Conclusion

Let’s get straight to the point: “The intuition of craftsmen” can now be systematized.

When a veteran worker at a small factory retires, the quality often plummets. This fear is something manufacturing business owners can relate to.

Currently, three technologies are emerging simultaneously, fundamentally changing this structure. The industrial process optimization AI agent “little m,” LLM-based synthetic data generation, and CNC transfer learning. Each of these technologies is interesting on its own, but when combined, they present a compelling case for small and medium-sized manufacturers: “You can eliminate dependency at a monthly cost.”

Let’s look at what’s happening in detail, supported by specific numbers.

—

1. “little m” — Solving Factory Optimization Problems with Natural Language

Manufacturing accounts for about one-third of global energy consumption. Yet, many small factories operate based on the “experience of veterans.” Temperature settings, pressure, timing—implicit knowledge that is not quantified drives the heart of the factory.

“little m” aims to convert this implicit knowledge into mathematical optimization models.

Traditionally, formulating industrial processes mathematically required the expertise of operations research specialists. Translating specifications written in natural language into precise mathematical formulas could cost hundreds of thousands of yen and take several months.

“little m” automates this translation by combining a domain-specific knowledge repository with an LLM (large language model). Research has shown that it significantly outperforms general-purpose LLMs (like GPT-4) in terms of the accuracy of optimization models.

The important change here is in the cost structure.

Traditionally, outsourcing an operations research expert would cost between 3 million and 5 million yen per project. Using an AI agent like “little m” via an API, the inference cost is only a few hundred to a few thousand yen per instance. Even if optimizations are run dozens of times a month, the total monthly cost would be around 50,000 to 150,000 yen.

In other words, it shifts from “hiring an optimization expert” to “subscribing to optimization services.” This is a game changer for small and medium-sized manufacturers.

However, there is a caveat. At this stage, “little m” is still a research project and cannot be contracted as a commercial service. Initial model building and domain adaptation will require organizing and feeding in your own process data. It is realistic to expect initial costs of around 500,000 to 1.5 million yen for this.

—

2. Synthetic Data Generation — “Lack of Data” is No Longer an Excuse

The first wall that small and medium-sized manufacturers face when trying to implement AI is: “We don’t have data.”

This statement is half true and half false. Data is often accumulated in sensors and PLCs. However, there is a lack of “labeled data in a format usable for machine learning.” Datasets that are tagged and organized for abnormal and normal conditions are typically nonexistent in small factories.

This is where LLM-based synthetic time series data generation comes into play.

Recent studies have proposed methods for generating realistic time series data by providing conditions of the manufacturing process (temperature profiles, processing speeds, material characteristics) as prompts to the LLM. By combining this with RAG (Retrieval Augmented Generation), the diversity of the data can be expanded while preserving the statistical properties of real data.

Research results indicate that models trained on this synthetic data perform equally or better in anomaly detection compared to traditional methods (like ARIMA or LSTM) that only use real data.

Here’s the cost breakdown:

• LLM fine-tuning costs: 100,000 to 300,000 yen using cloud GPUs
• Data preparation and preprocessing for RAG: 100,000 to 200,000 yen
• Total: You can overcome the “lack of data problem” for 200,000 to 500,000 yen.

Traditionally, creating an equivalent labeled dataset manually would cost between 2 million and 5 million yen, including sensor installation and annotation, and take over six months. This difference is significant.

Of course, synthetic data is not a panacea. There is a risk that discrepancies in distribution between synthetic and real data can affect the accuracy of downstream tasks. This is why a “hybrid learning” approach, combining a small amount of real data with synthetic data, becomes a practical solution. If one month of real data combined with synthetic data yields the learning effect equivalent to twelve months of real data, the barrier to implementation dramatically lowers.

—

3. CNC Transfer Learning — Deploying Knowledge Learned from One Machine to All

In small and medium-sized manufacturing sites, CNC (Computer Numerical Control) machines from different manufacturers and years are lined up. A lathe from Company A, a machining center from Company B, models from ten years ago and three years ago. The number and types of sensors also vary.

Traditionally, AI models needed to be trained individually for each machine, costing tens of thousands to 100,000 yen per machine. If there are five machines, that totals 500,000 yen. This is unaffordable for small businesses.

The focus here is on a transfer learning framework called “SAAC-JEPA (Schema-Adaptive Action-Conditioned Joint-Embedding Predictive Architecture).” While the name is long, the concept is simple: a technology that allows models learned from one machine to be applied to another machine with a different sensor configuration.

If there are partially overlapping sensors (for example, both machines have vibration sensors, but only one has a temperature sensor), the model can be adapted using the common elements as a foothold. It maintains prediction accuracy while absorbing differences in dynamics and interfaces.

The cost structure changes as follows:

• Traditional: Model building for each machine → 800,000 to 1.5 million yen per machine
• After introducing transfer learning: Base model construction costs 1 million to 2 million yen + adaptation for additional machines costs 100,000 to 200,000 yen per machine.

For installing AI in five machines, the traditional cost would be 4 million to 7.5 million yen. With transfer learning, it would be 1.4 million to 3 million yen. This is less than half the cost.

Even more importantly, there are cost reductions during the operational phase. Standardizing machine state predictions eliminates the dependency where “only A-san understands that machine.” If the early detection of anomalies is automated, the veteran’s “I can tell by the sound” can be replaced with a dashboard that anyone can use. The estimated monthly maintenance cost reduction is 50,000 to 150,000 yen per machine.

—

What Happens When You Combine the Three

Now, let’s get to the main point. It’s meaningless to look at these three technologies in isolation; we need to observe the structural changes when they are combined.

Step 1: Solve the “lack of data” issue with synthetic data (200,000 to 500,000 yen)
Generate synthetic data from a small amount of real data using LLM. Quickly build the dataset needed to train anomaly detection models.

Step 2: Deploy “full machine coverage” with CNC transfer learning (base 2 million + 200,000 yen for each additional machine)
Expand the model created from one machine to other machines in the factory. Adaptation is possible even with different sensor configurations.

Step 3: Optimize the entire process with “little m” (monthly cost of 50,000 to 150,000 yen)
Not only monitor the state of individual machines but also optimize the energy efficiency and production schedule of the entire factory through mathematical optimization.

Assuming a small factory with 30 employees and five CNC machines:

• Initial investment: Total 3 million to 5 million yen
• Monthly operational cost: 150,000 to 300,000 yen
• Expected monthly cost reduction: Maintenance costs of 250,000 to 750,000 yen + energy optimization of 100,000 to 300,000 yen

The payback period for the initial investment is 6 months to 1 year. In monthly terms, the cost of eliminating dependency effectively becomes zero or even positive.

—

So, What Should Be Done?

I’m not saying to implement everything at once. A realistic step for small businesses is as follows:

First, start with synthetic data generation. This has the lowest cost (200,000 to 500,000 yen) and allows for the quickest verification of effectiveness. You can confirm whether an anomaly detection model works on your one machine within 2 to 3 months.

Next, expand horizontally with transfer learning. Apply the model that worked on one machine to others. This breaks the traditional notion of “creating AI individually for each machine.”

Finally, layer on process optimization. Once monitoring data for individual machines is in place, proceed to optimize the entire factory.

Following this order, the first step can be taken for under 500,000 yen.

—

Risks to Consider

It’s pointless to only discuss optimistic scenarios. I’ll also outline the current risks:

1. Maturity of Technology: Both “little m” and SAAC-JEPA are still in the research phase. Until they are stably provided as commercial services, the involvement of integrators will be necessary.
2. Data Quality Barrier: While synthetic data is convenient, if the quality of the underlying real data is low, the synthetic data will also be unusable. A minimum level of sensor maintenance and a few weeks of real data collection are prerequisites.
3. Resistance on the Ground: It’s essential to explain that “AI distributes the wisdom of veterans to everyone” rather than “AI will take away jobs.” If this is misunderstood, no matter how good the technology is, it won’t take root in the workplace.

—

Dependency Has Become a “Cost Issue”

In the past, eliminating dependency was a matter of “education” and “manual creation.” Asking veterans, creating procedure manuals, and teaching younger employees took ten years.

Now, dependency is shifting to a cost issue of “how much do we need to pay monthly to eliminate it?” By learning from synthetic data, expanding with transfer learning, and optimizing with AI agents, when these three elements come together, the implicit knowledge that resided in the mind of one veteran can be transformed into a system costing 150,000 to 300,000 yen per month.

I want to ask: When your factory’s veteran retires next year, will that knowledge remain? Is 300,000 yen a month too high to ensure it does?

The answer is probably already clear.