The Era of AI Training AI Has Arrived: An In-Depth Look at Three Cutting-Edge Trends

Introduction: AI Development Enters a New Phase of Self-Iteration

The field of artificial intelligence is undergoing a quiet yet profound transformation. According to multiple research developments disclosed in Issue 449 of the prominent AI newsletter ImportAI, current AI technology is accelerating along three critical paths — large language models (LLMs) are beginning to train other LLMs, distributed training at the 72-billion-parameter scale has become a reality, and the challenges of computer vision are proving far more complex than generative text. At the same time, the potential "political interregnum" that AI could trigger has become a hot topic in academic circles.

Core Trend One: LLMs Training LLMs — The Self-Evolution Flywheel Begins to Spin

In the past, training a large language model required massive volumes of human-annotated data and carefully designed human feedback mechanisms. However, the latest research shows that using a powerful LLM to generate training data, evaluate output quality, and guide the learning process of another LLM has become a viable and efficient new paradigm.

This "AI training AI" approach offers multiple advantages. First, it dramatically reduces dependence on human annotation, significantly accelerating model iteration cycles. Second, teacher models can generate targeted training data addressing the student model's weak points, enabling more precise capability improvements. More importantly, this approach opens the door to continuous AI self-improvement — when models can effectively train the next generation, the pace of technological progress could grow exponentially.

However, researchers have also flagged potential risks: if training data is generated entirely by AI, models may fall into an "echo chamber effect," continuously amplifying their own biases and errors. Striking the right balance between efficiency and quality will be the central challenge facing this paradigm.

Core Trend Two: 72-Billion-Parameter Distributed Training Breaks the Compute Barrier

For a long time, training ultra-large-scale models has been monopolized by a handful of tech giants with access to top-tier computing infrastructure. The recently completed 72-billion-parameter distributed training experiment is beginning to shake up this landscape.

The experiment successfully completed the full training of a 72-billion-parameter model by distributing training tasks across multiple compute nodes in different geographic locations. This means that even without centralized supercomputing clusters, research teams can train competitive large models by coordinating distributed resources.

The significance of this breakthrough extends beyond the technical level. It signals that the "democratization" of AI training is accelerating — more research institutions, startups, and even open-source communities can expect to participate in the development of frontier large models. The maturation of distributed training technology could fundamentally reshape the current AI competitive landscape dominated by a few giants.

Of course, distributed training also faces a series of engineering challenges including communication latency, gradient synchronization, and fault tolerance mechanisms. But the successful validation at the 72-billion-parameter scale has undoubtedly injected strong confidence into the entire industry.

Core Trend Three: Computer Vision Is Far More Difficult Than Generative Text

Amid the global wave of generative AI, progress in text generation has been plain for all to see. However, analysis in ImportAI Issue 449 reveals an easily overlooked fact: the complexity of computer vision tasks far exceeds that of generative text.

Text generation is essentially sequence prediction within a discrete vocabulary space, whereas computer vision requires processing continuous pixel spaces, complex spatial relationships, and multi-level semantic understanding. From object detection to scene understanding, from 3D reconstruction to visual reasoning, the problem dimensions involved in computer vision are far more complex than simply "generating the next word."

This cognitive gap is also reflected in industrial applications. Although text generation tools like ChatGPT have already been widely commercialized, fields dependent on computer vision — such as autonomous driving, industrial quality inspection, and medical image analysis — still face numerous unresolved technical bottlenecks. Researchers are calling on the industry to devote more attention and resources to visual AI, cautioning against letting the "text generation hype" overshadow the equally critical visual intelligence track.

Deep Analysis: AI Could Trigger a "Political Interregnum"

Beyond technical breakthroughs, AI's potential impact on socio-political order has also sparked serious discussion. Some scholars have proposed that the rapid development of AI could lead to a "Political Interregnum" — a gap phase in which existing governance frameworks and institutional arrangements have not yet adapted to the upheaval brought by AI, while new governance paradigms have yet to be established.

During this transitional period, structural changes in labor markets, deep reshaping of the information ecosystem, and redistribution of power could all deliver shocks to the existing political order. If policymakers fail to respond to these changes in a timely manner, society may experience a period of turbulence filled with uncertainty.

This warning is far from alarmist. From the EU's AI Act to executive orders in the United States, governments around the world are already accelerating AI governance efforts. Yet a difficult-to-bridge gap persists between the speed of technological development and the lag in regulatory response.

Outlook: Balancing Technological Acceleration with Governance

Taken together, current AI development presents a complex picture of accelerating technology alongside mounting challenges. The rise of the LLM self-training paradigm will speed up model iteration, breakthroughs in distributed training will lower barriers to participation, and the deep challenges of computer vision remind us that the boundaries of AI capability are far from reaching their ceiling.

Looking ahead, the technology community must place greater emphasis on safety alignment, data quality, and social impact while pursuing performance breakthroughs. Policymakers, for their part, need to establish more agile governance mechanisms to address the profound social transformations brought about by AI technology. Only when technological innovation and institutional development advance hand in hand can we ensure that AI development truly benefits all of humanity.