For many years, discussions about artificial intelligence have largely centered around one question. It’s a way to build bigger, more capable models. But as AI increasingly moves beyond screens into robots, autonomous systems, and real-world machines, other questions are emerging. Can reliability, safety, and decision-making challenges in physical environments be solved at model scale alone, or does physical AI require a fundamentally different approach?
Physical AI challenges common assumptions about AI
With the recent rise of generative AI, deep learning has become the main face of artificial intelligence. Large-scale language models, multimodal systems, and increasingly powerful underlying models are driving much of the industry’s attention, investment, and technological ambition.
However, some engineers working on physical systems argue that: Equating AI with deep learning risks oversimplifying the broader field.
According to raymond kimFounder and CEO of aid allis a Seoul-based robotics company focused on autonomous mobility technology, but the industry is now increasingly understanding AI. Shaped by market narratives Rather than a complete history of intelligent systems.
“Since ChatGPT, ‘AI’ has become almost synonymous with large-scale deep learning models.”
Mr. Kim said this in an exclusive interview with KoreaTechDesk.
“That equation is a market shortcut, not a definition.”
Kim then cited technologies such as fuzzy logic control systems, anti-lock braking systems (ABS) in automobiles, fly-by-wire systems in aircraft, and Mobileye’s Responsible Safety (RSS) framework. AI-related systems have been operating in safety-critical environments for decades. No need to rely on the latest deep learning approaches.
This distinction will become increasingly important as AI systems begin to interact directly with the physical world.
“Shortcuts work on screen, but they stop working the moment the system enters the physical world.”
Kim said.
Why physical AI raises all kinds of technical questions
physics AI Generally refers to AI systems embedded in machines. Recognize, reason, and act within yourself genuine environment. This includes: autonomous robots, industrial systems, mobility platforms, others intelligent machine Operating outside of a controlled digital setting.
Unlike purely software-based applications, physical AI systems must function under strict constraints get involved power consumption, latency, environmental uncertainty, and Real-time decision making.
The past year has seen a significant increase in industry attention regarding physical AI. Nvidia Explained physical AI as a new category of intelligence Enabling autonomous systems to understand and interact with the real worldMeanwhile, researchers and technology companies continue to explore architectures optimized for robotics and edge computing.
For Kim, the challenge isn’t just finding one good model.
“At AidALL, we treat this as an architectural problem, not a philosophical one.”
Go beyond deep learning alone
One of the emerging themes in physical AI research is the growing interest in: Combine multiple AI paradigms Rather than relying solely on deep learning systems.
Kim said that AidALL’s internal architecture combines: Deterministic computation, neuromorphic design principles, and neurosymbolic reasoning layers.
“We are combining deterministic computation, neuromorphic design principles, and layers of neurosymbolic reasoning into a single system.”
he explained.
Neurosymbolic AI is gaining increasing attention among researchers seeking to combine the pattern recognition strengths of neural networks with the reasoning power of symbolic systems. Research organizations such as IBM describe the neurosymbolic approach as a way to: Improve explainability, reasoning, and decision transparency With complex AI systems.
Neuromorphic computing, on the other hand, takes inspiration from biological nervous systems. improve efficiency and Reduce computational demands. Researchers increasingly see neuromorphic architectures as a promising direction for robotics, edge devices, and resource-constrained environments where energy efficiency is critical.
These approaches do not necessarily replace deep learning. Instead, they suggest that future physical AI systems may require: A broader toolbox.
“Change in physical AI is not about finding new models; it’s about asking which paradigm fits which problem.”
Why reliability is as important as intelligence
Conversations become more complex when an AI system is expected to: The decision directly impact physical consequences.
Publicly available AI benchmarks often focus on model accuracy, inference performance, or task completion rate. However, real-world implementation introduces additional requirements that benchmark scores may not fully capture.
The International Organization for Standardization’s ISO/IEC TR 5469 guidance on AI and functional safety notes that machine learning systems can pose challenges regarding explainability, repeatability, and assurance for safety-related applications.
Mr. Kim argues that Some important system characteristics cannot simply be revealed through additional data or larger models.
“The most repeated misconception is that ‘scale is the answer.'”
he believes in certain requirements It sits outside the normal scaling curve of modern AI development.
“Determinism is not something that can be achieved by adding statistical resources; it is a property of the computational structure itself.”
This perspective reflects broader discussions occurring across the robotics, autonomous systems, and edge AI communities. While larger scale models continue to improve perception, language understanding, and generalization, engineers are increasingly faced with questions about the traceability, predictability, and reliability of these systems as they interact with the physical environment.
Extensive discussion on the future of AI
The debate about physical AI is ultimately not about choosing one AI paradigm over another.
Deep learning continues to play a central role in many of the industry’s most important advances. At the same time, researchers and engineers working on real-world systems are increasingly conducting investigations. How different computational approaches can complement each other.
That change is already visible in areas such as robotics, autonomous mobility, industrial automation, and edge intelligence.
And for companies building systems that need to operate beyond controlled digital environments, the next stage of AI development will likely depend on more factors than just model size. How different forms of intelligence are combined into a reliable architecture.
Architecture questions
The history of AI is often explained through breakthroughs in models. But today, physical AI could propel the industry toward a different conversation.
As intelligent systems move toward machines that navigate, assist, and operate in real-world environments, the central challenge may no longer be about how big a model can be. Maybe so How reliably can the entire system recognize, reason, and respond? when the situation is unpredictable and carry mistakes genuine result.
And ultimately, for engineers building the next generation of physical AI systems, that distinction is rapidly becoming a defining engineering reality, determining which technologies succeed in the physical world and which remain confined to the lab.
Important points
- Physical AI shifts attention Beyond model scale System architecture, reliability, and actual operations.
- Raymond Kim of AidALL argues: AI should not be treated as a synonym for deep learningEspecially in physical systems.
- Deterministic computing, neuromorphic computing, neurosymbolic AI has emerged as an important complementary approach in physical AI development.
- Research and industry interest in physical AI continues to grow As AI expands into robotics, autonomous systems, and edge computing.
- According to Kim, some important properties, such as deterministic guarantees, This cannot be achieved simply by using larger models or more training data.
- Industry-wide discussions are increasingly focused on: Which AI architecture is best suited for real-world deployment? Not just the scale of the model.
🤝 Do you want to connect with established Korean companies operating all over the world?
Browse carefully selected company profiles and request a direct introduction. beSUCCESS connect.
– Stay one step ahead of the Korean startup scene –
Get real-time insights, funding updates, and policy changes shaping South Korea’s innovation ecosystem.
➡️ Follow KoreaTechDesk LinkedIn, X (Twitter), Threads, Bluesky, Telegram, Facebookand WhatsApp channel.
