignment research aims to ensure that AI systems remain interpretable, correctable, and human-compatible as they become more capable — which is why organizations like Anthropic, OpenAI, and DeepMind invest heavily in this work alongside capability research.
Key Takeaways
- Recursive self-improvement describes an AI system that can meaningfully improve its own capabilities and use those improvements to iterate further — closing a loop that currently requires human researchers to operate.
- Jack Clark’s 60% estimate for 2028 is one of several near-term predictions from people close to frontier AI development. The uncertainty is real, but the probability assigned is significant.
- The prerequisites for RSI — deep self-modeling, automated evaluation, sufficient agency, and alignment stability through modification — each represent active research areas. Progress on all four simultaneously is the threshold to watch.
- The intelligence explosion scenario follows directly from RSI: compounding capability gains that accelerate beyond normal development timelines.
- Safety and alignment research isn’t separate from this conversation — it’s the central question. How AI goals remain stable through self-modification is the hardest unsolved problem in the field.
- You don’t need to wait for 2028 to engage with this. Understanding how agentic AI workflows work today is practical preparation for a world where AI capabilities continue to increase rapidly.
The honest position is this: we don’t know whether 2028 is the right year. But the people building these systems think it’s a plausible one — and that’s worth taking seriously.
