"Generative artificial intelligence (AI) is trapped behind an '80% wall' where accuracy fails to exceed 80%. No matter how high the benchmark scores are, it does not understand the field, so it fails to affect the productivity of corporations. This is why proven 'physical AI' that actually runs on corporate sites and delivers results is becoming more important."
On the 3rd at the Westin Seoul Josun Parnas, Chief Executive Yoon Seong-ho of MakinaRocks said in a keynote at the AI event "Attention 2026," "According to a study by a U.S. economic research institute, about 90% of corporations that adopted AI said they had not experienced a measurable productivity change on their income statements over the past three years."
Yoon said that while the intelligence of Generative AI models from leading AI corporations such as OpenAI, Anthropic, and Google is doubling every seven months, they still have not meaningfully improved corporate productivity. He cited McDonald's as an example, which introduced AI at more than 100 drive-thru (DT) locations in the United States in 2021 but halted it after three years. As AI repeatedly misrecognized customer orders and, for example, entered quantities for items like hamburgers and French fries as 200 instead of 20, McDonald's ultimately ended AI ordering services in 2024.
He said, "This is a representative case of AI not working properly in the field," adding, "McDonald's moved quickly to withdraw, but if such AI errors occur in a factory that produces cars every few minutes, in a semiconductor production line where equipment worth billions of won is deployed, or on a battlefield where not even the slightest error is tolerated, the result could be enormous damage."
He pointed out that today's widely used large language model (LLM)-based AI solves the easy 80% well but is stuck in the long-tail problem—the so-called '80% wall'—making mistakes on the complex 20%. Yoon said, "If you ask AI to analyze a file that organizes semiconductor design data and draw a trend line, it will draw it accurately, but if you ask in a production setting how to respond when a specific sensor value has risen above 40, it cannot answer because it has no sense of the numbers' context," adding, "The field can never operate at 80% accuracy."
He went on, "Cutting-edge AI corporations fixate on benchmark performance of AI models, but that has nothing to do with corporate productivity," adding, "Because they have never experienced field data and do not know how the work is actually done, they do not solve long-tail problems at all."
Know-how that field experts have accumulated over decades—such as drawing interpretation, reading sensor values, and detecting equipment anomalies—has never been made public online, meaning Generative AI has not learned it. He said, "That is why, no matter how high the scores and even if it passes bar and medical exams or solves math challenges, it still cannot properly take simple orders at a McDonald's drive-thru."
Yoon argued that corporations must secure "AI sovereignty" to strengthen AI competitiveness. He said corporations should turn their accumulated proprietary data and work know-how, field experts' decision criteria, and AI training outcomes into a core corporate asset. He added that, at the same time, they need the option to freely switch and combine models and infrastructure as needed, without being locked into a specific foundation model or graphics processing unit (GPU) and cloud infrastructure.
He said MakinaRocks' AI operating system "Runway" realizes corporations' "AI sovereignty" by unifying distributed environments—from data centers to factory servers to edge devices—and centrally managing the entire process of developing, deploying, and operating AI models and agents. Yoon said, "As for outcomes MakinaRocks has achieved in the field, we saved more than 1,000 hours per year in design document change reviews, and we became the first corporation to pilot AI on a military battlefield network during the U.S.-Korea combined training (UFS)."
MakinaRocks has applied AI at more than 80 customer sites across 32 cities in five countries through its FDE organization in charge of on-site deployment. This year, MakinaRocks' FDE has spent 70,000 hours in the field and has deployed more than 6,000 AI models in total.
Yoon said, "MakinaRocks has focused on building AI that breaks through the 80% wall and runs in the field," adding, "From factories to battlefields, we will compete with AI that delivers results in the roughest and most unpredictable environments." To that end, he said the company has launched a project to build the initial shape of a fully autonomous manufacturing plant.
Marking its third year, "Attention 2026" is an industrial AI event that MakinaRocks has held annually since 2024. Under the theme "AI Has Landed," about 700 attendees from manufacturing, energy, semiconductors, defense, and the public sector joined this year's event. Speakers from major corporations, including Samsung Electronics, HD Hyundai, Hankook & Company, and Samsung SDS, discussed how AI has moved past the proof-of-concept (PoC) phase and begun delivering results on actual industrial sites.