Tiếng Việt
Artificial Intelligence (AI) — Concepts, nature, and practical applications explored through a conversation with the Gemini chatbot. The author's personal observations are highlighted in blue-purple note blocks.
Lately, everyone seems to be talking about AI, so I decided to ask a chatbot to see what the buzz is all about.
1. Definition of AI
Artificial Intelligence (AI) is a branch of computer science focused on building machines and systems capable of simulating human intelligence. The primary goal is to enable machines to perform tasks that typically require human cognition and intellectual capabilities.
Comments: Based on the aforementioned definitions, Artificial Intelligence (AI) can be broadly summarized as follows: Artificial Intelligence (AI) is a branch of computer science focused on developing methods to solve problems where inputs, outputs, or processing rules cannot be fully described, measured, or defined using traditional programming approaches, or problems that cannot be fully captured by static, fixed rules. For these types of problems, absolute 100% accuracy cannot be guaranteed; instead, the system delivers solutions with the best possible accuracy or effectiveness based on the trained data and models. In practice, AI is particularly valuable for problems with massive volumes of data and state spaces so large that defining explicit rules through traditional programming becomes entirely inefficient.
Successive iterations or versions of an AI model do not inherently guarantee across-the-board superiority over earlier releases, as performance hinges on multiple variables. The lifecycle of an AI model generally consists of developing a Foundation Model followed by subsequent fine-tuning via software pipelines combined with human feedback. Creating the baseline Foundation Model accounts for 80–95% of the total training cost (however, this ratio varies over time and depends on criteria such as model size, model type, data, GPU, etc). Conceptually, this phase resembles arranging tokens for a fill-in-the-blank word game - in terms of visualization, it can be easily visualized to be organizing closely related data points adjacently within a multidimensional vector space. Even the model developers cannot fully control every nuance of this spatial alignment, as the optimization routine runs millions of iterations and stops when reaching maximum computational limits or training steps. AI model updates strive to improve overall evaluation quality with each version iteration.
Typical application scenarios include predicting human behavioral trends, Q&A systems, language translation, computer vision, speech recognition, and modern conversational AI chatbots. Conversely, for problems that can be solved deterministically using classic algorithmic logic, applying AI rarely yields meaningful performance or accuracy advantages. A defining trait of modern AI is its heavy reliance on statistical methods, Machine Learning, Deep Learning, and data science methodologies to train models from source data. Post-training and optimization, AI models can be deployed to serve end-users through virtual assistants, chatbots, or enterprise integrations. Today, some tech providers offer broader definitions, asserting that any software or hardware system capable of self-learning, expanding training data, or improving reasoning constitutes AI. However, this definition presents governance, safety, and legal challenges. Unchecked model learning mechanisms can lead to unexpected behaviors, complicating transparency, accountability, and safety. For instance, models trained on biased or toxic data may output harmful content. In highly autonomous physical systems like robotics or autonomous vehicles, training data distribution shifts can result in unintended decisions. This stems directly from the probabilistic nature of contemporary AI models: they operate on statistical inference, making their outputs probabilistic rather than absolute, particularly when encountering out-of-distribution scenarios.
As of now, AI software can assist in handling tasks ranging from roughly 10% to 60% of raw, unoptimized task workloads depending on domain complexity (across overall industries, though office/knowledge work may see higher ratios). However, human review remains indispensable to decide whether to adopt, modify, or discard AI-generated outputs.
Capabilities and Characteristics of AI
AI capabilities encompass:
- Learning and Adaptation: AI learns from data, detects implicit patterns, and refines performance over time without requiring explicit rule programming for every case. This forms the foundation of Machine Learning and Deep Learning.
- Reasoning and Problem-Solving: AI can evaluate information, draw logical inferences, and solve complex problems akin to human thought processes.
- Language Understanding and Generation: AI leverages Natural Language Processing (NLP) to comprehend text or speech and synthesize text, poetry, or source code.
- Image and Speech Recognition: AI analyzes visual inputs to recognize faces or objects and processes spoken audio commands.
- Decision Making: Guided by data and algorithmic logic, AI can make decisions across varied contexts, ranging from recommendation engines to medical diagnostics.
- Creativity: Modern generative models produce synthetic art, music compositions, and diverse textual contents.
How Does AI Work?
AI operation relies heavily on complex algorithms and massive datasets.
- Artificial Neural Networks (ANN): A computational architecture inspired by biological brain structures. ANNs consist of interconnected artificial neurons processing data and learning from pattern representations. ANNs established the foundation for Deep Learning and serve as the ancestral lineage of modern LLMs. Deep learning evolved from multi-layer neural networks, while modern LLMs represent a major branch of deep learning, particularly those based on the Transformer architecture.
- Machine Learning (ML): A subfield of AI that enables systems to learn automatically from data without explicit task-specific programming. It encompasses supervised, unsupervised, and reinforcement learning paradigms.
- Deep Learning (DL): An advanced branch of ML utilizing multi-layered neural networks to process and analyze large, high-dimensional datasets, proving particularly effective in computer vision, speech recognition, and NLP.
Comments: Most modern AI systems utilize Machine Learning models, particularly Deep Neural Networks, to learn from data and execute tasks such as recognition, prediction, or content generation. The fundamental difference between modern AI and the classical AI era (prior to around 2012/2017) lies in how the system processes and leverages data. The year 2017 marked a significant milestone with the advent of the Transformer—an innovation that subsequently paved the way for the development of today's large language models—making it possible to view this period as the transition to modern AI.
It is easy to conceptualize training data as an address book or a map. During the classical AI era, the utilization of training data was relatively straightforward: the system primarily looked up entries explicitly recorded in the book—what was listed was correct, and what was absent was wrong. "Learning" at that time was essentially appending new addresses to the instruction manual, and when encountering a real-world scenario, the system simply queried the record to select a path.
In contrast, in the modern AI era, training data offers far greater value. Modern machine learning models do not merely memorize input data; they capture underlying rules, patterns, and hidden statistical relationships. Consequently, they possess the ability to infer and predict "new addresses"—scenarios that have never appeared verbatim in the training data—with highly impressive accuracy.
Competitive differentiation among technology firms stems not merely from dataset volume, but from data quality, neural architecture design, compute infrastructure, and training methodologies. Major tech firms construct large-scale models capable of capturing nuanced relationships, yielding superior generalization capabilities on unseen data.
Modern AI training is far from fully automated. Following massive pre-training, models undergo fine-tuning using human-labeled datasets and feedback (such as RLHF). Human annotators curate reference data, grade output quality, and align training objectives to make model responses accurate, helpful, and safe.
Nevertheless, neural networks have strict reliability boundaries. They are not exact biological brain replicas, but mathematical models inspired by neural abstractions. Their efficacy remains bound to training data fidelity, architecture design, loss functions, and compute resources. While achieving high accuracy across many domains, they remain prone to hallucination, fallacious reasoning, or inaccurate outputs when navigating edge cases outside their training distribution.
Scaling parameter counts generally improves representation capacity for complex tasks. However, parameters do not contribute uniformly. As parameter size scales, training, storage, and inference costs scale non-linearly. Balancing model capacity, data quality, compute efficiency, and response reliability remains a central engineering challenge in modern AI.
History of AI Development
AI was officially born in 1956 at the Dartmouth Conference, and computer scientist John McCarthy coined the term "artificial intelligence." Since then, AI has progressed through distinct eras, including Expert Systems in the 1980s, evolving into a core technological pillar of today's Industry 4.0 revolution.
AI is reshaping how we work, study, entertain, and interact with the world, introducing new paradigms while impacting employment, investment strategies, and fundamental societal workflows.
2. Everyday AI Products and Services
- Virtual Assistants: Voice assistants like Apple's Siri, Google Assistant, and Amazon Alexa process natural voice commands to set alarms, query information, and control smart home ecosystems.
- Mapping and Navigation: Google Maps and Waze evaluate real-time traffic telemetry to compute optimal routing, estimate arrival times, and broadcast traffic congestion alerts.
- Content Recommendation: Platforms like Netflix, YouTube, and Spotify analyze consumption history to recommend tailored media. E-commerce platforms like Amazon similarly personalize product recommendations based on purchasing patterns.
- Spam Filtering and Smart Replies: Email clients employ AI models to detect spam/phishing attempts and suggest context-aware quick responses.
- Grammar and Spell Checkers: Tools like Grammarly or native smartphone autocorrect engines rely on AI to detect and fix syntactic and grammatical errors.
- Facial Recognition: Applied for device authentication, identity verification in security infrastructure, and automatic tag recommendations in media management.
- Smart Home Infrastructure: Smart security cameras and adaptive thermostats utilize AI to automate environmental functions based on user habits.
- Autonomous Driving: Self-driving systems utilize computer vision and sensor fusion AI to analyze real-time environments and make vehicle control choices.
3. Enterprise and Industrial AI Applications
- Customer Experience:
- Chatbots and Virtual Agents: Handle routine inquiries, order management, and 24/7 support, reducing service agent workload while accelerating response times.
- Sentiment Analysis: Evaluates customer feedback text and voice interactions to measure satisfaction and brand sentiment.
- Banking and Finance:
- Fraud Detection: Analyzes transaction streams in real time to detect anomalous patterns indicative of credit card fraud or money laundering.
- Risk Management: Assesses creditworthiness and models market trends to support investment risk evaluations.
- Robo-Advisory: Generates automated portfolio advice matched to individual risk profiles and financial goals.
- Healthcare:
- Diagnostics: Evaluates medical imagery (X-rays, MRIs, CT scans) and clinical records to assist clinicians in detecting early-stage pathologies (e.g., oncology, cardiology).
- Drug Discovery: Accelerates pharmaceutical R&D by analyzing biological molecular datasets.
- Patient Monitoring: Wearable IoT devices leverage AI to track patient vitals remotely and signal anomalies.
- Manufacturing and Supply Chain:
- Predictive Maintenance: Monitors equipment sensor telemetry to forecast maintenance schedules, minimizing unplanned downtime.
- Supply Chain Optimization: Forecasts inventory demand and optimizes logistics routes for operational efficiency.
- Quality Control: Inspects production line output via computer vision to identify defects automatically.
- Logistics and Transportation:
- Route Optimization: Factors in traffic conditions and weather telemetry to determine optimal fleet routing.
- Autonomous Logistics: Powers autonomous freight trucks and delivery drones.
- Education:
- Personalized Learning: Adapts educational content and pacing dynamically to student performance data.
- Intelligent Tutoring Systems: Offers tailored explanations and problem-solving hints.
- Automated Grading: Streamlines evaluation for standardized tests and coursework.
- Marketing and Advertising:
- Campaign Personalization: Segments consumer data to deliver targeted marketing campaigns.
- Content Generation: Assists in drafting promotional copy, visual imagery, and marketing video assets.
4. Military and Defense AI Applications
Weaponized AI applications represent severe societal risks; deploying autonomous military AI robotics during armed conflicts presents terrifying potential consequences. However, these applications remain constrained by training iteration cycles, real-time edge processing limits, and vulnerability to adversarial exploitation once operational heuristics are identified.
Comments: Not included in Gemini AI's initial response.
5. Do AI applications primarily serve as decision support, or can they be deployed in domains requiring 100% precision?
Many AI applications focus on decision support. AI processes data, identifies patterns, and generates forecasts or recommendations to assist human decision-makers.
However, "precision" must be contextualized broadly. Beyond advisory roles, AI is actively integrated into high-precision, highly automated workflows, sometimes executing decisions autonomously without human intervention.
AI in High-Precision Domains
AI achieves high operational precision across domains processing massive datasets and complex visual patterns:
- Healthcare Diagnostics:
- Medical Image Analysis: AI scans X-rays, MRIs, and CT scans to highlight minute lesions with high precision, assisting in early oncology detection. Clinical studies indicate AI models can attain up to 94% accuracy in tumor detection, matching or exceeding radiologist baseline performance in specific tasks.
- Predictive Prognostics: Evaluates electronic health records and genomic data to forecast disease progression.
- Robotic Surgery Support: AI-assisted surgical robotics perform micro-precision maneuvers, reducing human tremor errors.
Comments: Nevertheless, clinical AI outputs still require validation by qualified medical specialists before intervention. Relying uncritically on AI in medicine is suboptimal; the human body is a resilient biological system capable of buffering minor dosage variations, but unverified AI prescriptions risk delivering zero-value or unsafe outcomes.
- Finance and Banking:
- Fraud Detection: Requires high operational precision. Models analyze millions of real-time financial transactions to flag fraudulent activity, reducing monetary loss.
- Risk Management: Calculates loan default risk and analyzes market exposure.
- Industrial Manufacturing:
- Quality Assurance: Computer vision models inspect manufacturing lines at high speed to spot micro-defects invisible to the naked eye.
- Predictive Maintenance: Monitors machine telemetry to forecast failure timelines, preventing costly operational shutdowns.
- Transportation:
- Autonomous Vehicles: A critical example demanding extreme accuracy. Autonomous vehicle AI must identify surrounding objects (pedestrians, vehicles, signage), predict trajectories, and make split-second control decisions safely.
- Logistics Routing: Computes delivery routes based on real-time traffic and weather data.
- Security and Surveillance:
- Biometric Recognition: Deployed in access control systems for high-accuracy identity verification.
Autonomous AI Decision-Making
AI applications extend beyond decision support into full task automation and autonomous execution, particularly in controlled environments or where latency is critical:
- Factory Automation Robotics: Performs material handling and assembly tasks without human intervention.
- High-Frequency Trading (HFT): Algorithms execute thousands of market trades per second based on programmatic strategies, operating entirely without real-time human intervention.
- Smart Grid Management: AI dynamically balances electrical distribution grid load to optimize efficiency and mitigate power outages.
Comments: Power grid management directly impacts public welfare, meaning AI should strictly serve early-warning and recommendation functions rather than full autonomous control. Similarly, industrial factory robots do not necessarily require AI: conventional robotics rely on deterministic, hardcoded motion scripts (repeating set spatial trajectories like left, right, up, down, spin). For dynamic obstacles, deterministic sensor-triggered feedback loops often suffice without non-deterministic AI.
6. Factory automation robots aren't purely AI; isn't AI's actual share in robotics relatively small?
That is an accurate observation: factory automation robotics are not always AI-driven, and in many instances, AI components represent only a small portion of the overall system architecture.
To clarify, we must distinguish between Robotics and Artificial Intelligence:
- Robotics: The engineering domain covering the mechanical design, fabrication, actuation, and physical control of robots. A robot is a physical hardware mechanism that performs work via mechanical linkages and control systems.
- Artificial Intelligence: The computational discipline focused on developing software algorithms capable of learning, reasoning, pattern recognition, and natural language understanding. AI exists purely as software logic.
Comments: Mission-critical physical operations requiring 100% deterministic accuracy should avoid non-deterministic AI modules. For instance, a washing machine executes hardcoded state-machine commands (power on, fill, spin, drain). Introducing AI into such core mechanical functions adds unnecessary non-determinism; AI in appliances is often limited to user interaction gimmicks rather than core mechanical control. AI should be integrated into this field as a warning and error-detection system to assist manufacturers in investigating and identifying root causes for improvement.
The Distinction and Relationship Between Robotics and AI
Not all robots utilize AI, and not all AI requires a physical robotic hardware body.
- Traditional Industrial Robots: Most factory robotic arms (such as welding or pick-and-place arms) execute fixed, deterministic instruction loops with high repeatability, but possess no perception, learning, or reasoning capabilities. They perform exactly what they were pre-programmed to do within static, predictable environments. They represent deterministic automation rather than AI.
- AI-Integrated Robotics: Integrating AI endows robotic hardware with adaptive capabilities:
- Perception and Environmental Awareness: Computer Vision allows robots to identify arbitrary objects, inspect visual defects, or determine spatial coordinates dynamically.
- Reinforcement and Supervised Learning: ML enables robots to refine trajectory planning over time for speed or energy efficiency.
- Environmental Adaptability: Enables robots to adjust trajectories dynamically when environmental conditions shift (e.g., misaligned components or unexpected obstacles).
- Autonomous Decision-Making: Autonomous Mobile Robots (AMRs) in logistics warehouses use AI SLAM algorithms to map facility floors, dynamically route around obstacles, and transport payloads without physical guide rails.
Proportion of AI in Industrial Robotics
The share of AI in legacy industrial robotics remains modest. However, integration is accelerating:
- Legacy Robotics: Industrial robots installed in past decades rely on rigid PLC deterministic logic with minimal or zero AI integration.
- Modern and Next-Gen Robotics:
- AI Convergence: Chipmakers and industrial automation firms are building dedicated edge-AI hardware and software toolchains to embed computer vision and ML models into robot controllers.
- Industry Trends: The International Federation of Robotics (IFR) lists Artificial Intelligence among top robotics trends, reflecting its growing role.
- Market Adoption: Machine Learning and Computer Vision represent growing segments in robotics market share, driven by demand for adaptive material handling and automated quality inspection.
- Cobots (Collaborative Robots): Cobots operate alongside human personnel; AI assists cobots in sensing human proximity and modulating motion trajectories for workplace safety.
While not all industrial robots incorporate AI today, embedding adaptive AI models into robotic systems is an active industry trend, making automation hardware more flexible, adaptive, and efficient.
Comments: It is crucial to recognize that industrial robotics prioritize strict spatial accuracy and repeatability. Therefore, AI acts as a supplemental plugin module to enhance sensory perception rather than an "AI-first" replacement for core deterministic motion controllers.
7. If AI is viewed as a feature/function, is the industry trend moving toward building AI-native coding frameworks, or integrating AI as plugin|layer tools?
Viewing AI as an augmented system feature is an accurate architectural perspective. The integration of AI into software development frameworks is progressing along two distinct vectors:
1. AI as "Plugins" or "Developer Assist Tools" (AI-augmented Development)
This is the most widespread paradigm today, where AI serves as an external assistance layer embedded directly into existing IDEs and software toolchains.
- Characteristics: AI functions as a developer copilot, automating boilerplate code generation, suggesting completions, and catching syntax errors. It does not alter the underlying architectural design of programming languages or application frameworks.
- Examples:
- AI Code Completion/Generation: Extensions like GitHub Copilot, Amazon Q Developer, Tabnine, and native IDE integrations (VS Code, JetBrains IntelliJ IDEA) leverage LLMs to:
- Provide context-aware inline code completion.
- Synthesize functions and classes from natural language prompts.
- Refactor existing codeblocks for efficiency.
- Generate unit tests automatically.
- Produce inline documentation and docstrings.
- AI Debugging/Static Analysis: Scans codebases to catch vulnerability patterns, memory leaks, or performance bottlenecks.
- AI-Powered Low-Code/No-Code Platforms: Translates natural language directives into visual workflows, schema definitions, or backend scaffolding.
- Pros: Frictionless adoption, zero requirement to redesign legacy architectures, and immediate developer productivity gains.
- Cons: AI remains strictly auxiliary; structural software engineering and system design decisions remain manually directed by human engineers.
2. AI-Native Frameworks (AI-First Architecture)
An emerging paradigm where entire software frameworks are architected around LLMs and AI Agents as foundational primitives. Instead of manually coding deterministic control flows, developers define high-level intents, constraints, and execution goals.
- Characteristics: Framework primitives center on model orchestration. Developers specify agent roles, tool interfaces, and goal states, leaving the framework's AI engine to resolve plan generation, API invocations, and state management dynamically.
- Examples:
- AI Agent Frameworks: Frameworks such as LangChain, LlamaIndex, CrewAI, AutoGPT, Semantic Kernel, and AutoGen facilitate multi-agent orchestration, equipping models with memory, planning heuristics, and external tool execution APIs:
- LLM Context & Chain Management: Sequences multi-step model calls and handles schema transformations.
- Agent Memory Management: Maintains short-term conversational context alongside long-term vector database embeddings.
- Tool/API Invocation: Grants agents autonomy to execute web searches, query SQL databases, or execute shell commands.
- Autonomous Multi-Agent Collaboration: Coordinates specialized agents (e.g., architect, coder, tester agents) working in concert toward an engineering milestone.
- AI-Native Cloud Platforms: Cloud orchestrators designed to auto-scale, deploy, and optimize microservices based on AI telemetric analysis.
- Generative AI Application Frameworks: Domain-specific SDKs focused on building RAG pipelines, multi-modal generation tools, and streaming AI endpoints.
- Pros: Enables high operational automation, streamlines complex agentic workflows, and simplifies building software centered on non-deterministic model logic.
- Cons: Steeper learning curve, requires a shift in software engineering design patterns, and can introduce complex debugging and state management challenges compared to traditional deterministic frameworks.
Conclusion on Industry Trends
Both paradigms will continue to coexist and develop in parallel:
- AI as Plugins/Tools: Will remain the dominant approach for standard software development, enhancing developer throughput across established stacks without breaking existing codebase conventions.
- AI-Native Frameworks: Will capture market share in building agentic software, complex RAG applications, and domains where non-deterministic reasoning forms the core business logic of the application.
8. Distinguishing Between Products with Genuine AI and Those Merely "Labeled as AI"
As AI becomes increasingly popular, many companies are branding their products as AI-powered to attract customers. This practice is known as AI washing. Gemini suggests the following criteria for evaluating AI products:
| Criteria |
Genuine AI |
"AI-Washed" Products |
| 1. Learning and Adaptation |
Improves its results as more data becomes available and learns from errors or past experience. For example, a recommendation system becomes more accurate as you continue using it. |
Relies on fixed rules and predefined scenarios, such as "if A, then B." It cannot improve on its own or adapt to new situations. |
| 2. Data |
Clearly explains how data is collected, processed, and used to train its models. |
Provides vague information about data sources and models. Claims that it can handle complex tasks without requiring any data are a warning sign. |
| 3. Handling Complex and Ambiguous Situations |
Can analyze situations involving multiple variables and incomplete information. For example, it can analyze multiple symptoms and medical records to assess possible health conditions. |
Can only solve simple problems in controlled environments. When faced with situations outside its predefined scenarios, it may fail or produce meaningless results. |
| 4. Case Studies and Real-World Results |
Provides concrete implementation case studies and measurable performance indicators, such as improvements in productivity, cost reductions, or increases in accuracy. |
Offers only generic claims, lacks supporting data, fails to explain how AI contributes to the results, and avoids detailed technical questions. |
| 5. Technology and Architecture |
Is willing to disclose the technologies used, such as Deep Learning, Natural Language Processing (NLP), Computer Vision, and Reinforcement Learning, as well as the models involved (e.g., LLMs such as GPT-4 or Llama) and how they are integrated. |
Uses vague terms such as "intelligent algorithms" or "proprietary technology" without providing concrete technical explanations. |
| 6. Customization and Scalability |
Can be customized to meet specific customer needs and industry requirements, and can scale as data volumes and business requirements grow. |
Offers a closed solution that is difficult to customize and becomes less effective as data volumes or operating contexts change. |
| 7. Demo and Trial |
Allows customers to test the product using their own data and real-world scenarios while maintaining satisfactory results. |
Relies on carefully staged demos that showcase only ideal scenarios. When customers test the product themselves, the results often fall short of expectations. |
Note
Chatbots typically do not learn continuously in real time during conversations. Improvements generally occur through subsequent training, fine-tuning, or system updates. Therefore, the "learning" criterion should be understood as the system's ability to improve over time, rather than its ability to learn from every individual conversation.
9. Overall Assessment
9.1. Terminology
The concept of AI has become widely popular thanks to media coverage, so asking a chatbot "What is AI?" is an easy way to quickly grasp the basics. However, some technical terms differ significantly from their everyday meanings. For example, "inference" in AI refers to the process by which a trained model uses input data to generate predictions or responses. In essence, it is a computational process with inherent limitations. If the results are incorrect, improvements typically require further training or fine-tuning in a subsequent version, or additional context and reference materials supplied to the model. By contrast, human reasoning involves thinking, analyzing, and drawing conclusions based on knowledge, experience, and context. Humans can also adjust their reasoning immediately when they recognize that something is right or wrong. Although the two concepts share the same name, they are fundamentally different and should not be treated as equivalent.
The term "artificial neural network" is also heavily influenced by marketing: while it draws inspiration from the workings of biological nervous systems, it is, in practice, a mathematical model designed for information processing—a model far simpler than the biological brain and nervous system. Consequently, it cannot be compared to the sophistication of the human brain, which is the product of billions of years of natural evolution.
9.2. Where Is AI Useful?
Media coverage often portrays AI as an extraordinarily powerful technology. However, after exploring the subject through chatbots, I have come to believe that AI is primarily useful for problems involving large datasets, problems without a single definitive answer (such as forecasting and even fortune-telling), and tasks such as programming and learning. AI should not be overused or applied inappropriately, especially when doing so merely complicates matters unnecessarily.
- As of the time this article was written (2025), one of AI's most successful applications is AI chatbots. Students, office workers, businesses, and individuals are using them as highly effective assistants for their work. In uncertain situations involving chance or ambiguity, AI can also help improve decision-making. In other words, AI is best applied to problems where improvements are possible, even when finding the absolute best solution or a definitive answer is impossible. This is essentially an application of optimization theory.
- The growing popularity of asking questions and receiving answers through AI chatbots may be contributing to a decline in the volume of searches performed through Google Search.
- Training models is expensive, and correcting mistakes can require considerable time and effort. Therefore, AI products often combine model training with the reuse of existing components, such as data retrieval modules. They may also incorporate additional techniques throughout the AI product development process without having to retrain the model from scratch.
9.3. AI on Standalone Devices (Robots and UAVs)
For robots or UAVs to operate autonomously using AI, they may require substantial memory for real-time processing and considerable storage capacity, making practical deployment challenging. Currently, on-device AI is primarily used for inference rather than learning. It operates based on predefined datasets and instructions, and after several interactions with the environment, its working memory may be cleared instead of retaining information for long-term analysis and dataset expansion.
When a robot operates independently of a server, with its own CPU and an integrated AI system capable of learning, there is a risk of incorrect training leading to undesirable behaviors, such as violence, abusive language, or bizarre beliefs. AI behavior is shaped by its input data, so deliberately manipulated or misleading training data can produce corresponding behaviors.
Uncontrolled AI systems may also violate privacy and copyright laws when collecting data. The current environment is relatively permissive, but these practices may eventually face stricter legal restrictions, potentially resulting in certain AI hardware devices being banned from the market.
For non-agentic AI applications, companies providing AI modules need substantial hardware and software infrastructure to develop their products. However, when these products are integrated into applications, they are simply plug-in modules or components that provide AI capabilities. Therefore, AI should be used to solve software and hardware problems that conventional methods cannot adequately address. These are primarily problems involving datasets too large to exhaustively search, or problems for which exhaustive search is possible but produces response times that are too slow for practical deployment.
9.4. Is It Necessary to Learn ML?
II haven’t studied, and have no motivation to explore, machine learning (ML), which is often considered the brain of AI. In my opinion, developers building applications that integrate with AI to any degree don’t necessarily need to learn ML. Understanding the basic concepts covered above is enough to get started. You can also look up some important terms using Google AI Mode, such as: large language model (LLM), token, prompt, context window, temperature, hallucination, fine-tuning, RAG (Retrieval-Augmented Generation), and embedding.
After completing an AI product, data scientists provide API functions that other applications can call when needed. To the end user, AI is a “black box”: as long as the output works properly, there’s no need to understand what’s inside. Typically, the backend application and the API of an AI-first provider sit within the same internal network of a cloud provider, so the latency of an AI API call is very low.
A common example today is individuals and small and medium-sized enterprises (SMEs) integrating ChatGPT into their applications or websites through the OpenAI API. This allows customers to ask questions and get answers from AI right on the business’s website, without having to visit ChatGPT separately.
If you want to go deeper into AI and have your own “mini ChatGPT,” you can download DeepSeek-R1 to your machine, install it, try it out, and explore. It’s a great way to understand how large language models work without investing heavily in ML theory right from the start.
9.5. Will AI Become More Intelligent Than Humans and Escape Human Control?
Some people worry that AI will become more intelligent than humans and eventually become uncontrollable. In my opinion, AI may well become more intelligent in certain respects, but whether it will become uncontrollable is a separate question. Chatbot-based AI and applications running on servers are essentially large software systems and tools for increasing productivity. At present, only a handful of major companies around the world operate at the forefront of this field.
By contrast, physical AI systems that operate independently of servers, can "learn," and take the form of robots are constrained by memory capacity and computational power. They can potentially freeze or crash when memory is exhausted. Most of the computational power is concentrated in servers, so cramming every AI capability into a standalone robot would make it slower, more expensive, and harder to sell on a mass-market scale.
Ultimately, whether a product contains zero AI, some AI, or is almost entirely AI-powered, its commercial success depends on how well the product or service meets customer requirements and satisfies customer needs.
9.6. Can AI Coding Agents Replace Software Developers?
It is true that AI agents can generate a rough, preliminary codebase, but they still have significant limitations:
- Designing complex systems and making difficult architectural decisions.
- Understanding ambiguous or frequently changing requirements and accommodating the needs of multiple stakeholders.
- Lacking deep knowledge of specific business domains and struggling to evaluate trade-offs between alternative solutions.
- Finding original and innovative solutions, developing new algorithms, and handling rare edge cases.
- Source code security: Individual customers and startups may be willing to hire AI coding agent services, but large enterprises tend to prefer internally deployed AI agents. Even internal solutions can quickly become outdated as requirements change.
- When requirements change or new features must be added to an existing codebase, agents cannot always update the code correctly.
- Over the next few years, even as AI agents improve, generated code will still contain defects and require knowledgeable developers to review and fix it. Completely entrusting software development to AI agents is little different from playing the lottery.
- As an application evolves over time, especially in projects involving multiple developers, an AI-assisted low-code approach may reveal significant limitations in performance and maintainability. A trial-and-error workflow, in which developers repeatedly prompt AI to fix errors and then wait for the results, not only slows development but also makes the codebase increasingly difficult to control. Without sound architecture and a disciplined development process, the application may become a patchwork of poorly integrated components that is difficult to extend and ultimately has to be discarded and rebuilt from scratch, particularly if its runtime performance becomes painfully slow.
9.7. Topics for Further Investigation
- The relationship between AI, malware, and computer viruses.
- AI-generated images and the proliferation of low-quality, misleading, or meaningless content.
- Uncontrolled AI behavior caused by manipulated input data or a mismatch between real-world conditions and expectations.
- The impact of AI on employment, occupational structures, and future skill requirements.
9.8. Conclusion
Despite the mixture of challenges and benefits discussed above, AI has reached a relatively mature stage and is already being used effectively to increase productivity across many industries, particularly software development. However, achieving a revolutionary breakthrough that fundamentally transforms how people work, learn, create, and organize society may still be a long way off. No one can yet say with certainty which direction AI will take, or whether a superior new technology will emerge in the future to complement or replace it.
Appendix: GPUs, CUDA, and the Road to AI
This section explains why the GPU became the hardware foundation of modern AI.
The GPU
Definition: A GPU (Graphics Processing Unit) is a massively parallel processor made up of thousands of small cores that perform the same operation on many pieces of data at once (the “single instruction, multiple data” model), paired with high-bandwidth memory (VRAM).
History: Dedicated graphics chips existed as early as the 1970s and 80s (arcade machines, graphics workstations), but they were not called GPUs. In August 1999, NVIDIA launched the GeForce 256 and marketed it as “the world’s first GPU.” It moved geometry transformation and lighting (T&L) into hardware instead of leaving them to the CPU. The term “GPU” became widespread from then on, although the label was NVIDIA’s marketing choice.
Role: GPUs were originally built to render graphics. Today they serve both graphics and general-purpose parallel computing (AI, science, video, and more).
Comparison with CPUs:
| CPU | GPU |
|---|
| Cores | Few, powerful | Very many, simple |
| Strength | Sequential tasks, complex logic | The same operation on huge volumes of data |
CUDA
Definition: CUDA (Compute Unified Device Architecture) is a platform that lets developers program the GPU as a general-purpose computer.
Timeline: CUDA was announced in late 2006 alongside the G80 chip, and version 1.0 was officially released in June 2007. That release can be considered CUDA’s birth.
Role: Before CUDA, scientists who wanted to compute on a GPU had to disguise their work as graphics commands, which was very difficult. CUDA bridged that gap by allowing GPUs to be programmed in ordinary C/C++. Frameworks such as PyTorch and TensorFlow later placed Python on top of CUDA, so today’s AI practitioners do not need low-level GPU programming skills.
From the GPU to AI
Early period (2000-2007): GPUs and CUDA were not aimed at AI. Their main users were computational scientists and simulation engineers who needed to run large volumes of parallel calculations.
Neural networks, 2000-2010: Neural networks had existed for decades but received little attention and were not yet a mainstream direction in the field.
Discovering the fit (around 2008-2009): A few researchers, such as Andrew Ng’s group at Stanford and Geoffrey Hinton’s at Toronto, began experimenting with GPUs for neural networks. They realized on their own that GPUs suited a neural network’s need for billions of independent matrix multiplications.
Proof (2012): AlexNet, trained on GPUs, won the ImageNet image recognition competition. The result demonstrated the practical payoff and opened the large market for deep learning and today’s AI.
Hardware built for AI (from 2017): NVIDIA added Tensor Cores, dedicated hardware units for matrix multiplication.
Key milestones
| Year | Event |
|---|
| 1999 | The GPU is born for graphics (GeForce 256) |
| 2007 | CUDA 1.0: the GPU becomes a programmable computer, serving scientific computing |
| 2008 - 2009 | AI researchers begin trying GPUs for neural networks |
| 2012 | AlexNet proves the approach, launching the deep learning era |
| 2017 | Tensor Cores: hardware optimized specifically for AI |
