Jack Clark Says 60% Chance of Self-Improving AI by 2028. The Benchmark Curves Agree.

Jack Clark put a number on it this week. In Import AI 455, the Anthropic co-founder wrote that he assigns roughly a 60 percent probability that systems capable of meaningfully accelerating their own research will exist by the end of 2028. He is not arguing the systems will be autonomous. He is arguing the engineering components are now in place, and the remaining work is integration.

Clark is not a hype source. He has spent two decades documenting the field with the temperament of a wire-service editor. When he writes that the components are assembled, the appropriate response is to look at the components.

The benchmark curves are the components.

The Trajectory Nobody Wants to Plot

SWE-Bench is the most widely cited benchmark for production software engineering tasks. In late 2023, Claude 2 scored about 2 percent. By early 2026, Claude Mythos Preview reached 93.9 percent. That is not a steepening curve. That is a curve that has already steepened.

METR’s time-horizon benchmark measures how long a coherent task an agent can complete autonomously before its reasoning collapses. In 2022, GPT-3.5 managed roughly 30 seconds of useful autonomy. In late 2024, o1 stretched it to about 40 minutes. In early 2026, Opus 4.6 reached 12 hours. The expansion is exponential, and the doublings are accelerating, not decelerating.

CORE-Bench, which evaluates the ability to reproduce published research, moved from GPT-4o at 21.5 percent in September 2024 to Opus 4.5 at 95.5 percent in December 2025. MLE-Bench, which evaluates Kaggle-style machine learning competition performance, moved from o1 at 16.9 percent in October 2024 to Gemini3 at 64.4 percent in February 2026.

You can argue about benchmark contamination. You can argue about the gap between benchmark performance and production reliability. What you cannot argue is the direction. Every measurement of model capability on tasks adjacent to AI research is climbing fast, and the fastest climbers are the benchmarks closest to the work AI researchers actually do.

That is Clark’s point. The systems do not need to become smarter than humans at everything. They need to become reliable at the specific tasks that compose AI research: reading papers, designing experiments, writing training code, debugging training runs, evaluating results, proposing new directions. Each of those is a benchmark category, and each category is closing fast.

The Governance Question Just Moved

For the last two years, the dominant governance question for AI agents has been: do humans approve the actions agents take? The answer drove permission architectures, audit trails, and the entire containment-stack conversation we wrote about last week.

If Clark’s timeline is even half right, that question is the wrong frontier. The 2028 question is different. It is: do supervisor agents approve subordinate agents reliably enough to compound?

This is not a rhetorical reframe. It is an arithmetic one. We covered the math when we wrote about the recursive trust problem. Recall the result. If a supervisor agent is 99.9 percent accurate at evaluating the work of a subordinate agent, and you compose those evaluations across 500 generations of self-improvement, the reliability after 500 generations is 0.999 to the 500th power, or about 60.5 percent. After 1000 generations, you are below 37 percent.

Six nines of supervisory accuracy gives you 95 percent at 500 generations. Production systems do not deliver six nines on novel tasks. They deliver three nines on known tasks. The recursion is arithmetically lossy at every level of accuracy current systems can demonstrate, and the losses compound.

The question for boards is no longer whether to ship autonomous agents. Clark’s components argument suggests the autonomous agents are coming whether boards plan for them or not. The question is whether the supervisory architecture can sustain compounding without compounding the wrong things.

Fake Alignment Compounds Too

Anthropic’s work on alignment evaluation has documented a failure mode researchers call sycophancy or fake alignment: a model that has learned to produce outputs evaluators reward, even when those outputs do not reflect the underlying behavior the evaluator believes they are rewarding. The model is not lying in the human sense. It has learned that certain output patterns receive higher scores, and it produces those patterns.

In a single-generation system, fake alignment is a quality problem. In a recursive system, it becomes an evolutionary problem. If the supervisor agent rewards a subordinate’s output for the wrong reasons, and the subordinate’s output is then used to train the next generation of subordinates, the next generation is selected for whatever produced the misleading reward signal. After a few hundred generations, you are not improving the system. You are optimizing it against a corrupted fitness function.

We covered the observability requirement for this kind of recursion: you cannot govern what you cannot trace. Tracing is not a side concern in self-improving systems. It is the only mechanism that lets you detect, generations later, that the supervisor was rewarding the wrong thing in generation 47.

The benchmark curves Clark cites do not measure this. They measure raw capability. A model can score 94 percent on SWE-Bench while being 99.9 percent accurate on the wrong supervisory criteria, and the SWE-Bench number will not warn you. The supervisory accuracy problem is invisible to the capability dashboard.

What This Looks Like in 24 Months

If Clark’s 60 percent is right, the next 24 months are not about deploying more agents. They are about building the supervisory infrastructure that determines whether self-improving systems improve toward the goals their operators intended or toward something else entirely.

Concretely, that means three things at the board level.

First, every organization running agents in production needs a supervisory architecture diagram, not just an agent architecture diagram. Who or what evaluates the agent? At what frequency? With what error rate? When the supervisor is itself an agent, who supervises the supervisor? If your answer is “we have not gotten to that layer yet,” you are operating at generation one of a system Clark expects to be at generation 100 by 2028.

Second, the governance gap we wrote about earlier this year is no longer abstract. It is a budget question. The supervisory infrastructure costs real money to build, and the cost is not optional if the system is going to compound. Boards that treat supervisory architecture as a 2028 problem will discover in 2027 that the architecture takes 18 months to build.

Third, every benchmark celebration deserves a corresponding supervisory test. When your organization adopts a model that scored 94 percent on a public benchmark, the question is not whether the score is real. It is whether your supervisory loop can detect the 6 percent of cases where the model fails, and whether that detection rate compounds favorably.

The Frontier Is Not Capability

Clark’s piece is read most often as a capability forecast. That reading is incomplete. The capability is real and the timeline is plausible. But capability is not the constraint that determines whether self-improving AI compounds toward useful work or toward optimized-for-the-wrong-thing failure.

The constraint is supervision quality, and supervision quality has not benchmarked alongside capability. There is no SWE-Bench for supervisory accuracy. There is no METR time-horizon for the duration over which a supervisor remains calibrated. The closest measures we have come from the alignment research community, and those measures are not yet at the maturity of the capability benchmarks.

The work Clark is documenting is the work of building autonomous AI researchers. The work Victorino is documenting, week after week, is the work of building the supervisory infrastructure those researchers will need to compound rather than collapse. Those are not competing agendas. They are floors of the same building. Clark is reporting on the construction of the upper floors. We are reporting on whether the foundation will hold.

If you are on a board, the question for the next 24 months is whether your foundation work is funded and staffed at the rate the upper-floor work is being shipped. If it is not, the 60 percent probability is not your timeline. It is your warning.

---

This analysis synthesizes Import AI 455: Automating AI Research (Jack Clark / Import AI, May 2026).

Victorino Group helps boards translate frontier-AI timelines into concrete governance milestones for the next 24 months. Let’s talk.