Produced by I Xiahai (fallsea)

Written by I Hu Buzhi

On July 9, 2026, OpenAI officially opened up full access to the GPT-5.6 series of models to global users. The three models are named the flagship Sol, the balanced Terra, and the lightweight Luna.

Amid this information-packed release, the attention of the vast majority was drawn to two key events. The first was Sol's staggering 91.9% score on the programming benchmark Terminal-Bench 2.1, outpacing Anthropic's Claude Fable 5 by over 8 percentage points. The second was the halving of prices, with Luna's input cost dropping to just $1 per million tokens, directly overturning Silicon Valley's pricing system.

But deep within the technical community, what truly caused a seismic shift was a casual remark largely overlooked by most.

OpenAI mentioned in its technical documentation that Luna, the smallest model in the suite, was trained autonomously by its big brother, Sol. Specifically, Sol independently sought out available GPUs, determined training configurations, wrote launch scripts, and confirmed task execution—all without human engineer intervention.

The significance of this statement far exceeds any benchmark scores or reviews.

It means that tasks previously requiring human researchers to lead—such as data cleaning, reward model design, knowledge distillation, and hyperparameter search—can now be independently completed by the flagship model. AI is no longer just a tool in human hands; it has begun training its own successors.

In the field of AI safety, this concept has an unsettling scientific name: recursive self-improvement. When AI becomes powerful enough to autonomously restructure, test, and even fine-tune its next-generation models, the flywheel predicted by futurists for decades has finally begun to spin.

This is not a routine version update. It is a public declaration of a paradigm shift.

What Is AI Training AI?

To understand the disruptive nature of recursive self-improvement, one must first grasp the traditional model of large-scale AI training.

During the GPT-4 and GPT-5 eras, training a top-tier large model was essentially a highly human-dependent craft.

The entire process roughly unfolded as follows. First, human researchers Filter ed (selected) and cleaned trillions of tokens of training data from the internet. Then, they designed pre-training architectures and adjusted thousands of hyperparameters. Next, through human-feedback-based reinforcement learning, they aligned the model with human preferences—a step requiring tens of thousands of human annotators to score the model's responses. Finally, they designed various evaluation benchmarks and repeatedly debugged to ensure the model did not become dumb or out of control (go out of control).

In this chain, AI was merely the object of training, while human researchers were the absolute subjects. Every leap in model capability was backed by the efforts of hundreds of top PhDs over months or even years. Human labor costs accounted for as much as 30% to 40% of overall training costs.

The bottlenecks of this model are evident. The energy, judgment, and creativity of human researchers became the ceiling for model evolution.

GPT-5.6's recursive self-improvement model completely rewrote this chain.

According to technical details disclosed by OpenAI, during Luna's training, the flagship model Sol acted as an automated researcher. It no longer passively accepted human-fed data and parameters but actively participated in four core stages.

The first stage was autonomous data selection. Sol could independently assess the quality, diversity, and potential biases of vast candidate datasets, deciding which data should enter Luna's training set and which should be excluded. This task previously required a data team of over a dozen people working for months.

The second stage was autonomous experiment design and execution. Sol could propose hypotheses for training strategies, design controlled experiments, run complete training processes on clusters, and analyze results. Whereas a junior researcher might run two or three experiments a day at most, Sol could run hundreds in parallel daily.

The third stage was autonomous knowledge distillation. As a lightweight model, Luna needed to inherit core capabilities from Sol while compressing parameter scale. This teaching-and-learning process was overseen by Sol itself as the teacher. It decided which knowledge was most important, how to compress it, and how to verify the compressed effects.

The fourth stage was autonomous evaluation and iteration. Sol could write its own evaluation test cases, identify Luna's weaknesses, then adjust training strategies and start over, forming a complete closed loop.

OpenAI released a set of intriguing internal data. Over the past six months, the company's internal computing resources for code reasoning have grown 100-fold, while token consumption for agent tasks has increased about 22-fold. On an internal benchmark measuring recursive self-improvement capabilities, Sol scored 16.2 points higher than the previous GPT-5.5 generation.

When Sol autonomously debugs training scripts in code repositories as an agent or analyzes results on evaluation platforms, it naturally consumes dozens or even hundreds of times more reasoning compute power and tokens than manual human operations.

A former OpenAI researcher, speaking anonymously, put it this way: 'We used to train a student; now we're training a teaching assistant. This TA never tires, doesn't need a salary, doesn't make rookie mistakes, and everything it learns can be immediately taught to the next student.'

Can Competitors Keep Up?

As OpenAI transforms recursive self-improvement from a lab concept into an engineered, mass-produced tool, where do its competitors stand?

Anthropic is OpenAI's strongest rival. Just days before GPT-5.6's release, Anthropic's Claude Fable 5 had held the world No. 1 spot on programming benchmarks for 17 days before being dethroned overnight by Sol.

But Anthropic's real crisis lies not in a single benchmark loss but in its relatively lagging Layout (deployment) of recursive self-improvement.

Anthropic's core strategy has always been a safety-first, premium route, conducting extremely rigorous safety alignment and red-teaming before model releases. This approach is highly effective in winning enterprise client trust but could become a drag in the self-improvement race. Because the essence of recursive self-improvement is enabling AI to act autonomously, while Anthropic's safety philosophy is precisely about limiting AI autonomy as much as possible.

However, Anthropic is not unaware of this issue. On the same day as GPT-5.6's release, Anthropic's co-founder publicly announced an industry-shaking decision: to stop hiring junior engineers. His exact words were, 'Large-scale experiments that once required armies of junior researchers can now be done by Claude itself. Our hiring standard now is senior intuition—we only recruit experienced individuals capable of directional judgment, not execution-level staff.'""This decision itself is the most direct confirmation of the arrival of the recursive self-improvement era. When Anthropic begins using Claude to replace junior researchers for experiments, it has effectively stepped onto this track—albeit at least six months later than OpenAI.

Google DeepMind possesses the world's most abundant AI R&D computing power and the deepest research accumulation. In May 2025, DeepMind released AlphaEvolve, which used the Gemini model to generate candidate algorithms, Filter ed (selected) superior solutions through automated evaluation and evolutionary search, and applied this method to data center scheduling, chip design, and AI training process optimization.

But DeepMind's problem lies in corporate bureaucracy. Every new technology's journey from lab to product requires layers of approval, safety reviews, and cross-departmental coordination. While OpenAI's Sol was already iterating Luna daily in production environments, DeepMind's related projects might still be awaiting quarterly reviews.

In March 2026, DeepMind launched the AutoML-X project, attempting to establish a fully automated model training pipeline. However, according to insiders, the project remains in limited experimental stages, far from achieving OpenAI's level of engineered mass production. For Chinese large model companies, the advent of the recursive self-improvement era may mean an even more brutal dimensionality reduction strike.

First is the computing power bottleneck. The core of this model is trading reasoning compute power for R&D efficiency, allowing flagship models to run experiments, write code, and conduct evaluations 24/7 as agents. Behind OpenAI's 100-fold increase in internal code reasoning compute power over six months lies its foundation of tens of thousands of top-tier chips. Domestic companies, constrained by chip export bans, already struggle with inadequate computing power reserves and can hardly support such extravagant AI-researching-AI models.

Second is the generational gap in deployment. While OpenAI and Anthropic are already using flagship models to train next-gen lightweight models, most domestic companies' main models are still chasing GPT-4-level foundational capabilities. This creates a virtuous cycle for the strong: the stronger your flagship model, the better its trained sub-models become. The better the sub-models, the faster they can accelerate the flagship model's iteration. Once this flywheel starts spinning, laggards face not linear gaps but exponential widening.

A technical lead at a leading domestic large model company admitted anonymously, 'We're still using human labor to chase their previous-gen models while they've already started using AI to train the next generation. It's like we're hand-sewing clothes while they've moved to fully automated production lines. The gap will only keep growing.'

The Twilight of Human Researchers

The engineered implementation of recursive self-improvement first impacts not competitors but AI industry practitioners themselves.

Over the past three years, the large model industry has spawned massive numbers of junior researcher positions dubbed 'hyperparameter tweakers.' Their daily work involves cleaning data, running ablation experiments, adjusting learning rates, and recording results—repetitive, low-creativity tasks that were nonetheless indispensable for model training.

Now, these roles are being replaced en masse by automated systems.

Anthropic co-founder's statement about 'no more junior engineers' is not prophecy but ongoing reality. According to Stanford research data, by mid-2025, employment of U.S. software developers aged 22–25 had dropped nearly 20% from the 2022 peak. In the first two months of 2026 alone, the global tech industry laid off over 150,000 workers, with AI being the primary cause of layoffs for three consecutive months—an average of 974 jobs lost daily.

A U.S. National Bureau of Economic Research survey of 750 CFOs revealed that AI-driven layoffs would reach about 500,000 positions in 2026, nine times the 2025 figure. In China, the situation is equally severe. Tech giants like Tencent, Alibaba, and ByteDance have generally cut 15% to 40% of their workforce, with entry-level and mid-skill roles hit hardest.

An engineer who worked as a junior researcher at a leading AI company recalled anonymously, 'When I joined in early 2025, our group had 12 junior researchers manually running experiments, adjusting parameters, and writing reports. By early 2026, the company implemented an automated experiment platform powered by its own models. Three months later, we were cut from 12 to 4. The remaining four could make directional judgments and design system architectures.'""He laughed bitterly, 'We spent two years training AI, and in the end, AI trained us out of the company.'""In stark contrast to the massive contraction of junior roles is the super-premium placed on top AI talent.

According to Maimai data from January–April 2026, AI scientists and team leads now average 132,800 RMB monthly—1.8 times the 74,400 RMB earned by second-tier algorithm researchers. OpenAI offered up to $445,000 annual salaries (over 3.2 million RMB) to hire safety experts researching how AI can safely train stronger versions of itself.

The logic is simple: when automated systems can replace 80% of foundational research work, the remaining 20% becomes disproportionately critical. These tasks include defining research directions, designing systems themselves, and judging whether AI outputs are reliable—roles requiring senior intuition and research taste, human capabilities machines cannot replicate short-term.

The AI industry's talent structure is rapidly reshaping from a pyramid to a dumbbell. Massive junior roles at the base are replaced by AI, while a small number of senior roles at the top see explosive salary growth, with the middle layer severely compressed.

Of course, technological change also creates entirely new roles. AI trainers are no longer annotators scoring model responses but advanced roles designing training strategies and defining behavioral boundaries for AI teaching assistants.

Demand for alignment engineers is exploding. When AI begins training AI, ensuring training processes don't deviate from human intentions or cause alignment failures becomes paramount—the cutting edge of AI safety.

Agent architects design multi-agent collaboration frameworks, enabling multiple AI researchers to divide tasks, verify each other's work, and avoid working in silos.

The problem is that these new roles have extremely high barriers, far beyond what laid-off junior researchers can easily transition into. A structural talent gap is quietly forming within the AI industry itself.

Final Scenario Projection

The recursive self-improvement flywheel has already started spinning. Where will it take the AI industry?

Within the next one to two years, automated training systems will become standard capability for all leading large model companies. Those without such systems will be completely outpaced in iteration speed, like factories lacking automated production lines.

OpenAI's Project Loop has invested $500 million, aiming to achieve fully autonomous model iteration by 2027. Anthropic and Google DeepMind are accelerating their pursuit. Domestic companies, constrained by computing power and talent bottlenecks, will likely fall behind in this race. The Sino-U.S. AI gap may widen from the current six months to one year to two to three years.

In the medium term, automated systems will take over more than 50% of large model training work. R&D teams will shrink dramatically, but per-capita output will increase 5–10 fold. A team of 10 top researchers plus an automated system will produce models that today require 100 people.

This means AI companies' organizational structures will fundamentally change, shifting from labor-intensive to capital-intensive R&D. Computing power investment and system quality will replace researcher headcount as the core variable determining model capability.

This is the ultimate question in AI safety. When systems become powerful enough to autonomously design next-generation systems, a true closed loop forms: AI designs better AI, which designs better training systems, which design even better AI. If this loop's iteration speed exceeds human capacity to understand and intervene, it could trigger so-called intelligence explosion.

AI capabilities have grown exponentially in an extremely short period, far surpassing human control. Anthropic divides recursive self-improvement into three stages. The first stage is AI-assisted coding, which is the current stage. The second stage is AI autonomously conducting experiments, which the industry is now entering. The third stage is complete AI autonomy in iteration, which has not yet arrived. Currently, the industry is at a critical transitional window from the first stage to the second.

The good news is that this model still faces several key bottlenecks. The first bottleneck is computational constraints. Each self-iteration requires massive amounts of inference computing power, and computational costs may become the physical ceiling limiting the flywheel's speed.

Alignment decay is the second bottleneck. When AI trains AI, each generation of models may introduce subtle alignment biases. If these biases accumulate and amplify over multiple generations of iteration, the final model may deviate from human intentions. The third bottleneck is the lack of research taste.

AI excels at executing well-defined experiments but still falls far short of top human researchers in proposing truly original research hypotheses or making counterintuitive directional judgments. These bottlenecks provide humanity with a precious window for intervention. However, this window is narrowing.

Conclusion:

On July 9, 2026, when the release page for GPT-5.6 refreshed across millions of screens worldwide, most people saw only stronger benchmarks, lower prices, and more user-friendly AI tools.

But only a tiny few noticed the truly historic detail: Luna was trained by Sol.

In science fiction, the technological singularity is often depicted as a thunderous explosion—machines awakening, humanity falling, and the world being reconstructed. However, in reality, the technological singularity is often announced in a casual remark, hidden in a technical document that no one reads closely.

It marks the moment when human researchers transition from being coaches to becoming referees. Over the past three years, the dominant narrative in the AI industry has been humans training AI. We have used more data, greater computational power, and more sophisticated algorithms to push model after model to new heights of intelligence.

But starting today, the narrative is quietly shifting toward AI training AI. The flywheel is already turning, and all humanity can do is ensure it still holds the brakes before the flywheel spins out of control. The singularity is not a thunderous explosion. It is a casual remark—and you almost missed it.