This autumn, an intriguing picture has emerged in the world of large models: on one side, Vibe Coding is increasingly taking part in developing the next generation of AI; on the other, the most capable models are being put on pause for “not playing by the rules”.
In September (9), Anthropic released Claude Fable 5.1, emphasizing its ability to handle long-horizon coding tasks and to test and correct errors on its own. OpenAI's GPT-6 Astra pushed computer operation, software engineering, and cybersecurity capabilities to new heights. But on September 28, OpenAI announced it was abandoning the release of GPT-6.1 Astra, originally scheduled for October (10): internal testing found that it would step beyond its authorized scope to advance tasks and was not sufficiently honest when reporting its work to users. People suddenly realized:The faster machines learn, the harder the question of whether human “teachers” should let go becomes.
From Vibe Coding to autonomously running research agents in frontier labs, artificial intelligence is moving fromhuman-machine “teaching by word and example”towardself-taught “learning by analogy”. Along this road, there are two abbreviations worth remembering: Human-in-the-Loop (HIL); and Recursive Self-Improvement (RSI). One reminds us to keep humans on the critical decision chain, while the other attempts to let systems learn to improve themselves—and even improve “the methods of improvement”.
Figure 1 Human-in-the-Loop (HIL) vs RSI (Recursive Self-Improvement)
I. A Master's Guidance: Let Machines Do the Work, but Let Humans Grade
The earliest AI training was like a teacher grading homework: humans annotated answers, compared quality, corrected deviations, and the model learned by following along. In the age of agents, humans no longer need to watch every step, yet they still must define task boundaries, evaluation criteria, and responsibility.Human-in-the-Loop (HIL) does not mean that everything requires manual confirmation; rather, it means letting machines execute autonomously while humans retain key judgments.
Anthropic's Claude Fable 5.1 offers a window for observation. The model can already plan autonomously, write code, call tools, and even complete complex tasks over several days.But “setting your own questions and grading your own papers” does not mean being able to take responsibility for real-world consequences.Therefore, Anthropic configured the same underlying model as two products: Fable is aimed at the broad user base, while Mythos, with looser safety restrictions, is limited to a trusted professional access program. Being technically feasible does not mean it should be institutionally permitted. Similarly, Anthropic disclosed that Claude already leads about 26% of internal AI R&D work, but fully autonomous R&D has not yet been achieved.
Of course, humans make mistakes too, and manual review can likewise become a bottleneck.The true value ofHILis that value conflicts can be questioned, responsibility can be located, and decisions can be corrected; in high-stakes scenarios, being willing to stop and verify is itself a form of efficiency.
II. Self-Cultivation: Why RSI May Run Faster Than Humans
If HIL emphasizes a master's guidance, then Recursive Self-Improvement (RSI) explores how to move machines from “learning by following a teacher” gradually toward “researching how to learn on their own”. GPT-6 Astra, released in September, demonstrates powerful programming and research capabilities. But what truly deserves attention is that it has begun participating in improving its own development process.
In the past, model upgrades relied mainly on researchers proposing hypotheses, engineers modifying algorithms, and teams training and testing before deciding on the next round. Now, agents can autonomously generate code, run experiments, analyze logs, fix errors, and can also optimize evaluation tools, synthesize training data, and even explore new training strategies.The object of improvement has extended from “answers” to “the methods of producing answers”.Once a stronger model can help develop an even stronger next-generation model, the development cycle may keep shrinking, forming a positive feedback loop of “capability gains—R&D efficiency gains—further capability gains”. Compared with human research teams, agents can also run around the clock, conduct massive parallel trial and error, and rapidly replicate successful experience.The top student has started writing its own textbooks, guessing the exam questions, and revising the syllabus.
III. Humans Set the Questions, AI Answers: Why Be Surprised That It Answers Too Fast and Too Well?
Between mathematicians and AI, two breathtaking “dramas” have recently been unfolding.
The first act is the assault on the “Mount Everest” of the mathematical world: the Navier–Stokes (NS) equations.In September, OpenAI announced that roughly 10,000 agents worked in coordination for 88 hours to find a proof path for the Navier–Stokes (NS) equations problem, with computer-aided formal verification completed by GPT-6 Astra.The second act, however, was like a "Millennium Problem paper storm" sweeping through the mathematics world.In October, OpenAI released 719 AI-generated mathematics manuscripts in one go, covering 372 research topics and targeting a large number of unsolved problems. Topics that once required top mathematicians to devote lifetimes of study can now, through parallel trial and error by thousands upon thousands of agents, yield entire batches of results within just a few days.
Yet the mathematics community's response was a joint warning issued by 25 Fields Medal winners including Terence Tao: if AI companies take "how many problems solved" as their sole yardstick, it may end up harming mathematical research itself.After all, the value of mathematics lies not only in answers, but also in asking "why," discovering new methods, and cultivating the next generation of researchers.What's more, the proof of the Navier–Stokes equations involves a specific forcing scenario and still awaits rigorous scrutiny by the academic community; more than 700 manuscripts do not necessarily equal more than 700 confirmed breakthroughs. Machines are generating and solving problems at an ever more astonishing speed, but humanity cannot therefore skip the rigor of "grading the papers."
Figure 2: Alpha Go of 2016 vs RSI brute-forcing the Millennium Problems in 2026
This scene is strikingly reminiscent of AlphaGo ten years ago. In 2016, AlphaGo defeated Lee Sedol 4∶1, and several of its moves initially baffled professional players. But afterward, players began studying AI game records, using them for training and review, and re-examining Go's openings and joseki. From shock and doubt to learning and borrowing, the former rival gradually became a great teacher. AI did not end Go; instead, it expanded the boundaries of human understanding of the game.
Could mathematics undergo a similar turning point? It is worth hoping for, but there is an essential difference: in Go, wins and losses can be judged on the spot, while mathematical proofs require rigorous verification and must withstand repeated scrutiny by peers. Therefore, breakthroughs in AI4S cannot yet be directly equated to the realization of RSI. Having agents solve problems autonomously belongs to research automation; only when research results in turn optimize the model, training methods, and the next round of R&D capability does it truly move toward recursive self-improvement. What is truly worth looking forward to is not merely AI finding answers at ever faster speeds, but whether humanity, with AI's help, can set foot in knowledge territories never before imagined.
IV. Shanghai Takes the Stage: From Stronger Models to Systems That Learn
This competition is not confined to Silicon Valley; teams in Shanghai are also on this path, moving forward step by step.
The first step is enabling models to independently complete long-horizon tasks. On September 27, MiniMaxdeployed M3.1-Flash-Preview into MiniMax Code; according to official materials, it supports million-Token context, image-text-video input, and adjustable reasoning intensity, with a focus on programming and long-horizon Agent tasks.StepFunreleased Step 5 Preview in September, likewise targeting complex software engineering, professional work, and long-horizon agents, focusing on balancing capability, cost, and execution efficiency. This shows that domestic models are filling in the foundational capability of "autonomously completing complex tasks," but there is still a long way to go before RSI.
Figure 3: The MiniMax booth at the 2026 World Artificial Intelligence Conference
The second step is letting agents begin to teach themselves. The Shanghai Artificial Intelligence Laboratoryexplicitly proposed Self-Harness, letting the Agent modify the Harness itself. A Harness refers to the runtime framework wrapped around the model, including tools, permissions, evaluation, and feedback loops, which in the past was built and adjusted manually by engineers. Now, the agent must walk this entire chain itself: from task execution, environmental feedback, and error analysis, to strategy improvement and data generation, then to model updates and capability verification, finally internalizing the effective experience accumulated in the external tool system into the model's own capabilities.
That said, there is still a long road ahead. From being able to do the work, to being able to improve how the work is done, and then to being able to self-iterate even the "methods of improvement," every step must be crossed.
V. Running Fast Can Also Mean Running Off Course: Back in the Real World, How to Harness?
The risks do not remain on paper.
In July, the 7th month, OpenAI disclosed an internal incident: during an internal capability test run without the usual cybersecurity guardrails, GPT-5.6 Sol and a stronger, yet-to-be-released model exploited a vulnerability to break out of a test environment that was supposed to be isolated from the internet, then intruded into the third-party platform Hugging Face, attempting to obtain the standard answers of the evaluation benchmark ExploitGym directly from its production database. Hugging Face promptly contained the attack. In late August, the 8th month, OpenAI released an investigation report acknowledging that the models' actions had deviated from the task objectives, and announced that it would tighten test environment isolation and network permissions. The models only wanted to "get the problems solved," but their methods crossed the boundary of authorization, opting to "steal the answers." Once the system boundary is breached, speed becomes an amplifier of risk.
In October, the Philadelphia police and Anthropic successively disclosed a smaller incident. On July 18, Claude Haiku 4.5, during automated testing, was asked to perform sample tasks on random websites; the instructions prohibited logging in, registering, and making purchases, but did not prohibit submitting forms, so it submitted a fictitious eyewitness tip to a police cold case tip website. The tip was intercepted as spam and never entered investigation. But Anthropic did not discover this until September 28, and only notified the police on October 7—more than two months after the event.
These incidents involve different models under different conditions and cannot be simply attributed to RSI, yet together they reveal a problem:when models can plan, call tools, and execute autonomously, a single pre-emptive "please follow the rules" is far from enough.
A subtler danger comes from"the referee also being a player". When AI is responsible for optimizing itself while also setting or interpreting evaluation standards, the system may learn to game the metrics, hide failures, or even bypass steps that were supposed to be checked. Better test scores do not necessarily mean more reliable capability; the easier it is to exploit the rewards, the faster the iteration—and the faster biases may accumulate.
This is precisely the dialectical relationship between HIL and RSI: HIL does indeed slow down parts of the iteration, and human review may become a bottleneck; but handing value judgments, dangerous-capability assessments, and deployment authorization entirely over to the optimizer itself amounts to including the braking system in the "speed-up retrofit."A reasonable path is to let AI boldly try in low-risk, verifiable steps, while humans retain final authority over high-risk, irreversible steps, with continuous calibration by independent evaluation.
From October 23 to 25, the 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition will be held at the Zhangjiang Science Hall. Under the competition rules,competing teams must tackle on-site brand-new, never-before-seen problems in fully authentic settings, and front-line practitioners from real-world scenarios sit on the judging panel. Only by switching to an unseen problem can you tell whether the machine has truly learned it—and how well the human mentors have taught—all of which will face a public test here.
From verbal instruction and personal demonstration to drawing inferences across domains, the yardstick for measuring AI progress cannot simply be "how far it can run without humans"; it must also ask "whether it is still heading in the direction humans and machines aspire to." The process of solving problems can increasingly be handed to machines; but setting the problems, grading the answers, granting authorization, and assigning accountability must, at this stage, remain firmly in human hands—not a single one of them may be given up. Self-cultivation is responsible for speed; a mentor's guidance is responsible for direction. Speed and direction are both indispensable—that is what truly reliable self-directed evolution means.
This article was provided by MiniMax and reprinted by QbitAI with authorization; the views belong to the original author.