Anthropic’s estimate that there is better than a 60% chance AI systems will autonomously train their successors by 2028 matters because it shifts recursive self-improvement from a speculative idea into a near-term planning problem for labs, companies, and governments. The important correction is that this is not yet a picture of runaway autonomy; today’s systems sit on a spectrum where AI does more of the work, but humans still set objectives, judge outputs, fund infrastructure, and decide what gets deployed.
Why the 2028 marker changes the planning horizon
Anthropic’s forecast gives a date and a probability to a transition that has often been discussed in abstract terms. If an AI system can train a successor model by 2028, the development loop for frontier systems could compress sharply, with economic and political effects arriving on ordinary budgeting and regulatory timelines rather than in some distant future.
That does not mean a clean handoff from humans to machines. Recursive self-improvement, in the practical sense, starts well before full autonomy: models already help write training code, debug systems, evaluate outputs, and assist with deployment operations. The threshold Anthropic is pointing to is the closing of more of that loop, not the disappearance of every human checkpoint.
Where the loop is already closing, and where it still is not
Google DeepMind’s AlphaEvolve is a useful marker because it shows both the progress and the limit. It uses large language models to evolve algorithms and improve designs, but humans still define the target and determine how success is measured, which means the system is powerful without being self-directed in the full sense.
At the company level, StrongDM’s “lights out” AI factory pushes further into operational autonomy by having AI agents write, test, and deploy software without human code review. That is a concrete deployment pattern for recursive self-improvement inside a bounded environment: the agents improve work products and compound local advantage, but they do so inside workflows, data access rules, and infrastructure owned by the firm.
Experimental systems such as Darwin Gödel Machines and the AI Scientist point in the same direction. They show that “seed improver” behavior, where a system can inspect and modify parts of its own process, is becoming technically plausible, but still under supervision and still dependent on human-designed evaluation, compute budgets, and safety boundaries.
Why full autonomy is harder than the forecast sounds
The main brake is not only model quality. Frontier AI development depends on costly training runs, specialized chips, data pipelines, networking, power, cooling, security, and teams that know how to keep large systems stable; those requirements run into the billions of dollars and cannot be abstracted away by saying the model can code.
The operational constraint is also organizational. Running a chip supply chain, a major data center fleet, or a model-serving platform requires distributed expertise and institutional control, so even a highly capable model does not automatically become self-sustaining. Nathan Lambert’s “lossy self-improvement” framing fits here: as systems become more complex, each additional improvement can become harder, noisier, and more resource-intensive rather than automatically compounding at the same rate.
| Stage | What AI does | What humans still control | Main limiting factor |
|---|---|---|---|
| Assisted development today | Code generation, debugging, evaluation support, deployment assistance | Goals, acceptance criteria, release decisions | Reliability and human review |
| Bounded autonomy | Algorithm search, software testing, deployment inside a defined environment | Objectives, environment design, risk limits | Operational complexity and domain constraints |
| Autonomous successor training by 2028, if achieved | Training substantial parts of the next model generation | Compute access, deployment authority, policy controls, external oversight | Compute cost, infrastructure dependence, alignment and governance |
The governance problem Anthropic is actually pointing to
Anthropic’s call for geopolitical crisis infrastructure, modeled on Cold War hotlines, is a sign that the company sees recursive self-improvement as a coordination problem as much as a research milestone. If one lab’s systems begin accelerating model development or autonomous research cycles, other firms and states may interpret that as a strategic shift and respond before they fully understand the technical details.
The governance need is therefore not limited to model safety testing. Governments and major labs may need mechanisms for rapid communication, incident reporting, compute monitoring, and agreed procedures for slowing deployment or limiting diffusion during an acute risk event. Without that layer, the pressure to keep up could outrun the ability to verify what a system is actually doing.
The next checkpoint for companies and regulators
The practical question over the next few years is not whether AI can contribute to building better AI; it already can. The real checkpoint is whether any system reaches a genuinely closed loop in which it can propose, train, evaluate, and meaningfully improve a successor with minimal human intervention, and whether that loop can operate repeatedly under real production constraints.
For companies, that means separating useful automation from a fantasy of total autonomy. StrongDM-style workflows may be valuable in tightly scoped software environments, but the economics only hold where firms have proprietary data, clear metrics, and enough operational control to trust machine-run iteration. For regulators, the checkpoint is different: which labs have the compute, data access, and organizational concentration to make autonomous successor training plausible by 2028, and what controls exist before that threshold is crossed.
Immediate questions
Is recursive self-improvement already here? In partial form, yes. AI systems already help improve tools, code, and research workflows, but that is still different from a fully autonomous runaway process.
What would count as a material step by 2028? A system that can train a successor model with much less human intervention than today, while handling more of the evaluation and iteration loop itself.
What is the main warning sign? Not just better benchmark scores, but a lab or company showing repeatable closed-loop improvement tied to real compute infrastructure and real deployment authority.
