AI systems are increasingly being used to write software, but a more consequential shift is beginning to emerge: using AI to improve the systems and infrastructure used to build AI itself. Chinese AI company ModelBest says a series of experiments with AI-generated training and optimization software demonstrates how recursive self-improvement could reduce the engineering and computing costs involved in developing new models.
The AI industry has spent years trying to make models better at writing code. The next step could be using those coding capabilities to improve the machinery that trains AI models in the first place.
That idea—often described as recursive self-improvement (RSI)—is gaining attention as coding agents become capable of working across increasingly complex software engineering tasks. Rather than simply generating snippets for human developers, these systems can potentially design, test and optimize parts of the infrastructure used to develop subsequent AI systems.
Chinese AI company ModelBest says it has demonstrated several examples of this approach through a technical paradigm it calls Forge Engineering.
The company’s recent work includes an AI-engineered pre-training framework, an agent that iteratively optimized model-training infrastructure, and an automated system for optimizing software stencils. ModelBest developed the systems in collaboration with the OpenBMB open-source community.
The experiments remain company-reported results rather than independently verified industry benchmarks. But they point to an important question for the AI industry: what happens when AI coding agents become capable of improving the tools used to build AI?
When AI starts working on the AI development stack
In conventional machine learning development, engineers design training infrastructure, write optimization code and tune systems manually.
That process can be expensive.
Large-scale model development depends on sophisticated software stacks for distributed training, memory management, hardware utilization, data processing and inference. Improvements to those systems can translate directly into lower compute costs or shorter training cycles.
ModelBest’s approach attempts to put AI agents into that engineering loop.
At the end of May, the company and OpenBMB released ForgeTrain, described as a production-level pre-training framework engineered by AI without a human in the loop.
ModelBest says ForgeTrain was used to train the base model of MiniCPM5-1B, which the company describes as a leading small language model according to AA Index. The company reports that its training speed was approximately 10% faster than NVIDIA’s Megatron framework.
That comparison is significant because Megatron has become an important reference point for large-scale model training infrastructure. But benchmark comparisons can depend heavily on hardware, model architecture, training configuration and workload, making independent reproduction important before drawing broad conclusions.
An AI agent optimizes the training framework
ModelBest subsequently pushed the concept further.
At the World Artificial Intelligence Conference (WAIC) in July, the company says it developed an agent powered by GLM-5.2 that iteratively optimized its training framework over an 18-hour period.
The optimized framework was then used to train MiniCPM5-130M from scratch.
According to ModelBest, the resulting model reached performance comparable to Google’s Gemma 3 270M, despite having a smaller parameter count.
The important element is not simply the final model benchmark. The experiment represents an attempt to automate part of the optimization process that normally requires experienced machine-learning engineers.
An AI agent was effectively being asked to improve the machinery used to train another AI model.
That creates a feedback loop: an existing model helps improve development infrastructure, which is then used to create or train another model.
ForgeStencil takes the idea beyond model training
ModelBest’s August release, ForgeStencil, applies a similar philosophy outside conventional language-model training.
The company describes ForgeStencil as a fully automated system for optimizing and deploying stencils used in industrial and scientific software. ModelBest says the system optimized more than 100 stencils over a single week without a human in the loop.
Stencil computations are particularly relevant to scientific and engineering workloads, where performance can depend heavily on how operations are mapped to hardware.
The broader idea is that AI agents can generate highly specialized implementations rather than relying exclusively on generalized software frameworks.
That is the foundation of ModelBest’s Forge Engineering concept.
From general-purpose software to AI-forged systems
For decades, software engineering has relied heavily on general-purpose frameworks. Developers build reusable systems designed to serve a broad range of applications.
That approach offers significant advantages in maintainability and development speed, but generalized software can leave performance on the table because it is not optimized for every individual workload.
AI coding agents introduce another possibility.
If an AI system can understand a specific workload, generate specialized code, test different implementations and select the best-performing version, customized software could become cheaper to produce.
That is essentially the premise behind Forge Engineering: use AI to make specialized software economically viable at a much larger scale.
The concept is closely connected to trends already visible across the technology industry. NVIDIA is optimizing computing stacks around its GPU architecture, while companies such as Google, Microsoft, Anthropic and OpenAI are investing heavily in coding agents and AI development tools.
The difference is where ModelBest is placing the agent in the stack.
Instead of using AI primarily to accelerate human software development, the company is exploring whether AI can take over increasingly large portions of the optimization process itself.
Recursive improvement could change AI economics
The implications go beyond developer productivity.
Model training is constrained by the availability and cost of compute. Even modest improvements in training efficiency can become economically meaningful when applied to large workloads.
The AI industry is already investing heavily in infrastructure to address those constraints. Improvements in GPU utilization, distributed training and specialized accelerators can determine how quickly and cheaply a model can be developed.
If AI agents can autonomously identify and implement those optimizations, they could create a compounding effect.
A more capable coding agent could produce better training infrastructure. Better infrastructure could enable cheaper experimentation. More efficient experimentation could produce better models, which could in turn become more capable coding and optimization agents.
That is the basic logic behind recursive self-improvement.
It is also where the concept becomes more speculative.
The current demonstrations do not prove that AI systems can autonomously improve their own general intelligence indefinitely. They show something narrower and potentially more practical: AI models can already participate in tightly scoped engineering loops where software is generated, tested and optimized with reduced human intervention.
The human role may shift rather than disappear
There is another important distinction.
“AI-engineered” does not necessarily mean that human expertise becomes irrelevant.
Even highly autonomous development systems require choices about objectives, evaluation criteria, safety constraints, compute budgets and deployment environments. Human engineers also remain responsible for determining whether an optimization is genuinely useful outside a controlled benchmark.
For enterprises, this distinction will matter.
The near-term opportunity may be less about completely autonomous AI research and more about using coding agents to automate repetitive optimization tasks that currently consume specialist engineering time.
That could apply to model training, inference optimization, data pipelines, compiler workloads, scientific computing and enterprise software.
ModelBest’s experiments offer an early glimpse of what that future could look like.
If the approach continues to improve, the economics of software specialization could change alongside the economics of AI development. Instead of asking developers to adapt general-purpose systems to every workload, organizations could increasingly ask AI agents to forge the software they need for a specific task.
The result would not necessarily be machines improving themselves without limits. It could be something more tangible: an AI development ecosystem in which increasingly capable agents continuously optimize the tools, code and infrastructure used to build the next generation of AI.
Market Landscape
The development of AI coding agents is creating a new layer in the AI infrastructure market.
Companies such as Anthropic, OpenAI, Google and Microsoft are pushing coding agents toward longer-running software engineering tasks. At the infrastructure layer, NVIDIA and other chip and cloud providers are optimizing the systems that support model training and inference.
ModelBest’s Forge Engineering concept sits at the intersection of these trends.
Its premise is that AI-generated software can be specialized for individual workloads at a lower cost than traditional human engineering. If that assumption holds at scale, it could challenge the economics of general-purpose frameworks.
The potential market impact extends beyond LLMs. Autonomous optimization could eventually be applied to AI training frameworks, compilers, scientific computing, HPC workloads, inference engines and enterprise software.
However, reproducibility remains crucial. Performance claims involving training speed, model benchmarks and optimization gains need to be evaluated under comparable hardware and workload conditions before they can establish a broader industry benchmark.
Top Insights
- ModelBest is applying AI to AI development infrastructure, using coding agents to build and optimize training frameworks rather than simply assist human programmers.
- ForgeTrain reportedly improved training speed by 10% over Megatron, illustrating how AI-generated infrastructure could reduce compute requirements for model development.
- An AI agent optimized a training framework in 18 hours, then used the resulting system to train MiniCPM5-130M from scratch.
- ForgeStencil extends autonomous optimization beyond AI models, with ModelBest reporting more than 100 industrial and scientific software optimizations in one week.
- Forge Engineering challenges general-purpose software economics, suggesting AI could make highly customized software development cheaper and increasingly autonomous.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
