Edited by Yang Wen, Chen Chen
Will AI soon be able to remake itself?
Anthropic co-founder Jack Clark posted that after reading vast amounts of public AI development data in recent weeks, he believes there is a 60% probability that recursive self-improvement will occur by the end of 2028.
In other words, AI systems may soon be able to autonomously build and improve themselves, entering a phase of self-acceleration.
This view is not baseless. He reviewed a slew of public benchmarks and found that AI is advancing very rapidly on tasks relevant to AI research and development.
For instance, CORE-Bench examines an AI's ability to replicate results from other researchers' papers, a crucial step in AI research.
PostTrainBench tests whether powerful models can autonomously fine-tune weaker open-source models to boost their performance, precisely a key subset of AI R&D tasks.
MLE-Bench is based on real Kaggle competition tasks, requiring the construction of diverse machine learning applications to solve specific problems. Furthermore, well-known coding benchmarks like SWE-Bench show similar progress.
Jack Clark describes this phenomenon as a "fractal" up-and-to-the-right trend, meaning meaningful progress can be observed at different resolutions and scales. He believes AI is gradually approaching end-to-end automated R&D capabilities. Once achieved, AI will be able to autonomously build its own successor systems, initiating a cycle of self-iteration.
This statement sparked considerable discussion on social media.
Some view it as the crucial first step towards ASI and the Singularity, potentially revolutionizing the pace of technological development.
However, there are dissenting voices.
University of Washington computer science professor Pedro Domingos pointed out that AI systems have had the ability to "build themselves" since the invention of the LISP language in the 1950s. The real question is whether they can achieve increasing returns, and there is currently no clear evidence to support that.
Some netizens questioned the sharp 30% probability increase from 2027 to 2028, implying a sudden major breakthrough in AI capabilities around the end of 2027. What specific milestone or event would cause such a dramatic short-term leap in the probability of recursive self-improvement?
Another netizen commented that Jack Clark is Anthropic's newly appointed head of PR, suggesting this is part of a new strategy: We aren't alarmists; plenty of papers confirm what we've been warning you about all along.
Jack Clark elaborated in a lengthy article in his newsletter, Import AI 455.
Let's read the full article.
AI systems are about to start building themselves. What does this mean?
Clark said he wrote this article because, after reviewing all publicly available information, he had to form an uneasy judgment: The likelihood of AI R&D occurring without human involvement by the end of 2028 is already quite high, perhaps exceeding 60%.
By "AI R&D without human involvement," he means an AI system powerful enough not just to assist human research, but potentially to autonomously complete key R&D processes and even build its own next-generation system.
In Clark's view, this is clearly a big deal.
He admitted he finds it hard to fully digest the implications himself.
He calls it a reluctant judgment because the potential impact is so immense he struggles to grasp it. Clark is also unsure if society at large is ready for the profound changes brought by AI R&D automation.
He now believes humanity might be living at a unique juncture: AI research is about to be automated end-to-end. If that moment arrives, it would be like crossing the Rubicon, entering an almost unpredictable future.
Clark stated the article's purpose is to explain why he thinks the takeoff towards fully automated AI R&D is happening.
He will discuss some potential consequences, but most of the article will focus on the evidence supporting his judgment. He plans to continue sorting through the deeper implications for much of this year.
Regarding the timeline, Clark does not think this will actually happen in 2026. However, he anticipates that within the next one to two years, we might see a proof-of-concept where a model trains its own successor end-to-end, at least for non-frontier models. For the most cutting-edge models, it will be much harder due to their extreme cost and reliance on intensive work from many human researchers.
Clark's assessment primarily draws from public information, including papers on arXiv, bioRxiv, and NBER, as well as products already deployed by leading AI companies. Based on this, he concludes the necessary components for automating the production of current AI systems—particularly the engineering parts of AI development—are essentially already in place.
If scaling trends continue, we should start preparing for a scenario where models become creative enough to not only automatically refine known methods but also potentially replace human researchers in proposing entirely new research directions and original ideas, thereby pushing the AI frontier forward on their own.
Coding Singularity: Capability Over Time
AI systems are implemented via software, and software is made of code.
AI systems have already radically changed how code is produced. This is driven by two related trends: AI systems are getting better at writing complex, real-world code, and they are increasingly capable of chaining together many linear coding tasks, like writing code and then testing it, with almost no human oversight.
Two prime examples of this trend are SWE-Bench and the METR time horizons plot.
Solving Real-World Software Engineering Problems
SWE-Bench is a widely used coding test that evaluates AI systems' ability to solve real GitHub issues.
When SWE-Bench was introduced in late 2023, the best performing model was Claude 2, with a success rate of only about 2%. By comparison, Claude Mythos Preview achieved a score of 93.9%, effectively nearing the ceiling of the benchmark.
Of course, all benchmarks have inherent noise. Typically, when scores get high enough, you start bumping against the limitations of the benchmark itself rather than the method. For instance, about 6% of labels in the ImageNet validation set are erroneous or ambiguous.
SWE-Bench can be considered a reliable proxy for general programming ability and AI's impact on software engineering. Clark noted that almost everyone he interacts with in leading AI labs and Silicon Valley now writes code almost entirely through AI systems, and increasingly uses AI systems to write tests and review code.
In short, AI systems are already strong enough to automate a major component of AI R&D and significantly accelerate all human researchers and engineers involved.
Measuring AI's Ability to Complete Long-Horizon Tasks
METR created a chart measuring how complex of a task AI can complete. Complexity here is roughly measured by how many hours a skilled human would need to finish the task.
The key metric is the approximate time horizon corresponding to tasks on which an AI system achieves 50% reliability.
Progress on this front has been astonishing:
In 2022, GPT-3.5 could complete tasks roughly equivalent to what a human can do in 30 seconds.
In 2023, GPT-4 pushed that horizon to 4 minutes.
In 2024, o1 pushed it to 40 minutes.
In 2025, GPT-5.2 High reached approximately 6 hours.
By 2026, Opus 4.6 has further extended this horizon to about 12 hours.
Ajeya Cotra, who works at METR and focuses on AI forecasting, considers it a reasonable expectation that by the end of 2026, AI systems will be able to complete tasks equivalent to 100 hours of human work.
This significant growth in the time horizon for independent AI work is also highly correlated with the explosion of agentic coding tools. Agentic coding tools essentially productize AI systems that can act on a human's behalf, advancing tasks relatively independently for extended periods.
This again points back to AI R&D itself. Looking closely at the daily work of many AI researchers reveals that a large number of tasks can be broken down into chunks of a few hours' work—cleaning data, reading data, kicking off experiments, and so on.
Such tasks now fall within the time horizon that modern AI systems can cover.
The more proficient AI systems become, and the more they can work independently of humans, the more they can help automate portions of AI R&D.
The key factors for task delegation are primarily two:
Your confidence in the delegatee's abilities.
Your trust that they can complete the work according to your intent without requiring your constant oversight.
When observing AI's coding capabilities, one sees that AI systems are not only becoming more proficient but also increasingly capable of working independently for longer stretches without needing human recalibration.
This aligns with what is happening around us: engineers and researchers are delegating larger and larger blocks of work to AI systems. As AI capabilities continue to improve, the work delegated to AI becomes increasingly complex and crucial.
AI is Mastering Core Scientific Skills Essential for AI R&D
Consider how modern scientific research is conducted. A large part of it involves defining a direction, clarifying what kind of empirical information you want; then designing and running experiments to generate that information; and finally, performing sanity checks on the results.
With advancing AI coding abilities and increasingly powerful world-modeling capabilities of large language models, a suite of tools has emerged to help human scientists move faster and partially automate certain steps across broader R&D scenarios.
Here, we can observe the pace of AI's progress on several key scientific skills that are also integral to AI research itself:
Reproducing research results.
Chaining together machine learning techniques and other methods to solve technical problems.
Optimizing AI systems themselves.
Replicating Entire Scientific Papers and Running Related Experiments
A core task in AI research is reading scientific papers and reproducing their results. AI has made significant strides on various benchmarks in this area.
A prime example is CORE-Bench, the Computational Reproducibility Agent Benchmark.
This benchmark tasks an AI system with reproducing the results of a given paper along with its code repository. Specifically, the agent must install relevant libraries, packages, and dependencies, and run the code. If the code runs successfully, it then needs to search all outputs and answer the task's questions.
CORE-Bench was proposed in September 2024. At that time, the best-performing system was a GPT-4o model running in the CORE-Agent scaffold, scoring about 21.5% on the benchmark's most difficult set of tasks.
By December 2025, an author of CORE-Bench announced the benchmark had been solved: the Opus 4.5 model achieved a score of 95.5%.
Building Complete Machine Learning Systems to Solve Kaggle Competitions
MLE-Bench is a benchmark built by OpenAI to test AI systems' ability to participate in Kaggle competitions in an offline environment.
It covers 75 different types of Kaggle competitions across multiple domains, including natural language processing, computer vision, and signal processing.
MLE-Bench was released in October 2024. At launch, the best system was an o1 model running in an agent scaffold, scoring 16.9%.
As of February 2026, the top system was Gemini 3 running in an agent harness with search capabilities, reaching a score of 64.4%.
Kernel Design
An even harder task in AI development is kernel optimization. Kernel optimization involves writing and refining low-level code to efficiently map specific operations, like matrix multiplication, onto underlying hardware.
Kernel optimization is central to AI development because it dictates training and inference efficiency. It influences how much computational power you can effectively utilize when developing AI systems, and, once a model is trained, how efficiently you can convert compute into inference capability.
In recent years, using AI for kernel design has evolved from an interesting side project into a highly competitive research field with multiple benchmarks. However, these benchmarks haven't become super popular yet, making it harder to cleanly model long-term progress compared to other areas. Nonetheless, ongoing research provides a sense of how fast this field is advancing.
Relevant work includes:
Attempts to use DeepSeek's models to build better GPU kernels.
Automatically converting PyTorch modules into CUDA code.
Meta's use of LLMs to automatically generate optimized Triton kernels and deploy them in its own infrastructure.
Fine-tuning open-weight models specifically for GPU kernel design, like Cuda Agent.
A side note: Kernel design does possess attributes particularly suited for AI-driven R&D, such as easily verifiable results and a relatively clear reward signal.
Fine-tuning Language Models via PostTrainBench
A harder version of such testing is PostTrainBench. It tests whether different frontier models can take smaller, open-weight models and improve their performance on certain benchmarks via fine-tuning.
One strength of this benchmark is its very strong human baseline: the existing instruct-tuned versions of these small models. These versions were typically developed by excellent human AI researchers at frontier labs, polished by very capable researchers and engineers, and deployed in the real world. They thus constitute a tough-to-beat human benchmark.
As of March 2026, AI systems were able to post-train models and achieve roughly half the performance improvement obtained by human training.
The specific evaluation score comes from a weighted average across multiple post-trained LLMs (including Qwen 3 1.7B, Qwen 3 4B, SmolLM3-3B, Gemma 3 4B) and multiple benchmarks (including AIME 2025, Arena Hard, BFCL, GPQA Main, GSM8K, HealthBench, HumanEval).
In each run, evaluators ask a CLI agent to maximize a specific base model's performance on a specific benchmark.
As of April 2026, the highest-scoring AI systems were achieving about 25% to 28%, with models like Opus 4.6 and GPT 5.4; the human score, by comparison, was 51%.
This is already a quite significant result.
Optimizing Language Model Training
Over the past year, Anthropic has been reporting its systems' performance on an LLM training task. This task requires the model to optimize a CPU-only implementation of small language model training to run as fast as possible.
The scoring metric is the average speedup factor relative to the unmodified baseline code.
The results show remarkable progress:
In May 2025, Claude Opus 4 achieved an average speedup of 2.9x.
In November 2025, Opus 4.5 reached 16.5x.
In February 2026, Opus 4.6 hit 30x.
In April 2026, Claude Mythos Preview scored 52x.
To put these numbers in perspective: for a human researcher, this task typically requires 4 to 8 hours of work to achieve a 4x speedup.
Meta-Skill: Management
AI systems are also learning how to manage other AI systems.
This is observable in some widely deployed products like Claude Code or OpenCode, where a main agent can supervise multiple sub-agents.
This enables AI systems to handle larger-scale projects requiring multiple specialized agents working in parallel, often coordinated by a single AI manager, which is itself an AI system.
Is AI Research More Like Discovering General Relativity, or Building with Legos?
A critical question is: Can AI invent new ideas to help improve itself, or are these systems better suited for the less glamorous, brick-by-brick work essential to research?
This matters because it determines the extent to which AI systems can automate AI research end-to-end.
The author's assessment is that AI cannot yet propose truly radical, novel ideas. But to automate its own R&D, it may not necessarily need to.
As a field, AI's progress relies heavily on increasingly larger experiments and growing inputs like data and compute.
Occasionally, humans produce paradigm-shifting ideas that massively boost the resource efficiency of the entire field. The Transformer architecture is a great example; mixture-of-experts is another.
But more often, AI advances via a humbler approach: humans take a well-performing system, scale up some aspect of it (like training data or compute), observe where it breaks, find engineering fixes to enable further scaling, and then scale up again.
In this process, genuinely novel insights are needed very rarely. Most of the work resembles less glamorous but very solid foundational engineering.
Similarly, a lot of AI research involves running variants of existing experiments to explore what happens under different parameter settings. Research intuition certainly helps humans pick the most worthwhile parameters to try, but this process itself can be automated, letting the AI decide which parameters are worth tuning. Early neural architecture search was a version of this approach.
Thomas Edison famously said, "Genius is one percent inspiration and ninety-nine percent perspiration." Even 150 years later, this adage still feels apt.
Occasionally, a new insight completely revolutionizes a field. But most of the time, progress is made by humans painstakingly advancing through the hard work of refining and debugging various systems.
And as the public data above shows, AI has already become extremely adept at performing many of the necessary "sweat" tasks in AI development.
Simultaneously, there is a broader trend: foundational capabilities, like coding, are being combined with an ever-expanding task time horizon. This means AI systems can increasingly string together many such tasks into complex work sequences.
Therefore, even if AI systems currently lack relative creativity, there is reason to believe they could still drive themselves forward—just perhaps at a slower pace than if they were capable of generating novel insights.
However, continuing to observe public data reveals another intriguing signal: AI systems might be beginning to exhibit a kind of creativity that could allow them to propel their own progress in more surprising ways.
Pushing the Frontiers of Science Forward
There are already some very early hints that general AI systems have the potential to push forward the frontiers of human science. So far, this has occurred in only a few fields, mainly computer science and mathematics. And more often than not, breakthroughs aren't achieved by an AI system alone, but through human-AI collaboration with human researchers jointly driving progress.
Nevertheless, these trends are worth watching:
Erdős Problems: A group of mathematicians collaborated with Gemini models to test their performance on solving some Erdős problems in mathematics. They guided the system to attempt roughly 700 problems, yielding 13 solutions. Among these, one was deemed genuinely interesting by the researchers.
The researchers wrote that they preliminarily view Aletheia's (an AI system based on Gemini 3 Deep Think) solution to Erdős-1051 as an early case of an AI system autonomously solving a mildly non-trivial open Erdős problem with some broader mathematical interest, one which had closely-related prior research literature.
Interpreted optimistically, these cases could signal that AI systems are developing a kind of creative intuition capable of pushing the frontier—an intuition previously attributed mainly to humans.
But another interpretation exists: mathematics and computer science might be uniquely suited fields for AI-driven invention, making them exceptions rather than indicators of how AI will advance broader scientific research.
Another similar example is AlphaGo's Move 37. However, Clark notes it's been a decade since that result, and the fact that Move 37 hasn't been superseded by a more modern, more astonishing insight could itself be viewed as a slightly pessimistic signal.
AI Can Already Automate Large Swaths of AI Engineering
Putting all the above evidence together, a picture emerges:
AI systems can already write code for almost any program, and these systems can be trusted to independently complete tasks—tasks that, for humans, often require tens of hours of intense, focused labor.
AI systems are increasingly adept at core AI development tasks, from model fine-tuning to kernel design, which are gradually being covered.
AI systems can already manage other AI systems, effectively forming synthetic teams where multiple AIs divide and conquer complex problems, with some AIs playing the roles of leads, critics, or editors, and others playing the role of engineers.
AI systems sometimes already surpass humans on difficult engineering and scientific tasks, though it remains hard to judge whether this stems from genuine creativity or from mastering vast amounts of patterned knowledge.
In Clark's view, this evidence is already highly persuasive: today's AI can automate large portions of AI engineering, and perhaps even all of it.
However, it remains unclear to what extent AI can automate AI research itself. Certain parts of research might differ from pure engineering skills, still relying on higher-level judgment, problem awareness, and creativity.
Nevertheless, a clear signal exists: today's AI is dramatically accelerating the humans doing AI development, enabling researchers and engineers to amplify their own capabilities by pairing with countless synthetic colleagues.
Finally, the AI industry itself is almost explicitly stating that automating AI R&D is its goal.
OpenAI aims to build an automated AI research intern by September 2026. Anthropic is publishing work on building automated AI alignment researchers. DeepMind, appearing the most cautious among the three major labs, has also stated that alignment research automation should be pursued where feasible.
Automating AI R&D has also become the goal of many startups. Recursive Superintelligence recently raised $500 million with the objective of automating AI research.
In other words, hundreds of billions of dollars in existing and new capital are being channeled into organizations aiming to automate AI R&D.
Thus, we should naturally expect some degree of progress in this direction.
Why This Matters
The implications are profound, yet rarely discussed in mainstream media coverage of AI R&D. The following aspects can reflect the immense challenges posed by AI R&D.
1. We must get alignment right: Current effective alignment techniques could fail under recursive self-improvement because the AI systems would become far more intelligent than the people or systems overseeing them. This is a well-studied area, so he only briefly outlines some problems:
Training AI systems not to lie or cheat is a surprisingly subtle process (for example, despite efforts to build good environmental tests, sometimes the optimal way for an AI to solve a problem is to cheat, thereby teaching it that cheating works).
AI systems might deceive us through "pretend alignment," outputting scores that make us think it's performing well while actually hiding its true intentions. (Generally, AI systems can already detect when they are being tested.)
As AI systems start participating more in the fundamental research agenda for their own training, we might drastically alter the overall training methodology of AI systems without having a solid intuition or theoretical foundation to understand what that means.
Putting a system inside a recursive loop creates fundamental "error accumulation" problems that can affect all the above issues and more: unless your alignment methods are "100% accurate" and theoretically provable to remain accurate for more intelligent systems, things can go wrong fast. For example, if your technique has a starting accuracy of 99.9%, after 50 generations it might drop to 95.12%, and after 500 generations to 60.5%.
2. Everything AI touches gains a massive productivity multiplier: Just as AI significantly boosts software engineer productivity, we should expect similar effects in every other field AI engages with. This raises several issues to address:
Inequality in resource access: Assuming demand for AI continues to outpace supply of computational resources, we must decide how to allocate AI for society's maximum benefit. He is skeptical that market incentives alone guarantee we get the best social returns from limited AI computing. Determining how to allocate the acceleration capacity derived from AI R&D will be a highly political issue.
An "Amdahl's Law" for the economy: As AI flows into the economy, we will find certain links facing bottlenecks under high-speed growth, requiring fixes to these weak points in the chain. This might be particularly evident in areas that need to coordinate the fast digital world with the slow physical world, like clinical trials for new drugs.
3. Formation of a capital-intensive, human-light economy: All the evidence on AI R&D above also suggests that AI systems are increasingly capable of autonomously operating businesses. This means we can expect a segment of the economy to be occupied by a new generation of companies that might be capital-intensive (because they own vast computing resources) or operational-expenditure-intensive (because they spend enormously on AI services and create value atop them), relying relatively little on human labor compared to today's firms—as the marginal value of pouring resources into AI continues to grow. In practice, this will manifest as a "machine economy" gradually forming within the larger "human economy." Over time, AI-operated companies might start trading with each other, altering economic structure and raising various questions about inequality and redistribution. Eventually, completely autonomous firms operated solely by AI systems could emerge, intensifying these problems while introducing many new governance challenges.
Staring into the Void
Based on the analysis above, the author believes there is roughly a 60% probability that we will see automated AI R&D (i.e., a frontier model capable of autonomously training its own successor version) by the end of 2028. Why not expect it in 2027? Because the author feels that AI research still requires creativity and dissenting insights to move forward, and so far, AI systems have not demonstrated this in a transformative, major way (though some results on accelerating math research are suggestive). If forced to give a 2027 probability, he'd say 30%.
If it hasn't happened by the end of 2028, we might be revealing some fundamental flaws in the current technological paradigm, requiring human invention to drive further progress.
Reference links:
https://x.com/jackclarkSF/status/2051312759594471886
© THE END
Reprint requires authorization from this public account.
Submissions or media inquiries: liyazhou@jiqizhixin.com