1. AI can make users more passive or overly dependent on it. What approaches can help individuals maintain critical thinking and self-reliance in the age of AI?
2. Generative AI often follows strong logical patterns, which can sometimes lead our thinking in narrow or unintended directions. How can we prevent AI from shaping our perspectives too rigidly?
Answer
AI is merely a tool for acceleration, similar to how a calculator speeds up computation without eliminating human mathematical reasoning. To maintain self-reliance, we must shift our role from hands-on implementers to Designers and Validators of the output. By doing so, core critical thinking shifts and focuses on high-level factors like product vision, usability, and cost analysis.
How can organizations manage data for AI while staying compliant with GDPR and ISO 27001?
Answer
To ensure security and compliance (GDPR/ISO 27001), start by minimizing and anonymizing sensitive data before it is used for AI training. Next, enforce strict Access Controls and encrypt data both at rest and in transit. Finally, establish detailed auditing procedures on how data is used, logging every change and access to demonstrate continuous compliance.
To prevent AI from narrowing our thoughts, we must adopt a Critical Inquiry Strategy instead of merely accepting the output. Always validate the results by asking "Why is the answer this way?" and "Is there an alternative approach?". Most importantly, force the AI to present the contrarian view: What are the pros and cons of doing the opposite? to broaden the framework of thinking.
For businesses starting their AI Agent journey, what is the minimum checklist (team, data, processes, and infrastructure) required?
Answer
The AI Agent journey must begin by clearly defining the specific business problem and need the Agent will solve, ensuring real-world value. The next step is to redesign and organize proprietary internal data so it can be retrieved and used as context for the Agent (RAG). The company must recruit an initial core team focused on MLOps and Prompt Engineering. For infrastructure, prepare a flexible, cloud-agnostic computing environment that can scale quickly. Finally, establish a robust product development (SDLC) and maintenance process to manage the Agent’s lifecycle and evolution.
Context Engineering
Answer
Context Engineering focus on designing and managing the context data that AI models (like LLMs) need to provide accurate and relevant responses. This involves integrating real-time data and proprietary business knowledge into the Prompt using techniques like RAG (Retrieval-Augmented Generation). To transition, Full Stack Engineers should focus on mastering data retrieval systems, optimizing vector databases, and building high-performance data pipelines to serve context to the LLMs.
I am a member of I2FLabs Vietnam, a research group exploring practical AI applications.
We would like to ask: Where do you see Agentic AI in Vietnam over the next 5–10 years as government support and corporate investment drive its democratization across core sectors?
Additionally, how do you view the Internet of Agents (IoA) trend—where AI agents interact and exchange data? If Vietnam develops this model, what are the key opportunities and challenges for digital transformation, specifically in addressing traffic management and urbanization?
Answer
Over the next 5–10 years, Agentic AI will transition from the experimental stage to democratization and deep integration into core business processes in Vietnam. The AI Agent will become a standardized "logic layer", automating repetitive and complex tasks not only for large corporations but also for small and medium-sized enterprises (SMEs). Support from the State and businesses will focus on creating a flexible legal framework and investing in national data infrastructure to ensure Agents operate effectively and reliably.
The concept of the Internet of Agents (IoA), or Agent Mesh, unlocks enormous potential through a combinatorial effect. Each Agent will tackle a specific application, such as traffic management or climate forecasting, but when connected, they form a ""system of systems"" capable of solving highly complex problems. This model is ripe for areas like urban digital transformation, helping Vietnam move toward a "Cognitive Cities" model.
However, the greatest challenge for IoA lies in the complexity of the "mess of applications" within the network. Managing, maintaining, and ensuring the consistency/reliability across a vast network of Agents is extremely difficult from both a technical and organizational standpoint. This demands significant effort from engineers and end-users to clearly define interaction protocols, authorities, and the specific responsibilities of each Agent within the mesh.
How can we ensure the reliability of AI outputs in a production environment?
Answer
Watch the Data Always: We must constantly check if the data the model sees in production is the same as the data it was trained on. If the data changes (drift), the model breaks.
Explain the Choices: The model can't be a “black box”. We need the AI to give us a clear reason for every important decision it makes, so we can check it and trust it.
Test Automatically and Often: We need automated tests to make sure the model is fair, stable, and has no hidden errors before we let it run live.
Log Everything: Every AI prediction must be fully recorded (logged), including which model version produced the result, so we can easily track problems.
Design for Errors: The model must be strong enough to handle the messy, bad data it will definitely see in the real world, not just the clean data from the lab.
What do you think will be the next era after agentic AI, and can you share more about migration compared to rebuilding from scratch?"
Answer
Following the era of Agentic AI operating in digital environments, the next major shift will be to Physical AI, which focuses on embedding intelligence into robots and systems interacting with the real world, as championed by Elon Musk (Tesla) and NVIDIA. The era after that will likely be Symbiotic AI, where humans and AI cooperate deeply. Migrating current Agentic models to Physical AI is challenging because digital models do not inherently understand force and physics. Therefore, instead of migration, we primarily need to rebuild new Vision-Language-Action (VLA) models designed specifically for physical environments.
Where should developers start to qualify for AI roles or apply AI in business?
What does the learning roadmap look like? Please share reference roadmaps.
Answer
For any developer pivoting to AI, here’s your three-point checklist:
Code & Data Foundation: Master Python and the core data stack—that means Pandas and NumPy. You must handle data effectively before you can model it.
Practice is Priority: Get the ML theory, but prioritize end-to-end projects and competitive practice like Kaggle. Theory is static; execution is where you learn.
Deploy for Value: Leverage your Full Stack background for MLOps. This enables smooth production deployment. Crucially, always tie model accuracy to Business value—that is the final metric of success.
AI adoption in QA
Answer
AI adoption in QA means integrating AI/LLM technologies into the testing workflow to automate and enhance quality assurance activities.
This includes tasks such as:
Example:
In an AI Agent Platform, QA can use an AI model to generate test cases for each skill, validate rulebook search accuracy using semantic metrics, or simulate user queries to check how the agent triggers skills and returns responses.
What do you think about vibe coding?
Answer
Vibe coding is a new approach that helps accelerate the software development process. It allows developers to focus on ideas instead of technical details, and it’s especially useful for non-technical people, enabling them to quickly turn ideas into prototypes. Although vibe coding still has many limitations and imperfections today, I believe it will continue to evolve, become more mature, and eventually become an essential skill in how we build products.
Could the speaker share experiences and principles for implementing AI in software development while maintaining high security?
Answer
Deploying AI in software development can deliver significant productivity gains, but if it is not designed correctly from the beginning, AI can also become a new security risk. Therefore, the core principle is clear: AI must adhere to the same—if not stricter—security standards as traditional software systems.
Based on practical implementation experience, several key principles stand out:
First, apply Zero Trust to AI and data. AI, LLMs, and all input data should never be trusted by default. Every prompt, context, and output must be governed, with clearly defined access controls and usage boundaries.
Second, isolate AI from core systems. AI should operate as a sandboxed service without direct access to source code repositories, secrets, or production systems. All interactions must go through well-defined APIs and enforced security policies.
Third, strictly control the data and code provided to AI. Entire repositories, logs, or internal documentation should never be fed into LLMs. Only the minimum required context should be supplied, with sensitive information masked or removed before use.
Fourth, maintain human-in-the-loop for critical tasks. AI can assist with code generation, review, or analysis, but final decisions—especially those related to security—must remain the responsibility of humans.
Finally, observability and auditability are mandatory. It is essential to track what data AI accesses, what outputs it generates, and how those outputs impact the system, enabling effective auditing, traceability, and incident response.
In summary: AI can dramatically accelerate software development, but it is only safe when implemented with a security-first mindset, strict control over data and access, and clear human accountability at the end.
Multi-source data integration is key to smarter AI. How can we maintain data privacy and integrity when inputs come from various sources? Additionally, how do we ensure external data is clean and free from malware or vulnerabilities that could compromise the system?
Answer
Integrating multi-source data is essential for building intelligent AI systems, but it also introduces serious risks around privacy, integrity, and security.
To address this, AI platforms must treat all incoming data as untrusted by default and apply a Zero Trust approach to data.
Key practices include separating data ingestion from consumption, enforcing validation and sanitization pipelines, and classifying data early to control access and privacy exposure. Only verified and cleaned data should be made available to AI agents and LLMs.
Additionally, external data must be quarantined, scanned for malicious content, and governed by policy engines before use. End-to-end observability and auditability are critical to trace how data flows into prompts, impacts model behavior, and affects cost and system reliability.
In short, AI becomes truly trustworthy only when data is not just abundant, but controlled, validated, and observable.
