[카테고리:] AI & Technology

English articles about AI agents, LLMs, automation, developer tools, and technology trends.

What Is the OpenHuman AI Agent? The Promise and Limits of an Open-Source AI Assistant with Local Memory
OpenHuman AI agent is attracting attention because it is not simply “one more chatbot.” It aims at something larger than a single conversation window. It tries to bring a person’s files, notes, apps, web research and automation flows together into one AI work environment.

Its official GitHub repository says OpenHuman aims to be “a brain that builds local-first memory.” It also describes itself as an orchestrator running a fleet of agents and as a deep researcher. The phrasing is ambitious, but the direction is clear. It wants AI to stop asking from scratch every time. Instead, it imagines a work system that continuously remembers and acts within the user’s world.

What Kind of Project Is OpenHuman?

OpenHuman is an open-source AI agent platform published in the `tinyhumansai/openhuman` repository. According to the official repository, its primary language is Rust and its license is GPL-3.0. As of July 2026, the latest release is v0.61.8, and the README explicitly describes the project as Early Beta.

There are three core ideas.
1. A structure that remembers personal data with a local-first approach
2. An orchestration structure for running multiple agents and workflows
3. An execution environment that combines web, files, browser, voice and research tools
So OpenHuman is less like a standalone model such as ChatGPT or Claude. It is closer to an “agent harness” that connects multiple models and tools to real workflows.

First Difference: Local-First Memory

The point OpenHuman emphasizes most strongly is memory. The README mentions Memory Tree together with Obsidian Wiki. It describes a system that organizes user data into SQLite and a Markdown tree. It then connects that data to an Obsidian vault that humans can open and inspect.

This approach matters. Many AI tools say, “I remember.” But users often cannot tell where that memory is stored, how it is retrieved, or how to fix incorrect memories. At least in its direction, OpenHuman aims for an editable knowledge structure rather than black-box memory.

Of course, the actual quality still has to be verified separately. Even if a memory structure looks good, automatically collected information can become a burden if it accumulates as duplicate or outdated content. For a personal knowledge system, “organizing well and making corrections easy” matters more than “saving a lot.”

Second Difference: Agent Orchestration

OpenHuman emphasizes agent orchestration more than a single assistant. It aims to create workflows, run them based on triggers and build automation with approval gates. The README also mentions related open-source subcomponents such as `tinyflows` and `tinyagents`.

This direction fits well with the current AI agent trend. Work automation no longer ends with “how to write good prompts.” It is now more important to decide what should run when. Teams also need retry rules and clear points where user approval is required.

That is also why OpenHuman is interesting. If a personal AI assistant is to enter real work, it needs an execution structure more than conversational ability. Calendar checks, meeting summaries, file investigation, research, code work and report drafting need to connect into a single flow.

Third Difference: Combining Research and Execution Tools

OpenHuman introduces features such as web search, scraping, browser use, coding tools, voice and meeting agents. Its official documentation also uses the concept of SuperContext. The direction is to scan local memory, files and the web first. That secures context before the user has even finished asking.

If this structure works properly, the advantage is significant. You do not have to explain again from the beginning what the AI knows. It can answer after reviewing work context, file locations, recent conversations and related documents.

However, permission design is the core issue for this kind of tool-based agent. Connecting accounts such as Gmail, Slack, Notion and GitHub is convenient, but it also expands access to sensitive information. If the tool has automatic execution features, even more caution is needed. You need to check what data remains local and what requests are sent to external model APIs.

Who Might Want to Try OpenHuman?

OpenHuman is not a tool everyone needs. It becomes more worth considering if you are close to the following conditions.
- You already use Obsidian, Markdown or local-file-based knowledge management.
- You want to connect an AI assistant to work automation rather than simple Q&A.
- You want to handle context from multiple tools such as Gmail, Notion, GitHub and Slack in one place.
- You are interested in extensibility structures such as AI agent workflows, MCP and Skills.
- You can tolerate the instability of a beta tool and test it.
On the other hand, you may expect a work tool that you can install and use reliably right away. In that case, it is better to wait a little. The official repository also states that it is Early Beta. The more data access a tool has, the more important it is to start with small tests during the beta stage.

What to Check When Installing

The official INSTALL document primarily recommends native installation by operating system. macOS can use a Homebrew tap. Ubuntu and Debian can use the `.deb` from GitHub Releases. Windows can use an MSI installer.

There are also cautions. Script-based installation is convenient, but the official documentation also explains that separate signature verification is limited. For that reason, it is safer to use the operating system’s standard installation method or release files when possible.

Another issue is the Linux environment. The INSTALL document notes that AppImage may crash because of Wayland or some system library issues. On Debian/Ubuntu, installing the `.deb` may be the more stable choice.

Criteria for Evaluating OpenHuman

An agent platform like OpenHuman should not be judged only by its feature list. The real evaluation criteria are a little different.

1. Is the Memory Editable?

Users should be able to see what the AI remembers. They should be able to correct wrong memories and delete outdated context.

2. Are There Approval Points in Automation?

A good agent does not execute everything on its own. Tasks such as sending email, modifying files, posting externally and making payments require clear approval steps.

3. Does It Leave the Cause When It Fails?

Agents will inevitably fail. What matters is not pretending they do not fail, but showing where they got stuck and what logs or retry paths are available.

4. Is It Clear What Data Goes to External Models?

The phrase local-first is not enough by itself. You need to check what data stays local and what data is sent to external APIs.

5. Does It Conflict with My Existing Workflow?

AI tools change how work is done. That can be a positive change, but it can also create confusion in existing file structures and collaboration methods. It is best to test first in a small project.

Compared with Hermes Agent

OpenHuman gives the impression of being close to a personal desktop AI operating system. It puts local memory, Obsidian wiki, various app integrations and background loops at the front.

By contrast, Hermes Agent is strong in task execution, tool calls, skills, cron and workspace-based automation. It is closer to a method that continues from research and writing to uploading and verification. It works inside project folders and procedures the user has already defined.

The two tools differ more in perspective than in direct competition. OpenHuman is oriented toward building a “whole personal AI environment.” Hermes is closer to an “agent runtime that executes tools all the way through real work.” This difference will become more important in the age of AI agents. That is because harnesses and memory structures, more than models alone, will determine real productivity.

Conclusion: Worth Watching, but Start with Small Tests

OpenHuman is an interesting project that shows where AI agents are heading. It is an attempt to move beyond conversational chatbots and combine local memory, workflows, research and execution tools into one system.

But it is still Early Beta. Before connecting personal data and work tools, you need to check permissions, storage locations, the scope of external transmission, failure logs and the range of automatic execution. Rather than entrusting it with core work from the start, it is safer to begin with a separate test account and a small folder.

Ultimately, the question is not “Is OpenHuman good?” The more precise question is this: Are you ready to let AI continuously remember and execute around your knowledge and workflow? OpenHuman is a tool that poses that question quite clearly.

Further Reading
FAQ

Is OpenHuman an AI model like ChatGPT?

No. OpenHuman is closer to an AI agent platform that connects multiple models, tools, memory and workflows than to a specific model.

Is OpenHuman free?

The official GitHub repository is an open-source project under the GPL-3.0 license. However, actual use may require external model APIs, connected services or subscription features, so you should check the cost structure separately after installation.

What Is OpenHuman’s Biggest Strength?

It aims to combine local-first memory with agent orchestration. Rather than being a simple conversational AI, it is structured to connect the user’s files, apps, goals and automation flows.

Can I Use It for Real Work Right Now?

Because the official repository says it is Early Beta, it is better to test it first with a test account and limited data instead of applying it immediately to core work.

What Should I Be Most Careful About When Installing?

Account integration permissions and the installation method. When possible, use the official release’s native package, and connect sensitive accounts such as Gmail, Slack and GitHub with minimum required permissions.

References
Original Korean Article

This article is a full-fidelity English translation draft of the original Korean post: OpenHuman AI agent Korean article on Thinknote.
2026년 07월 21일
Will the AI Singularity Arrive Within Five Years? Five Questions for Preparing for the Agent Era
# Will the AI Singularity Arrive Within Five Years? Five Questions for Preparing for the Agent Era

Talk about the AI singularity usually flows in two directions. One is trying to guess the date when AI will surpass humans. The other is the more tangible question: when will my work and daily life actually change?

The EBS knowledge video “The latest it will arrive is five years from now” is closer to the second question. The core issue is not a grand prophecy about the future. It is what we should prepare for when AI moves beyond chatbots and approaches the agent stage, where it handles real work.

Scene from a discussion on the AI singularity

For the singularity, the “felt threshold” matters more than the date

In the video, the singularity is described as the turning point when artificial intelligence surpasses human intelligence. The speaker mentions that some AI scientists point to around 2030. That is why the phrase “it could arrive within five years” appears.

But if we focus only on the date, the discussion is easily exaggerated. There is a more important question: When will people begin to feel that AI is not just a simple tool, but a colleague at work or even a substitute?

That felt threshold is closer to agents than to the grand word “superintelligence.” When AI carries out multi-step tasks such as finding documents, comparing materials, calculating in Excel, sending emails, and coordinating schedules, people already feel, “This is a different phase.”

In relation to this topic, Thinknote’s summary of AGI and superintelligence risk is also worth reading. If that article looks at the larger risk landscape, this one focuses on changes felt in everyday work.

Why digital intelligence moves differently

One interesting point in the video is the difference between natural intelligence and digital intelligence. Human genius is bound to individuals. Even if the experience a person builds over a lifetime is recorded, it does not become another person’s ability as-is.

AI is different. The level one model reaches can become the starting line for the next model. A movement learned by one robot can also be copied across an entire fleet of robots. In the video, this is explained roughly as: “In AI, once an Einstein appears, that becomes the bottom line.”

Another difference is time. AI can simulate, in compressed time, the trial and error that humans would repeat over hundreds of years. That is why the singularity discussion is not simply about “smarter machines.” It is about changes in learning speed, replicability, and the way knowledge is transferred.

Slide showing questions for the AI era

What will change when the agent stage arrives?

As in OpenAI’s discussions of AGI stages, AI development is often described as moving from chatbots to reasoning, agents, innovators, and organization-level systems. The stage the public will most strongly feel first is the agent stage.

An agent does not stop at giving an answer. It receives a user’s goal, handles multiple apps and tools, checks intermediate results, and continues the necessary work. That is why how work changes in the agentic AI era has already become a practical topic for both individuals and companies.

Preparation must also change. Being good at prompts is not enough. You must design which tasks to entrust to AI, which data should not be entrusted to it, who will review the results, and how logs will be kept if something fails.

Hallucination is a risk and also a shadow of creativity

The video also spends considerable time on hallucination: the problem of AI producing answers that sound plausible but are wrong. In areas where errors cause serious harm, such as medicine, pharmaceuticals, law, and finance, this can be fatal.

But if hallucination is seen only as a bug, we miss something about the nature of AI. The video also introduces the view that “hallucination is not a bug but a feature.” New combinations and creative answers require some room for imagination and inference.

So the practical conclusion is not “Do not trust AI.” It is closer to use AI with verification mechanisms attached. Retrieval augmentation, source checks, calculation tools, expert review, and work logs should be used together. As discussed in the Obsidian deep-research automation article, the quality of AI use depends less on the answer itself than on the verification loop.

Embodied AI and the problems of the real world

In the latter part, humanoids and embodied AI appear. The question is whether AI that has learned only from text and images can truly understand the world. Experimenting in a lab, grasping objects, falling down, and readjusting are different from knowledge learned only through words.

Discussion of humanoids and embodied AI

Platforms such as NVIDIA Cosmos are attempts to solve this problem in virtual worlds. They simulate physical environments similar to reality and allow robots or autonomous-driving systems to accumulate large amounts of experience within them.

This point makes the singularity discussion more realistic. Rather than an AI surpassing humans suddenly appearing one day, the picture is closer to software agents and robots in the physical world developing at different speeds and entering various parts of society.

Five questions individuals and organizations should ask now

The conclusion of the video is closer to preparation than fear. AI may not be a tool that grows everyone equally. It can become a device that amplifies people who already have knowledge and resources even further.

That is why the following five questions are necessary.
1. What repetitive tasks in my work can AI already do instead? You need to separate small tasks first, such as report drafts, research, summarization, and schedule coordination.
2. What judgments should not be entrusted to AI? Human review is essential in areas with high error costs, such as legal responsibility, personnel evaluation, and medical or financial judgment.
3. Is there a loop for verifying AI results? If you do not check sources, calculations, logs, and reproducibility, AI can become a fast error-production machine.
4. Are our organization’s data and permissions designed safely? When agents manipulate real tools, permission management and work records become important.
5. Do I have enough background knowledge to ask AI good questions? As the video puts it, the AI era may be a comeback for broad knowledge. If the question is shallow, the answer will be shallow too.
Possibilities and anxieties of the AI era

Conclusion: Changes in how we work arrive before the singularity

No one can state with certainty exactly how many years remain before the AI singularity. The definition of intelligence is still not clear either. So it is better to read the number “2030” not as a prophecy, but as a warning signal.

What is clear is that the shift from chatbots to agents has already begun. AI is moving from an answering tool to an execution tool. This change touches personal productivity, organizational permission design, the direction of education, and debates over social distribution.

Closing scene from an AI singularity discussion

In the end, the core of preparation is one thing: not using AI more, but designing what to delegate, what to verify, and what questions to ask.

Recommended reading
FAQ

Will the AI singularity really arrive within five years?

The exact timing cannot be stated with certainty. However, the video emphasizes that the discussion timeline has moved forward enough for some experts to mention around 2030. More important than the date is the felt change when agentic AI begins to handle real work.

Are AGI and AI agents the same thing?

They are not the same. AGI refers to intelligence that can generally solve diverse problems like a human. An AI agent is closer to a system that receives a goal and executes multiple steps. However, many people may first experience changes that feel close to AGI at the agent stage.

Can AI hallucination disappear?

It is hard to assume it will disappear completely. Instead, risks can be reduced by adding retrieval augmentation, source checks, calculation tools, and expert review. The more important the domain, the more AI answers should be placed inside a verification loop rather than used as final judgments.

What should individuals prepare first for the AI era?

Work decomposition comes before a list of tools. You need to separate the repetitive parts of your work, the parts requiring judgment, and the parts with serious responsibility. Only then can you decide what to entrust to AI and what humans should review.

Why do broad knowledge and questioning ability matter in the AI era?

AI is strongly affected by the quality of the question. Background knowledge is needed to make good questions and judge whether the answer is correct. The ability to use AI well is therefore not merely prompt technique, but an ability to handle knowledge and context.

References
Image source: the captured images used in this article are used as quoted images from the original YouTube video for review, commentary, and educational purposes. Image copyrights belong to the original rights holders and the channel.

Original Korean article

Read the original Korean article
2026년 07월 21일

AI Web Design Workflow: How to Build a Landing Page with ChatGPT Mockups and Claude Design

# AI Web Design Workflow: How to Build a Landing Page with ChatGPT Mockups and Claude Design

AI web design is no longer just a story about “enter a prompt and a site comes out.” The more important change is a division-of-labor workflow: first visualize the design standard, then have another AI implement that standard.

Darrel Wilson’s video shows this flow well. First, he creates website screenshots with ChatGPT Sol, then passes a preferred mockup to Claude Design and turns it into a responsive web page. For beginners, it is a fast experimentation tool. For practitioners, it is a way to reduce the gap between a brief and implementation.

Example of a website mockup created by ChatGPT — AI first draws a finished-looking website mockup to create a visual reference point. Source: screenshot from Darrel Wilson YouTube video.

The core is “creating a visual standard,” not “generating code”

A common mistake when building a website with AI is to start with the prompt, “Create a cool landing page in HTML.” Even if the result looks plausible, the brand tone, image direction, section density, and typography can easily drift.

The video’s approach is different. It first asks ChatGPT Sol to create screenshots that look like finished websites by providing the industry, mood, image style, and layout requirements. Those screenshots become a visual brief that can be given to another AI.

Step 1: Create multiple design mockups with ChatGPT Sol

The text prompt should not simply say, “Make a website.” It should include specific design conditions, such as the following.

Industry and site purpose
Desired mood and brand tone
Image direction for the hero section
Layout characteristics such as vertical or horizontal text
CTA buttons, menus, and section structure
Whether to use high-resolution images

The purpose of this step is not to obtain code that can be deployed immediately. It is to compare several mockups quickly and choose the strongest direction.

Extracting image assets from a mockup — Images inside the screenshot are separated into high-resolution assets and passed to the implementation stage. Source: screenshot from Darrel Wilson YouTube video.

Step 2: Extract image assets separately and pass them to Claude

Even if a strong mockup is produced, the final result weakens sharply if the images turn into placeholders during implementation. That is why the video extracts the images inside the screenshot as high-resolution files and downloads them as a ZIP file.

The reason this process matters is simple. Claude Design can follow not only the layout, but also the image assets that created the mood of the original mockup.

In real work, one more check is needed at this stage. You must review the commercial usability of AI-generated images, human depictions, brand similarity, and copyright risk.

Claude Design implementation screen — The selected mockup is placed in Claude Design and translated into an actual web page structure. Source: screenshot from Darrel Wilson YouTube video.

Step 3: Implement the mockup as a responsive web page with Claude Design

The next step is to upload the selected screenshot and image ZIP to Claude Design. In the video, Wilson instructs Claude to follow the original mockup as closely as possible and, if necessary, turns off Claude’s default design system.

The prompt can be short. What matters is not saying only “Make a similar website based on this image,” but giving the following standards as well.

Preserve the original mockup’s layout first.
Do not replace the image assets with placeholders.
Consider both desktop and mobile responsiveness.
Connect the menu, CTAs, and section order like a real site.
Structure it so it can later expand into About, Services, and Contact pages.

Step 4: Do not stop at the homepage; expand the site structure

In the video, after creating the homepage, he generates additional pages such as About, Services, Insight, and Contact in the same design language. This part is important. Even if a single landing page looks beautiful, it is hard to use as a real website if the internal pages are empty.

AI website production should not stop at “the first screen looks pretty.” At minimum, you need to check the following.

Do navigation links lead to actual pages?
Do the menu and CTA work naturally on mobile?
Do the contact form, buttons, and external links function correctly?
Is the copy not duplicated across pages?
Are SEO titles and meta descriptions separated by page?

Adjusting content with prompts — Industry information and wording are entered again to customize page content. Source: screenshot from Darrel Wilson YouTube video.

Step 5: Refine copy and animation separately afterward

Claude’s first result is a starting point. In the video, industry information is entered again to customize all text, and animations such as birds, clouds, and human video elements are added.

Animation, however, should be handled carefully. Background videos and moving objects can improve the first impression, but they can also reduce mobile speed and accessibility. On small screens in particular, text readability comes first.

In practice, the following order is stable.

Complete the structure and sections first.
Have a human review the brand copy.
Fix the mobile layout.
Add animation in minimal units.
Check speed, accessibility, and SEO last.

Example of the deployment stage — The completed HTML is connected to hosting and published on a real domain. Source: screenshot from Darrel Wilson YouTube video.

What this method changes is the work sequence, not the tools

The message of the video is not “designers and developers are no longer needed.” Rather, it means the human role is moving further toward the front end and the back end of the process.

At the front end, people must create good visual briefs and judge which direction to choose among multiple mockups. At the back end, they must verify whether the result meets real service standards.

As AI becomes faster, the standards humans need to check must also become clearer.

Stage	What AI does well	What humans should check
Mockup generation	Suggesting varied layouts and image directions	Brand fit, differentiation, copyright
Implementation	Structuring HTML/CSS and generating a responsive draft	Code quality, accessibility, performance
Content	Drafting industry-specific copy	Accuracy, persuasion, legal wording
Deployment	Producing publishable files quickly	Hosting, domain connection, forms, security

Checklist for trying it right away

Do not try to finish everything at once; separate mockup creation, implementation, revision, and deployment.
Give ChatGPT the “design direction” and Claude the “implementation standards.”
Pass image assets along with the screenshot.
Check the mobile screen with a separate prompt.
Before actual deployment, check links, forms, speed, accessibility, and SEO.

FAQ

Q1. Can ChatGPT Sol complete a website by itself?

The core of the video is not to finish with Sol alone. ChatGPT Sol creates design mockups and image assets, while Claude Design implements those mockups as actual web pages. It is closer to a division of labor.

Q2. Why create a screenshot first instead of asking Claude to build it directly?

A screenshot is a visual brief that communicates layout, images, typography, and mood all at once. Compared with text alone, it gives the AI a much more concrete standard to follow.

Q3. Does this method replace Figma?

For simple mockups and landing page experiments, it can reduce some work before Figma. However, it is hard to say that it fully replaces team collaboration, component management, design systems, and detailed UX validation.

Q4. Can I use an AI-generated website commercially right away?

It is safer to use it after review rather than immediately. Image rights, responsive quality, accessibility, performance, personal-data handling in forms, and search optimization all need to be checked separately.

Q5. Is this workflow useful for beginners?

Yes. It is especially useful for beginners who lack design confidence because they can view multiple mockups first and choose a direction. However, they should not trust the final result as-is, but verify it step by step with a checklist.

References

Original Korean article

Read the original Korean article

2026년 07월 21일

Kimi K3 Shock and Controversy: Five Questions China’s Open AI Model Raises

Kimi K3 is not just another model announcement. It matters because Moonshot AI, a Chinese startup, has introduced an open 3-trillion-class model that challenges the competitive map of the AI industry.

According to Moonshot AI’s official documentation, Kimi K3 has 2.8 trillion parameters, a 1M-token context window, native multimodal understanding and a strong focus on long-horizon coding and knowledge work. CNBC, BBC and other major outlets have framed it as a Chinese open-model challenge to the closed frontier systems led by large U.S. technology companies.

But the Kimi K3 shock cannot be understood through hype alone. Its benchmark performance, the accuracy of the “open source” label, distillation allegations, chip-market reaction and enterprise adoption risks all need separate judgment. The real question is not simply, “Has China beaten the United States?” It is, “What is the new standard for AI competition?”

What Is Kimi K3?

Kimi K3 is Moonshot AI’s flagship model, announced in July 2026. The official documentation highlights four core specifications.

2.8 trillion parameters: Moonshot AI presents Kimi K3 as a first open model in the 3-trillion-parameter class.
1M-token context window: The model is positioned for long documents, codebases, meeting records and extended knowledge work.
Native multimodal capability: Kimi K3 is described as a model that can handle visual input as well as text.
Long-horizon coding and knowledge work: Its main use case is not only short question answering, but agentic coding and complex work execution.

The documentation also mentions Kimi Delta Attention, Attention Residuals and a Mixture of Experts architecture. The model reportedly activates 16 out of 896 experts, which suggests an attempt to combine very large scale with more efficient inference.

Why Did Kimi K3 Create Such a Shock?

The first reason is performance. Moonshot AI and several outside reports say Kimi K3 is close to, and in some task areas ahead of, top GPT and Claude-family systems in coding, web interface engineering and agentic tasks.

The second reason is cost and access. Several Korean and global reports argue that Kimi K3 is being positioned with a lower cost structure than leading U.S. frontier models. If a cheaper model performs well enough, companies will naturally ask whether they should remain locked into one premium API provider.

The third reason is the release strategy. Kimi K3 is described as open source or open weight, unlike closed API-first systems from U.S. labs. This distinction matters. Releasing weights does not automatically make training data, training procedures or safety evaluations fully transparent. For that reason, it is safer to treat Kimi K3 as a strategic open-weight model, rather than accepting the marketing phrase “open source” without qualification.

Controversy 1: How Much Should We Trust the Benchmarks?

Benchmarks are at the center of the Kimi K3 debate. On paper, Kimi K3 appears on the same leaderboard as top closed frontier models. Reports especially highlight its strength in coding and agentic work.

The problem is that benchmark scores do not capture every risk in real enterprise use. GovInfoSecurity, for example, argues that Kimi K3 shows the limits of AI leaderboards. A high test score does not automatically prove security, consistency, long-term reliability, sensitive-data handling, incident response or regulatory compliance.

So the practical question is not, “Which model won by a few points?” Companies should ask more concrete questions.

Does the model perform equally well on our own business data?
Does a long context window actually reduce hallucination and omission?
Does generated code pass tests and security checks?
Can we switch to another model if policies, access or pricing change?
Is the cost saving larger than the added review, security and governance cost?

Controversy 2: What Should We Make of the Claude Distillation Allegations?

Some media reports and social posts have claimed that Kimi K3 sometimes identified itself as Claude. They used that behavior as a basis for distillation allegations. In this context, distillation means using the outputs of a stronger model to train or improve another model.

The allegation is sensitive. U.S. AI companies are increasingly concerned that their model outputs may be used to train competitors. On the other side, Chinese officials and some analysts see these complaints as part of a broader geopolitical technology dispute.

A balanced view requires three points. First, a model misidentifying itself as another model is not enough to prove illegal distillation. Second, the use of model-output data is a gray area across the industry, not only a China-specific issue. Third, as TechCrunch reported, some experts argue that Kimi K3’s performance cannot be fully explained away by distillation alone.

The deeper issue is not whether Kimi K3 is “fake.” The bigger question is whether AI competition is moving from pure performance races toward disputes over training-data provenance, output rights and model supply-chain transparency.

Controversy 3: Is Kimi K3 Bad News or Good News for Chipmakers?

After the Kimi K3 announcement, Korean market commentary split over its possible impact on Samsung Electronics and SK Hynix. Some reports compared it with the earlier DeepSeek shock and asked whether a cheaper, highly capable Chinese model could weaken the investment logic behind massive GPU spending.

That is the bearish view. If China can produce strong models at lower cost, investors may wonder whether the demand for high-end AI chips will slow.

There is also a bullish view. If more high-performance open-weight models become available, more companies and developers may want to run their own inference infrastructure. That could expand demand for memory, servers, inference chips and data centers. In other words, the center of gravity may shift from training to inference, deployment and optimization.

Kimi K3 is therefore not simply a threat to semiconductor demand. It may be a signal that AI infrastructure demand is changing shape.

The Real Innovation Is Not Just That China Got Faster

If we read Kimi K3 only as a victory for Chinese AI, we miss the larger change. The real shift is the speed of open-model diffusion. High-performance models are no longer staying only inside closed APIs. They are moving faster into broader developer and enterprise ecosystems.

That creates three pressures.

Price pressure: Premium API pricing becomes harder to justify when open-weight alternatives improve.
Product pressure: Model companies must offer agents, tools and workflows, not just raw model access.
Policy pressure: Governments must think about AI access, open-weight release, data rules and export controls at the same time.

This connects directly to Thinknote’s earlier discussion of small language models and open source AI. The AI market may look like a winner-take-all race from the outside. In practice, it is becoming layered by model size, cost, openness and deployment location.

What Should Korean Companies Watch?

For Korean companies, Kimi K3 does not mean “use this model immediately.” It means that model selection criteria must change.

First, companies need a model portfolio. GPT, Claude, Gemini, Kimi and open-weight models should be evaluated by task. Locking every workflow into one API increases both cost and strategic risk.

Second, internal evaluation sets matter more than public benchmarks. Companies need to test models on their own documents, code, customer support cases, reports and data-analysis tasks.

Third, AI coding depends on the harness, not only the model. As Thinknote argued in The Essence of AI Coding Is Not the Model but the Harness, tests, reviews, deployment controls and rollback systems decide whether a powerful coding model becomes useful automation or risky automation.

Fourth, sovereign AI should be understood realistically. As discussed in Anthropic Mythos Shock, the point is not to reject foreign models entirely. The point is to secure alternative paths for strategically important work.

Fifth, the transition to agentic AI will accelerate. Kimi K3’s emphasis on long-horizon coding and knowledge work shows that AI is moving from chatbots toward work-execution systems. That connects with Thinknote’s broader argument about how work changes in the agentic AI era.

What It Means for Individual Users

Kimi K3 also matters for individual users because it expands the menu of choices. The important question is no longer, “Which chatbot is the smartest?” The better question is, “Which model mix fits my purpose?”

If you code, you should evaluate file editing, test execution and code-review flow. If you handle long documents, you should test whether a 1M-token context window actually improves summary quality. If you automate work, you should check cost, speed, privacy handling and log-retention policies.

The Kimi K3 shock is not a declaration that one model has won. It is a signal that AI users need to become more demanding buyers.

Five Criteria for Judging Kimi K3

Criterion	Question to Ask	Why It Matters
Performance	Does it work well on our own data?	Public benchmarks may not match real-world performance.
Cost	Is it still cheaper after input, output and caching costs?	Long context changes the real cost structure.
Transparency	What is disclosed beyond model weights?	Open weights and full open source are not the same thing.
Risk	Are data security, regulation and supply-chain risks manageable?	Chinese model adoption requires governance review.
Portability	Can we switch to another model easily?	Model dependence should be designed down from the start.

FAQ

Has Kimi K3 completely beaten OpenAI or Anthropic?

Not yet. Kimi K3 appears strong in some benchmarks and coding tasks, but overall performance and enterprise reliability still require independent verification.

Is Kimi K3 really open source?

Moonshot AI describes it as open source, but users should check what is actually released. Model weights, training data and full training procedures are different levels of openness.

Are the Claude distillation allegations proven?

No public evidence currently proves the allegation. There are reports and suspicious examples, but there are also expert views that Kimi K3’s performance cannot be explained only by distillation.

Should Korean companies adopt Kimi K3 right away?

Not immediately. They should first run internal evaluations that cover performance, security, cost, regulation and model-switching options.

Is Kimi K3 bad for semiconductor companies?

It may disturb short-term investor sentiment. Over the longer term, however, open-model adoption could expand inference infrastructure and memory demand.

Conclusion: Kimi K3 Shows the New Rules of AI Competition

Kimi K3 can be summarized in one sentence: top-tier AI may no longer be the exclusive territory of closed U.S. frontier models.

That signal should not be exaggerated. Benchmarks are only a starting point. Distillation allegations remain unproven. The “open source” label still needs careful interpretation.

The real change is the growth of choice. Companies and individuals now need to choose AI models by evaluation systems, data governance, cost structure and portability, not by model names alone. The winners of the next AI wave will not be the people who chase every new model announcement first. They will be the people who can compare, combine and govern those models safely.

Sources

Original Korean Article

This article is an English translation of the original Thinknote post: Original Korean article.

2026년 07월 21일

Parenting in the AI Era: Five Abilities Children Need Before a Good University
When an era arrives in which AI can study on a child’s behalf, what should parents leave their children with? In an interview with Jisik Inside, Professor Jo Byeok raises this question quite directly. If parents hold only to good universities, high scores, and more private education, they may fall one step behind the changes of the AI era.

The point is not “let’s stop studying.” Basic knowledge is still necessary. But using all of a child’s time to chase correct answers is becoming increasingly risky. In an age when AI can find correct answers quickly, the ability to ask questions, build relationships, and interpret one’s own experiences becomes more important.

Source: Screenshot from the Jisik Inside YouTube video

Why a good-university strategy is no longer enough

For Korean parents, the strategy for a child’s success has long been simple: get good grades, enter a good university, and secure a stable job. Professor Jo Byeok says this strategy was quite powerful in the past, but it may not work the same way in the AI era.

The video includes a striking analogy. In a family photograph taken about 100 years ago, three brothers are in the same place at the same time, yet they look as if they are living in completely different eras. One is holding on to old symbols of success, while another has moved toward the education of a new age. The scene asks today’s parents the same question: Is the “good path” we are holding on to really the path our children will live on?

Source: Screenshot from the Jisik Inside YouTube video

1. Character is not etiquette; it is a capability in the AI era

The first message in the video is that “character is also a capability.” Here, character does not simply mean being kind and polite. Professor Jo describes character as a uniquely human quality: communication, empathy, collaboration, and resilience, all human strengths that AI has difficulty replacing.

In the past, when knowledge and skills came first, character was sometimes treated as an extra. But as AI increasingly handles knowledge processing and the search for correct answers, the situation changes. People still have to work with people and solve ambiguous problems together. That is why character is no longer merely “nice to have,” but a core competency that helps a child endure over the long term.

2. The ability to ask questions is deeper than prompt technique

As the AI era begins, many people say that we need to “ask good questions.” The ability to write good prompts is certainly necessary. But the questioning ability discussed in the video is broader than that. It is closer to a child taking ownership of their own learning.

Children naturally ask many questions. But as they grow older, the number of questions decreases. That ability is suppressed by phrases such as “Don’t think about useless things; just study,” “Hurry up and do your homework,” and “Solve this problem first.” What parents need to do is not give children more answers, but restore an atmosphere in which it is safe to ask questions.

Source: Screenshot from the Jisik Inside YouTube video

3. Future literacy is not the ability to predict the future

Professor Jo does not describe “future literacy” as the ability to forecast the future. Instead, he describes it as the ability to create the future one wants to live in. This distinction is important.

AI may be better at prediction. Reading data and patterns and calculating possible scenarios are AI strengths. But deciding what future we want, what life we will choose, and what relationships we will protect is the human role. That is why parents should help children become not “children who guess the right answer,” but “children who design their own future.”

4. Unique matters more than best

Admissions competition keeps children comparing themselves with others. The standards become who scored higher, who entered a better school, and who got ahead faster. But in the AI era, competitiveness does not have to come only from beating others.

Professor Jo emphasizes “unique” over “best.” When a child has their own experiences, interests, questions, failures, and stories of recovery, they can become competitive without constantly competing. A person who quickly gives the same correct answer as everyone else will be compared with AI. But a person who sees problems from their own perspective and contributes through collaboration with others is not easily replaced.

Source: Screenshot from the Jisik Inside YouTube video

5. A parent’s first question should be about feelings, not scores

The final advice in the video is the most practical. When a child comes home from school, parents who ask “What did you learn today?” or “What score did you get?” may be moving against the AI era. A parent who instead asks “Did you have fun at school today?” connects with the child’s emotions.

This does not mean giving up on study. It means restoring the child’s vitality and relationships first. The parent-child relationship is not a project that disappears once college admissions are over. It is a lifelong relationship. When that relationship feels safe, children ask more questions, explore farther, and stand up again even after failure.

Source: Screenshot from the Jisik Inside YouTube video

Five questions parents can change today

Parenting in the AI era is not something that has to wait for sweeping institutional reform. Parents can begin by changing the questions they use at home.
1. Instead of “What score did you get today?” ask “What was the most interesting moment today?”
2. Instead of “Why don’t you even know that?” ask “Where did it start to feel confusing?”
3. Instead of “That dream is unrealistic,” ask “What experience would help you get closer to that dream?”
4. Instead of “Everyone else is doing it, so why aren’t you?” ask “What would you like to try in your own way?”
5. Instead of “Hurry up and say the correct answer,” ask “What would happen if we changed it into a different question?”
When the question changes, the way a child brings out their own thoughts also begins to change little by little. It may feel awkward at first. As the video says, it is not so much difficult as unfamiliar.

Recommended reading
FAQ

Will school study become less important in the AI era?

Basic knowledge is still important. However, spending all available time only on score competition is risky. On top of basic learning, children also need to develop questioning ability, AI literacy, empathy, collaboration, and resilience.

Does the character Professor Jo talks about mean only being nice?

No. In the video, character is closer to a uniquely human quality. It includes abilities that AI has difficulty replacing, such as communication, empathy, collaboration, and resilience.

How can parents support a child’s questions?

Rather than giving the correct answer immediately, first listen to what kind of thinking led to the child’s question. If you create an atmosphere that welcomes questions without judging them, children can gradually regain ownership of their own learning.

Is it wrong to aim for a good university?

The goal itself is not wrong. The problem is treating a good university as the only strategy for success. In the AI era, what matters more than the name of the university is what questions a child can create, how they interpret their experiences, and what contribution they can make.

References
- Jisik Inside, “Good universities will be discarded”: Conditions parents need to help children survive the AI era
- UNESCO, AI competency frameworks for students and teachers
Original Korean article: Parenting and education in the AI era with Jo Byeok

Image source: Captured images used in this article are stills from the original YouTube video. They are used for review, commentary, and educational explanation, and copyright remains with the original rights holders and the channel.
2026년 07월 17일
Build a One-Person AI Secretary with Claude Cowork: Automated Gmail, Notion, and KakaoTalk Briefings
If you use Claude only as a “chatbot that answers questions,” you will hit its limits quickly. The story changes when you build a structure that reads your email, organizes your tasks, checks your schedule and the weather, and reports to you every morning. Based on the flow of the Soso AI Beginner Notes video, this article explains how to connect Claude Cowork with Gmail, Notion, and PlayMCP to create a one-person AI secretary workflow.

Example of an AI daily briefing delivered through KakaoTalk · Screenshot from the original video

The key is not the prompt, but the flow of recurring work

The most important part of the video is not “one great prompt.” It is the full workflow: reading email in Gmail, drafting replies, organizing tasks in Notion, then connecting weather and KakaoTalk briefings through PlayMCP.

That difference matters. With one-off questions, a human has to explain the situation again every time. A work flow, on the other hand, can run at the same time each day, by the same standards, in the same output format once it has been designed. This is where the difference appears between people who use AI for work and people who build work systems with AI.

Step 1: Connect Gmail in the Claude desktop app

First, open the custom menu in the Claude desktop app and choose Connectors. In the video, after connecting the Gmail connector, the user asks Claude Cowork in collaboration mode: “Tell me about the four most recent emails.”

Claude Gmail connector setup screen · Screenshot from the original video

Two useful capabilities become obvious right away.
- It can summarize recent emails so you can quickly understand what needs attention today.
- For messages that need replies, it can create draft responses so the user only has to review and send them.
However, email contains a lot of sensitive information. At first, try connecting test emails or low-importance messages. It is also safer to create a separate work folder that Claude can access and keep important documents out of that folder.

Step 2: Automatically organize tasks in Notion

Checking email alone does not make an assistant. The assistant has to pull the work hidden inside those emails and move it into an actual task list. In the video, after connecting the Notion connector, the user asks: “Create this week’s work schedule as a to-do list page in my Notion.”

Building a work flow in the Claude collaboration screen · Screenshot from the original video

The advantage of this step is that the inbox and the task list stay connected. Usually, things get missed when we read email, make a separate note, and then move it again into Notion or a to-do app. If Claude reads and structures the email content, the human only has to decide priority and whether to act.

You do not need to build a complicated dashboard from the beginning. Simple fields like the ones below are enough.
- Task name
- Related email or requester
- Deadline
- Priority
- Status
- Next action
Step 3: Add KakaoTalk, weather, and map information with PlayMCP

If Gmail and Notion organize work information, PlayMCP expands the connection to external tools. In the video, the user adds KakaoTalk’s chat with myself, KakaoMap, and weather tools to the PlayMCP toolbox, then connects them with Claude.

Selecting weather and KakaoMap tools in PlayMCP · Screenshot from the original video

With this connection, the briefing becomes much more practical for daily life. For example, you can ask: “Check today’s weather forecast, find five good restaurants near the off-site work location mentioned in Gmail, and send them to me on KakaoTalk.” It becomes not just a work summary, but actual decision-making material for starting the day.

There is also something to be careful about here. Map, weather, and messenger tools are connected to permissions for external services. You must check which account is logged in, which permissions are granted to each tool, and where any automatic messages will be sent.

Step 4: Run it every morning with a skill and a scheduled task

The final exercise in the video is to create a daily briefing skill inside Claude and set it to run every day as a scheduled task. Once created, the skill remembers an order such as “check email → draft replies → update Notion → check weather and off-site information → report through KakaoTalk.”

Daily briefing result automatically sent through KakaoTalk · Screenshot from the original video

When a scheduled task is added, the user no longer has to type the prompt every time. While you are getting ready for work in the morning, AI can prepare the day’s materials first.

At this point, skipping approval or enabling automatic execution is convenient, but it also increases risk. Start by testing with manual runs, check that no wrong emails are sent and no sensitive information is exposed, and only then raise the level of automation.

Checklist before you follow along

Before applying this workflow directly, organize the following four points first.
1. Decide which work repeats every day. High-frequency tasks such as checking email, organizing schedules, drafting reports, and responding to customers are good candidates.
2. Separate the accounts and data ranges that AI may access. Instead of opening your entire personal mailbox immediately, define a test scope first.
3. Choose one channel where you will receive the output. Decide whether you will check it in KakaoTalk, Notion, or email so the flow stays simple.
4. Run at least three manual tests before enabling automatic execution. Check summary quality, omissions, permissions, and message recipients.
Who this method fits, and who should wait

This method works well for people who have a lot of recurring work and move between several tools. If you check email every day, have meetings or off-site work, and already use a task management tool such as Notion, you can see the benefits quickly.

On the other hand, if concepts such as Claude desktop, Notion, and MCP still feel unfamiliar, it is better not to connect everything at once. Start by connecting only Claude and Gmail and testing email summaries and reply drafts. Then add Notion, and finally add PlayMCP and scheduled tasks. That is the safer order.

Recommended reading
Frequently asked questions

How is Claude Cowork different from regular chat?

Regular chat is conversation-centered. Cowork, or collaboration mode, is closer to building a workflow that uses specific folders, files, connectors, and tools together. In the video as well, the user designates a work folder through the collaboration button in a new chat, then continues with email and file work.

Is it safe to connect Gmail and Notion?

The scope of permissions and your usage habits matter more than the tool itself. Do not automate every email and document from the beginning. Start with a test folder and low-risk work. For automatic sending features in particular, it is safer to keep a manual review step.

Is PlayMCP required?

To connect tools outside Claude’s basic connectors, such as KakaoTalk, KakaoMap, and weather, you may need an external tool-connection layer like PlayMCP. If you only want email summaries and Notion organization, you can start without PlayMCP.

Can I turn on automatic execution every morning right away?

It is not recommended. First, check the results with manual execution and confirm that sensitive information is not exposed and messages are not sent to the wrong chat room. After that, it is safer to add a scheduled task and automate it.

Is this article a direct transcript of the original video?

No. It is based on the practice flow of the original video, but it reduces the promotional latter section and reorganizes the setup order and cautions from the perspective of practical work automation.

References
Original Korean article: Claude Cowork one-person AI secretary workflow

Image source: Captured images used in this article are stills from the original YouTube video. They are used for review, commentary, and educational explanation, and copyright remains with the original rights holders and the channel.
2026년 07월 17일
Metacognition in the AI Era: How to Check Your Thinking Before Trusting Smart Answers

Metacognition is the power to step outside your thoughts and look again at the state of your thinking.

Before leaving work, you ask ChatGPT to draft a report. The answer comes quickly. The sentences look plausible. But something bothers you.

“Is this right?”

In the past, the ability to find answers mattered. Now it is different. Answers appear too easily. The problem is noticing whether I truly understand the answer, whether I can trust it, and whether I have adapted it to my situation.

The needed ability here is metacognition. Simply put, metacognition is “knowing what I know and what I do not know.” It may sound like the secret of good students, but today it is becoming a basic capability for office workers, creators, educators, and AI users.

## Metacognition is the ability to look at thinking once more

Metacognition sounds like a difficult psychology term, but in daily life it is familiar.

When solving a problem, you may realize, “I thought I knew this concept, but I cannot explain it.” In a meeting, you may pause and ask, “Am I stating a fact or a guess?” While writing, you may notice, “The sentences are smooth, but the logic is empty.”

All of these moments connect to metacognition. The core is stepping back. Do not remain trapped inside thought; look again at the state of your thinking.

Metacognition is therefore not simple self-reflection. More precisely, it is a technique for adjusting judgment. It distinguishes what you know from what you do not know, checks the gap between confidence and evidence, and changes strategy when necessary.

## Why metacognition matters again now

Metacognition is an old concept, but it has become important again because generative AI is changing our thinking process.

In a 2025 CHI paper, researchers from Microsoft Research and Carnegie Mellon analyzed 936 generative-AI use cases from 319 knowledge workers. A notable result appeared: the more users trusted AI, the less critical thinking they tended to perform; the more confident they were in their own task, the more critical thinking they tended to perform.

It would be too simple to read this as “AI makes people think less.” The more important message is that people who use AI well neither reject AI answers unconditionally nor accept them unconditionally. They verify answers, integrate them into their own context, and keep final responsibility.

UNESCO also released AI competency frameworks for students and teachers in 2024. These frameworks treat AI not only as tool-use skill but as human-centered judgment and responsible use. Education is shifting from “Can you use AI?” to “Can you check your thinking with AI?”

Plausible answers can help understanding, but they can also create the illusion of understanding.

## The illusion that grows as AI becomes smarter

The biggest danger in the AI era is not only wrong answers. A subtler danger is the illusion that “I understood.”

When you read text organized by AI, your head feels clearer. The summary is neat and examples are included. But when you try to explain it to someone, you may be unable to speak.

At that moment, you may possess information without understanding it.

Recent arXiv studies discuss similar concerns. AI can raise the level of individual creative output, but group-level diversity of ideas may decline. Long reasoning traces or explanations from LLMs can increase user confidence, but do not always improve actual task performance.

Some of these papers are still preprints, so they should be read carefully. Still, the direction is clear: AI explanations can help understanding, but they can also create the feeling of understanding.

That is why metacognition is needed. Do not ask only “Is the answer good?” Ask “To what level do I understand this answer?”

Good questions lead us to check the evidence and gaps in our own judgment instead of simply trusting AI answers.

## Five questions that build metacognition

Metacognition is not a matter of innate intelligence. It is closer to a habit. These five questions alone can improve the quality of thinking.

### 1. What am I mistaking as knowledge right now?

The first thing to check is illusion. Familiar words feel known, but familiarity and understanding are different.

A good method is one-sentence explanation. After reading a concept, explain it in one sentence as if to an elementary-school student. If you get stuck, it is not yet your knowledge.

AI answers are the same. Do not copy them as-is; ask, “How would I say this in my own words?”

### 2. Does my confidence come from evidence or atmosphere?

People trust content more easily when sentences are smooth. AI answers are especially like this. A confident tone, organized lists, and expert terms quickly create trust.

Metacognition asks where confidence comes from. Is my certainty based on data, experience, authoritative sources, or merely plausible sentences?

For work reports, sources must be checked. For investment, policy, and health topics, this matters even more.

### 3. Could opposing evidence change my judgment?

When metacognition is weak, people protect their own thoughts. When it is strong, people test them.

The same attitude is needed with AI. Ask, “What are the objections to this claim?” “Under what conditions could this conclusion be wrong?” and “How could this be interpreted from another perspective?” The quality of the answer changes.

The point is not to add objections formally. Your judgment must be able to change in practice.

### 4. Am I looking for an answer, or do I want to stop thinking?

The busier we are, the more we want answers. More precisely, we want to end thinking. AI satisfies this desire very well.

The problem is that fast closure is dangerous for important judgments. Hiring, strategy, curriculum design, writing, and business planning do not end with one right answer. They contain context, purpose, and stakeholders.

The metacognitive question is simple: “Do I need a conclusion now, or do I need exploration?” Distinguishing those moments is important.

### 5. Can I verify this with the next action?

Good thinking becomes verifiable action. Metacognition is weak if it remains only internal reflection.

If you wrote something, have one person read it. If you made a lecture plan, test it with a five-minute explanation. If AI recommended a strategy, try a small experiment first.

When you move from “it seems right” to “let’s check it small,” thinking becomes real capability.

A good AI-use routine includes verification, reconstruction, and final judgment, not only fast answers.

## A metacognitive routine for work and learning

Metacognition does not require grand training. Put it into the day as a short routine.

Before starting work, write three things: what I know, what I do not know, and what I need to check. Before a meeting, write your assumptions. After a meeting, leave one line about what changed in your thinking.

When using AI, the routine should be clearer:

– First, write a short draft of your own.
– Ask AI to improve it.
– Separate facts, interpretations, and suggestions in the AI answer.
– Mark parts that need sources.
– Rewrite the final sentence with your own judgment.

The order matters. If you hand everything to AI from the beginning, your own standard disappears. If you make your own draft first, AI becomes a checker rather than a replacement.

## Metacognition is a human speed in the AI era

AI is fast. So we feel we must become faster. But not every thought should become faster.

Important work needs slow zones: time to pause, doubt, explain again, and verify through small experiments.

Metacognition protects that slow zone. It is not lazy hesitation; it is an intentional pause for better judgment.

People who use AI well in the future will not only know many prompts. More important will be the ability to see the state of one’s own thinking: what I know, what I do not know, when to trust AI, and when to check again.

That is metacognition. Today it is becoming central not only to study methods but also to how we work and learn.

## Further reading

– [Human Value in the AI Era](https://www.thinknote.co.kr/ai-era-human-value/)
– [What Will Winners Prepare in the AI Era?](https://www.thinknote.co.kr/ai-era-winner-preparation/)
– [Creative Thinking Has Become More Important in the AI Era](https://www.thinknote.co.kr/creative-thinking-kim-jung-woon/)

## FAQ

### What is metacognition?

Metacognition is the ability to notice what you know and do not know and adjust learning or judgment strategies accordingly. In simple terms, it is the ability to look at your own thinking once more.

### Does high metacognition help study?

Generally, yes. People with strong metacognition find what they do not know quickly and can change learning methods. Knowing where to check is more important than simply studying longer.

### Why is metacognition important in the AI era?

AI quickly gives plausible answers. Users may therefore think they understand things they do not understand. Metacognition helps verify AI answers and judge them again in one’s own context.

### How can metacognition be trained?

The easiest method is a questioning habit: What do I know? What do I not know? What is the evidence? What is the opposing evidence? How can I test this in a small way?

### Does using AI weaken metacognition?

Not always. If AI is used only as an answer provider, thinking may shrink. But if it is used for draft review, objections, source checking, and experiment design, it can strengthen metacognition.

## References

– [Microsoft Research, The Impact of Generative AI on Critical Thinking, CHI 2025](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/)
– [UNESCO Digital Library, AI competency framework for students, 2024](https://unesdoc.unesco.org/ark:/48223/pf0000391105)
– [UNESCO Digital Library, AI competency framework for teachers, 2024](https://unesdoc.unesco.org/ark:/48223/pf0000391104)
– [arXiv, Individual Gain, Collective Loss](https://arxiv.org/abs/2606.05532)
– [arXiv, Explaining Too Much?](https://arxiv.org/abs/2605.25856)
– [arXiv, Guided Sensemaking](https://arxiv.org/abs/2606.02260)

[Original Korean article](https://www.thinknote.co.kr/metacognition-ai-thinking-checklist/)

2026년 07월 10일
In the AI Era, What You Need to Learn Before Prompts Is Your Own Language

Where does the difference come from between people who use AI well and those who do not? Many answer, “Do they know good prompts?” In reality, the issue is different. The core is not a few prompt sentences, but whether I can say what I want with context and criteria included. The Ildangbaek video “Human intelligence expressed delicately through language: the beginning of AI prompt engineering” shows this point well. The starting point is the book *AI Language Lessons for Intellectual Conversation*, but the conversation is less a book introduction than a question about language sense in the AI era. For Korean users, the question is even more important: when talking with AI in a language rich in omission and nuance, what must we make clearer? ## The AI-use gap comes from “language resolution,” not tool operation
A good prompt begins with organizing thoughts and criteria before sentence technique.
Many people first say to AI: “Organize this for me.” “Do not make it too long.” “Do not sound stiff.” “Do not draw an image; just show the prompt.” Between people, this often works because we read surrounding context, facial expressions, previous conversations, organizational culture, and tone. AI, however, guesses context the user did not provide. When the guess is right, it is convenient; when it is wrong, the result goes off track. Prompt guides from OpenAI and Anthropic commonly emphasize clear instructions, enough context, and the desired output format. A good prompt is not a magic sentence. It is a sentence that reduces what AI must guess. The key difference appears here. People who use AI well do not simply write longer questions. They structure context: purpose, reader, constraints, examples, preferences rather than bans, output format, and verification criteria. ## Why Korean is harder for AI
Korean omissions and nuance require clearer context explanation for AI.
One of the video’s most interesting points is the high-context nature of Korean. Korean often omits subjects and objects. A single particle changes focus. Honorifics may be processed on the surface, but sarcasm and irony are different issues. For example, sentences equivalent to “Cheolsu went to school” with different Korean particles may look similar but have different focus. “It is okay” may mean acceptance or refusal. Mixed emotions such as relief and regret vary by context. Humans read these differences from the situation. AI mostly receives text. Korean users therefore need to provide more context. “Do it appropriately” is convenient for humans but information-poor for AI. This also appears in translation. Anthropic’s interpretability research suggests large language models can connect multilingual inputs to shared internal concept spaces. But that does not mean Korean nuance is perfectly preserved. Emotion, omission, irony, and speaker intention can be lost when moving between languages. ## Prompt engineering is not “asking well”; it is operating a system
In organizations, prompt engineering goes beyond asking questions and becomes quality and operations design.
The video distinguishes prompts from prompt engineering. Everyday users can ask AI conversationally. But business systems, customer service, automation, and content pipelines are different. Prompt engineering is not simply “the skill of asking beautifully.” It examines how models differ in response tendencies, analyzes why wrong answers appear, designs structures that reduce cost, connects multi-step work reliably, and controls consistency. Writing may require creativity, but customer notices or legal and policy guidance should not vary every time. In such cases, generation settings like temperature, example-based output, verification steps, and retry conditions are needed. Prompt engineering is therefore both a language skill and an operations skill. It starts from an individual’s questioning habits, but in organizations it expands into quality and cost management. ## “Do this” is stronger than “Do not do that”
For AI, it is more stable to give desired direction and criteria than only prohibitions.
One practical tip repeated in the video is to use positive statements rather than negatives. “Use everyday words” is better than “Do not use technical terms.” “Write within three sentences per paragraph” is better than “Do not make it long.” “Write as a short explanatory passage” is clearer than “Do not use a list.” AI does not always handle negative instructions reliably. In image, video, or multimodal models, negative words can blur the desired result. Even in text models, “do not” can place the banned element at the center of context. | Common request | Better request | |—|—| | Do not write too difficultly | Use everyday words a middle-school student can understand | | Do not make it long | Explain only three key points within 600 characters | | Do not sound like AI | Mix short and long sentences and reduce repeated expressions | | Organize it for me | Organize it in the order of background, key issue, and action items | | Do not include subjective opinions | Separate confirmed facts from interpretation | This difference looks small, but the result changes greatly. Reducing the space for AI to guess reduces revision time. ## The productivity debate is about what you delegate, not how much you use AI
AI productivity depends on the ability to divide what AI should handle from what people must judge.
Opinions differ on whether AI truly improves productivity. Some research already observes concrete effects. NBER’s “Generative AI at Work” found that generative AI tools increased average productivity in customer support and especially helped less-experienced workers. The ILO’s analysis of generative AI and jobs, by contrast, suggests many jobs are more likely to have some tasks automated or assisted than to be fully replaced. This matches the video’s conclusion: AI does not simply eliminate all work; it redistributes components of work. The question is not “How much do you use AI?” It is whether you can decide what to delegate and what humans should judge. Simple summarization, drafting, format conversion, and repetitive replies are easy to delegate. Reading customer anxiety, judging field context, and cautiously confirming unspoken needs still remain heavily human. ## Five prompt principles for Korean users ### 1. Restore omitted subjects and objects Before writing “organize it,” state what should be organized, for whom, and for what purpose. Omission is natural in Korean conversation, but it becomes a blank for AI. ### 2. Turn negative statements into positive ones Say “Use a friendly but not exaggerated tone” rather than “Do not sound stiff.” Goals are more stable than bans. ### 3. Decide the output format first A table, list, paragraph, report, blog post, email, and presentation script are different outputs. Without a format, AI gives an average answer. ### 4. Separate context and criteria Give background as background, requirements as requirements, and verification criteria as verification criteria. If everything is mixed into one sentence, AI may miss priorities. ### 5. Do not try to finish in one turn Good AI use is closer to multi-turn collaboration than a single command. Receive a draft, strengthen criteria, revise again, and verify at the end. ## A prompt is ultimately a conversation habit, not just a technique UNESCO’s AI competency framework treats AI-era capability not as mere tool operation but as human-centered thinking, ethics, critical judgment, and practical use. Prompts are similar. They are not shortcuts to memorize. Talking with AI is a process of making your own thinking clearer. If you do not know what you want, AI does not know either. If you do not give criteria, AI produces an average. If you omit context, AI guesses. That is why the video’s core is deeper than “write better prompts.” Competitiveness in the AI era belongs not to people who know many technical tricks, but to people who can examine their own language and design context. In that sense, future AI literacy may be a language issue before it is a coding issue, especially for Korean users. The words we naturally omitted, the atmosphere we relied on, and the “do it appropriately” we handed over must all become sentences again in front of AI. ## Further reading – [Metacognition in the AI Era](https://www.thinknote.co.kr/metacognition-ai-thinking-checklist/) – [In the AI Agent Era, How Knowledge Workers Must Change](https://www.thinknote.co.kr/ai-agent-valuable-education/) – [Six Habits of People Who Get Smarter the More They Use AI](https://www.thinknote.co.kr/ai-smarter-use-six-habits/) ## References – [Ildangbaek video](https://youtu.be/OcmSp1Cn5rg?si=LQk_SQmJmDuqaETH) – [OpenAI, Prompt engineering](https://platform.openai.com/docs/guides/prompt-engineering) – [Anthropic, Prompt engineering overview](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview) – [Anthropic, Tracing the thoughts of a large language model](https://www.anthropic.com/research/tracing-thoughts-language-model) – [Brynjolfsson, Li, and Raymond, Generative AI at Work](https://www.nber.org/papers/w31161) – [International Labour Organization, Generative AI and Jobs](https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and-quality) – [UNESCO, AI competency framework for teachers](https://www.unesco.org/en/articles/ai-competency-framework-teachers) ## FAQ ### Are Korean prompts disadvantaged compared with English prompts? Not always. But Korean relies heavily on omission, particles, honorifics, and context, so AI often has to guess intention. When writing in Korean, it is better to state situation and criteria more clearly. ### Do I have to learn prompt engineering? Everyday users do not need grand engineering. But if you use AI for work, you need the basic habit of giving purpose, context, output format, and verification criteria. ### Why does “do not do this” often fail with AI? Negative statements place the prohibited object inside the context. Some models do not reliably reflect the ban. It is usually better to specify the desired behavior in positive terms. ### Can AI writing be made to feel human? To some degree. Adjusting sentence length, repeated expressions, omissions, rhythm, and concrete situations can reduce a mechanical feel. But AI does not possess real experience or judgment. ### What should humans do in the AI era? Humans must interpret context, set criteria, and make final judgments. Customer emotion, field situations, organizational tacit knowledge, and ethical judgment remain difficult to standardize fully. [Original Korean article](https://www.thinknote.co.kr/ai-korean-prompt-literacy/)
Image source: Captured images used in this article are stills from the original YouTube video. They are used for review, commentary, and educational explanation, and copyright remains with the original rights holders and the channel.

2026년 07월 10일
Seoul Learn Generative AI Service Support: Free Access for 1,000 High School and Older Students

The Seoul Metropolitan Government is launching **generative AI service support** for Seoul Learn members. The core is simple: among Seoul Learn students in high school or older, the first 1,000 selected participants can use paid generative AI services without separate subscription costs.

Official promotional image for Seoul Learn generative AI service support. Source: Seoul City and Seoul Learn guide materials, as shown in The Fact article image.

AI in learning has moved beyond simple search. It can summarize writing, guide problem solving, refine English sentences, and help with career-exploration questions. Paid AI services, however, can be a cost burden for students. This Seoul Learn program focuses on lowering that barrier.

## What is Seoul Learn generative AI service support?

**Seoul Learn generative AI service support** helps students participating in Seoul Learn use the latest AI services for study. According to the promotional image, selected participants can use a total of nine paid generative AI services, including ChatGPT, Claude, Gemini, and Perplexity, for free.

Rather than simply handing out AI accounts, it is closer to an education-support program that reduces learning gaps and broadens AI-use experience. High-school-and-older students can use AI for assignments, admissions preparation, personal-statement drafts, career exploration, and foreign-language learning.

## Recruitment and support period

| Category | Details |
|—|—|
| Recruitment period | June 9 to June 26 |
| Support period | June 2026 to February 2027 |
| Number recruited | First 1,000 participants |
| Target | Seoul Learn participating students in high school or older |
| Application method | Online application on the Seoul Learn website or participation through the pre-diagnosis QR code |

Because the support period runs until February 2027, selected students can use AI services for a relatively long period, not just a short-term semester trial.

## Who can apply?

The target is 1,000 Seoul Learn participating students in high school or older. The important condition is Seoul Learn membership. This should be understood as an education-support program for Seoul Learn participants rather than a public event open to everyone.

The image also says, “Check eligibility requirements and application on the Seoul Learn website.” Before applying, confirm participation eligibility, grade criteria, selection method, and pre-diagnosis conditions on the Seoul Learn site.

## Which AI services can be used for free?

The guide mentions nine paid generative AI services, including ChatGPT, Claude, Gemini, and Perplexity. Each has different strengths.

– ChatGPT: useful for writing, summarizing, organizing problem-solving approaches, and expanding ideas.
– Claude: strong for reading long text, summarizing materials, and refining sentences.
– Gemini: useful for information exploration and document support connected with the Google ecosystem.
– Perplexity: a search-style AI useful for research and checking sources.

For students, the important question is not “Which AI is best?” but “Which AI should I use for this assignment or learning situation?” For example, use explanatory AI for first understanding a concept and search-style AI for source research.

## How to apply

The image describes two routes.

– **Online application on the Seoul Learn website.** The image shows the address http://slearn.seoul.go.kr.
– **Participation through the pre-diagnosis QR code.** The guide says final selection follows review of the AI Ollie guide and participation in a pre-competency diagnosis.

This means the process may not be just pressing an application button. AI-use guide review and pre-diagnosis may be part of selection, so read the application page through the end.

## How should students use it?

If selected, use AI as a learning coach rather than an answer generator.

For a math problem, instead of immediately asking for the answer, ask, “What concept should I check first in this problem?” For English writing, ask, “Please make this sentence more natural and explain why you revised it that way.” That creates more learning value.

Recommended uses include:

– Have an unfamiliar concept explained at a level a middle-school student can understand.
– Summarize long passages or textbook content into key sentences.
– Receive English-sentence corrections and check the reasons.
– Create lists of questions about careers or majors.
– Structure presentation materials for performance assessments.
– Check sources and evidence during research.

Submitting AI-generated answers as-is is risky. The answer may contain errors or may not match school assignment rules. Use AI responses as drafts and references; final judgment and expression should be the student’s own.

## Checklist before applying

– Am I a Seoul Learn participating student?
– Am I included in the high-school-or-older criterion?
– Am I applying during June 9–June 26?
– Is a pre-diagnosis QR or AI-use guide procedure required?
– Which AI services can I use after selection, and until when?
– Do I know the rules for using AI in school assignments or test preparation?

Because recruitment is described as first-come for 1,000 students, interested students should check quickly within the application period.

## Conclusion

Seoul Learn generative AI service support is an opportunity for students to experience the latest AI tools without cost burden. The important point is not only free access, but the experience of properly connecting AI to learning.

AI is not a tool that studies instead of you. But if you ask good questions, it can be a strong assistant for concept understanding, writing, research, and career exploration. Seoul Learn students in high school or older should check the recruitment period and conditions.

## FAQ

### Who is the Seoul Learn generative AI service support for?

According to the image, it targets Seoul Learn participating students in high school or older, with 1,000 participants recruited.

### When is the recruitment period?

The recruitment period is shown as June 9 to June 26.

### Which AI services can be used for free?

The guide says participants can use a total of nine paid generative AI services, including ChatGPT, Claude, Gemini, and Perplexity.

### Where do students apply?

The guide indicates online application on the Seoul Learn website or the pre-diagnosis QR code included in the image.

### Are students selected immediately after applying?

The image says final selection follows review of the AI Ollie guide and participation in a pre-competency diagnosis. Final conditions must be checked on the Seoul Learn website.

## References

Image source confirmation: [The Fact, Seoul Learn provides free ChatGPT and Claude access to 1,000 members](https://news.tf.co.kr/read/life/2330694.htm)

– [Seoul Learn website](https://slearn.seoul.go.kr)
– Seoul Learn generative AI service support guide image

## Further reading

– [In the AI Era, What You Need to Learn Before Prompts Is Your Own Language](https://www.thinknote.co.kr/ai-korean-prompt-literacy/)
– [Metacognition in the AI Era: How to Check Your Thinking](https://www.thinknote.co.kr/metacognition-ai-thinking-checklist/)
– [In the AI Agent Era, How Knowledge Workers Must Change](https://www.thinknote.co.kr/ai-agent-valuable-education/)

[Original Korean article](https://www.thinknote.co.kr/seoul-learn-generative-ai-service-2026/)

2026년 07월 10일
Anthropic Mythos Shock: As AI Becomes a Strategic Asset, What Should Korea Prepare?

Anthropic’s “Mythos” issue is hard to treat as simple news about a new AI model. The colder point is that frontier AI models are now both cloud services and strategic assets.

Like semiconductor equipment or advanced GPUs, access to models itself is becoming a subject of diplomacy and security. Korea cannot dismiss this as another country’s regulatory news.

## What is the core of the Mythos issue?

The Mythos issue signals that access and control, beyond AI performance competition, have become matters of national strategy.

Anthropic introduced Claude Mythos 5 as a model strong in cybersecurity and biological research. Through Project Glasswing, it planned to use the model to find and defend against vulnerabilities in critical software.

According to the official explanation, early partners used Mythos Preview to find more than 10,000 high-risk or critical vulnerabilities in important software. For defensive purposes, that is a highly attractive result.

The problem is that the same capability can be used offensively. A model that quickly finds vulnerabilities is a weapon for defenders, but if control collapses, it can also become a weapon for attackers.

That is why Mythos was provided only to limited partners from the beginning. When the U.S. government then directed a suspension of foreign access to Fable 5 and Mythos 5, the issue moved from technology news to national-strategy news.

## Why did people say “AI is also a strategic asset”?

The U.S. directive showed that access to cutting-edge AI models can become a national-security judgment. The logic of semiconductor export controls has moved toward the model itself.

A key shift is underway. In the past, computing resources, chips, and equipment were the bottlenecks. Going forward, model weights, API access, the ability to remove safety measures, and data-retention conditions may also become targets of control.

For companies, this is more complicated. A model available yesterday may suddenly be blocked today. In high-risk sectors such as public administration, finance, healthcare, defense, and R&D, this becomes an operational risk, not just inconvenience.

## Three risks Korea must watch

Korea’s AI strategy must examine foreign model dependence, the dual-use nature of security AI, and the realism of sovereign AI together.

### 1. Dependence on foreign models

Korean companies and public institutions have quickly adopted global AI models. From a productivity perspective, that is natural. But if core work becomes deeply tied to a specific foreign model, supply interruption or access restriction can become work stoppage.

Areas connected to national functions—administration, defense, security, healthcare, energy, and finance—need separate standards. This does not mean using only domestic AI. It means areas that cannot be interrupted need fallback routes.

### 2. The dual-use nature of security models

Powerful security AI improves defensive capability, but without control it can also be converted into offensive capability.

The hardest question Mythos raises is: “If powerful security AI is widely released, does the world become safer or more dangerous?”

Vulnerability-detection AI is a major advantage for defense teams. But if verification, disclosure, and patching cannot keep up, vulnerability lists may simply pile up faster. Anthropic also explained that after vulnerabilities are found, verification, disclosure, and patching become bottlenecks.

Korea should not focus only on detection models when building AI security capability. Coordinated vulnerability disclosure, patch responsibility, supply-chain response, and incident-response training must be designed together.

### 3. The practicality of sovereign AI

Sovereign AI must not end as a slogan. It is not only about making one Korean-language model. Public data governance, domestic computing infrastructure, high-risk AI evaluation, industry standards, and procurement systems must be connected.

Korea is preparing systems and infrastructure such as the AI Basic Act, the National AI Committee, the AI Safety Institute, and the National AI Computing Center. The direction is right, but the Mythos issue demands more speed.

## Korea’s future strategy: before “securing a model,” build a controllable AI system

The core is not owning a specific model but having an AI operating system that can be stopped, changed, and verified when necessary.

Korea’s response should not end with “let’s build our own frontier model.” The more important question is: in which areas, at what level of control, and at what cost should control rights be secured?

### First, classify AI dependence in national core areas

Public institutions and critical industries should classify the AI services they use by work importance. A simple document-writing tool and a cyber, medical, or administrative decision-support tool cannot be judged by the same standard.

Core areas need at least three conditions: an alternative model, inference paths inside Korea or a trusted region, and manual operating procedures for failure.

### Second, make Korea’s AI safety evaluation system practical

AI safety evaluation must not end with document review. In areas with real potential harm—cybersecurity, biology, financial fraud, disinformation, and privacy leakage—red-team evaluation and repeated testing are needed.

For high-performance models, there must be steps between “ban use” and “open without limits.” Limited partner access, log retention, high-risk query routing, independent evaluation, and incident reporting must move together.

### Third, the National AI Computing Center must become strategic infrastructure

The government is pursuing a National AI Computing Center worth up to two trillion won. This infrastructure should not simply rent GPUs; it should connect Korean models, safety evaluation, and public-sector AI demonstrations.

Accessibility is crucial. If only large companies can use the infrastructure, resilience across the whole industry will not increase. Universities, startups, security research organizations, and public agencies must be able to use it in practice.

### Fourth, cooperate internationally but assume a blocking scenario

Korea cannot build every AI technology alone. Cooperation with the United States, Europe, Japan, Singapore, and others remains necessary. But cooperation is not the same as dependence.

Contracts should include clauses on data location, model-access suspension, emergency patching, switching to alternative models, and audit rights. Public procurement should evaluate not only “the best-performing model” but also “the model that can operate during a crisis.”

## What should companies and individuals check?

Companies should inventory the AI tools they currently use: which work is connected to which model, where data is stored, and whether replacement is possible within days if the service stops.

Individuals can look at it more simply. The ability to use AI well matters. But trusting the answer of one specific model as-is is risky. In the AI era, having one’s own language and judgment criteria comes before prompts.

Related Thinknote articles worth reading include [In the AI Era, What You Need to Learn Before Prompts Is Your Own Language](https://www.thinknote.co.kr/ai-korean-prompt-literacy/) and [Metacognition in the AI Era: How to Check Your Thinking](https://www.thinknote.co.kr/metacognition-ai-thinking-checklist/). For AI-agent trends, also see [AI Agent Evolution](https://www.thinknote.co.kr/ai-agent-evolution-openclaw-action-oriented-ai/) and [AI-Native Workflows](https://www.thinknote.co.kr/ai-native-workflows-digital-brain-ai-agents/).

## Conclusion: Korea’s AI strategy must prepare for the politics of access

The message of the Mythos issue is clear. AI competition will not be only performance competition. It will also be competition over who can access models, who can adjust safety measures, and who can maintain service during failure.

Korea should use global models, but in core areas it needs controllable alternatives. Sovereign AI is not isolation; it is insurance. That insurance works only when models, data, compute, safety evaluation, and procurement move together.

## FAQ

### Can ordinary users use Anthropic Mythos?

No. Anthropic described Mythos 5 as a restricted-access model strong in cybersecurity and biological research. Fable 5 was intended for more general knowledge work with safety measures, but access was also suspended after the U.S. government directive.

### Does the Mythos issue immediately affect Korean companies?

Not all companies are affected immediately. But it is a warning signal for companies that rely on overseas frontier AI models for core work. They should check access rights, data location, alternative models, and failure-response plans.

### Does sovereign AI mean not using foreign AI?

No. Sovereign AI means securing control and choice in necessary areas. Global AI can be used, but public, security, and industrial core areas need replaceability and domestic operating capability.

### What is the Korean government already preparing?

The AI Basic Act, the National AI Committee, the AI Safety Institute, and the National AI Computing Center are being prepared. The computing center is especially important infrastructure for domestic AI research and industrial use.

### What should individuals prepare?

Do not depend on only one model. Verify important judgments through multiple sources and practice explaining AI answers again in your own words.

## References

– [Anthropic, Claude Mythos](https://www.anthropic.com/claude/mythos)
– [Anthropic, Statement on the U.S. government directive to suspend access to Fable 5 and Mythos 5](https://www.anthropic.com/news/fable-mythos-access)
– [Anthropic, Claude Fable 5 and Claude Mythos 5](https://www.anthropic.com/news/claude-fable-5-mythos-5)
– [Anthropic, Project Glasswing](https://www.anthropic.com/glasswing)
– [Anthropic, Project Glasswing: An initial update](https://www.anthropic.com/research/glasswing-initial-update)
– [Korea Policy Briefing, National AI Computing Center](https://www.korea.kr/news/policyNewsView.do?newsId=148938942)
– [MSIT, AI Basic Act passed at the National Assembly](https://www.msit.go.kr/eng/bbs/view.do?sCode=eng&mId=4&mPid=2&bbsSeqNo=42&nttSeqNo=1071)
– [NoCutNews, the warning from the U.S. Anthropic block](https://www.nocutnews.co.kr/news/6533333)

[Original Korean article](https://www.thinknote.co.kr/anthropic-mythos-ai-strategic-asset-korea/)

2026년 07월 10일
15 Frontend Basics Every Vibe-Coding Beginner Should Know
# 15 Frontend Basics Every Vibe-Coding Beginner Should Know
Once you start vibe coding, unfamiliar terms arrive quickly: React, Next.js, API, CSR, SSR, NPM, build, bundling. At first, it feels like you must memorize everything.
But the important part is not memorizing terms. It is understanding why each technology appeared. Then you can see where AI-generated code runs and what to ask it to fix.
This article summarizes beginner-friendly frontend concepts based on a Vibe Coding University video for non-developer vibe coders.

When you start vibe coding, you often meet words like React. The point is flow, not memorization.
## 1. The internet and the web are different
The internet is the global communication network connecting computers. The web is a service on top of that network for exchanging documents and moving through links. A browser is the app used to view the web.
- Internet: the road connecting computers
- Web: documents and screens exchanged on that road
- Browser: the app used to view the web
## 2. The web began with documents and links
The early web created by Tim Berners-Lee connected research documents through hyperlinks. Three foundations remain important today: HTML for structure, URL for addresses, and HTTP for the agreement between browser and server.
## 3. HTML is the skeleton, CSS is the clothing, JavaScript is the behavior
HTML decides what exists on screen: headings, paragraphs, images, buttons, inputs. CSS controls appearance: color, size, position, spacing, fonts. JavaScript creates behavior: menus open, input is checked, cart quantities change.
## 4. The browser turns code into pictures
Frontend code is material for the browser. The browser reads HTML and CSS, builds structure, calculates size and position, and paints pixels on screen. This is rendering. Recalculating layout is reflow; repainting pixels is repaint.
Knowing this lets you ask AI more precisely: “change the button color” differs from “fix the layout shift when the button is clicked.”
## 5. jQuery and React solved different problems
Frontend thinking moved from DOM manipulation to state and components.

jQuery made DOM manipulation easier and more consistent across browsers. Later, web apps became more complex: login state, shopping carts, notifications, and real-time data. React handles complexity through state and components. When state changes, the screen is redrawn to match it; reusable UI pieces become components.
## 6. Node.js and NPM turned frontend into an ecosystem
JavaScript originally ran only in browsers. Node.js allowed it to run on servers and in development tools. NPM installs and manages JavaScript packages.
- Node.js: an environment for running JavaScript outside the browser
- NPM: a warehouse for installing code packages
- package.json: the list of packages the project depends on
## 7. Build means preparing development code for deployment
Build prepares developer-friendly code in a form users can load quickly.

Modern frontend projects contain many JavaScript, TypeScript, CSS, image, font, and library files. Build processes organize that work into code users can load quickly.
- Transpiling: converting newer syntax for broader browser support
- Bundling: grouping many files appropriately
- Tree shaking: removing unused code
- Optimization: reducing file size and improving load speed
## 8. MPA and SPA change pages differently
SPA is a major pattern that makes websites feel like apps.

Traditional sites fetch new HTML from the server whenever pages change. This is an MPA, or Multi Page Application. An SPA first loads an app shell and then changes screens by exchanging only necessary data.
- MPA: simpler structure and easier for search engines
- SPA: smoother experience, but initial loading and SEO can be harder
## 9. CSR, SSR, and hydration are about who draws first
CSR means the browser receives JavaScript and draws the screen. SSR means the server creates HTML first and sends it to the browser. Hydration attaches JavaScript behavior to HTML the server already rendered.

CSR and SSR differ in where the screen is first rendered.

Frameworks such as Next.js are useful because they mix CSR, SSR, and static generation depending on the page.
## 10. An API is the agreement between frontend and backend
The frontend handles what users see. The backend handles data storage, authentication, payment, and permissions. An API is the agreement for exchanging data between them: which address, which format, which response.
## Core checklist for vibe-coding beginners
- Does this technology handle structure, design, behavior, or data?
- Is the problem in the browser or server?
- Is the issue slow screen rendering, missing data, or tangled state?
- Do you need feature work, build-error fixing, or deployment optimization?
- Did you tell the AI the problem location and expected result?
With these distinctions, “it doesn’t work” becomes “React state changes but the screen does not update; please find the cause.”
## Related reading
## FAQ ### Do I have to learn frontend to vibe code? You do not need to memorize every syntax detail, but basic concepts such as HTML, CSS, JavaScript, APIs, and rendering help you ask AI better questions. ### Are React and Next.js the same? No. React is a UI library; Next.js is a broader framework built on React with routing, rendering, and deployment structure. ### Why do CSR and SSR matter? They affect initial loading speed, search visibility, and user experience. ### What should I check first when a build error occurs? Separate package installation, syntax conversion, type checking, and bundling issues. Give the AI the error message and command you ran. ### Which frontend concepts should beginners learn first? Roles of HTML/CSS/JavaScript, browser rendering, state, APIs, build, and the difference between CSR and SSR. ## References
Frontend can look like a field full of terms to memorize, but it is really the history of the web solving problems: documents gained design, then interaction, then app-like behavior, and later concerns about search and speed.
Vibe-coding beginners need this flow. Once you see it, you can read AI-generated code better, ask more specific questions, and modify projects more safely.

Original Korean article

Image source: Captured images used in this article are stills from the original YouTube video. They are used for review, commentary, and educational explanation, and copyright remains with the original rights holders and the channel.
2026년 07월 10일
The Essence of AI Coding Is Not the Model but the Harness: Matt Pocock’s Agentic Engineering
# The Essence of AI Coding Is Not the Model but the Harness: Matt Pocock’s Agentic Engineering
Tech Bridge thumbnail about Matt Pocock’s agentic engineering workflow

When people discuss AI coding, they usually name models first: Claude, Codex, Gemini CLI. Models matter, but Matt Pocock points elsewhere. The real difference comes from the harness.
The harness is the working environment around the model: prompts, skills, codebase structure, tests, documentation, sandboxes, GitHub Actions, and review flow. It is like judging not only the engine of a car but also the chassis, pit crew, and track operations.
This matters because model performance is hard for us to control, while the harness is something we can design.
## AI has eaten tactical programming
Pocock borrows John Ousterhout’s distinction between tactical and strategic programming. Tactical work includes writing code, fixing bugs, making commits, and matching syntax. Strategic work is deciding what structure will be maintainable, how to divide work, and where the codebase should go.
AI has already absorbed much tactical programming: small features, tests, refactoring drafts, and documentation updates. As AI handles more tactics, human value moves toward strategy: clear goals, narrow scope, completion criteria, and tests.
## The work environment matters more than the latest model
The strongest line is that everyone obsesses over models, but we should care more about the harness. For example, if you want to reduce token cost, one answer is shorter prompts. Matt’s answer is a codebase that is easy to change. Clear structure, tests, and current documentation let AI work with less context. A tangled codebase makes even expensive models wander.
## Skills should be managed as procedures, not piled on
A skill is a reusable bundle of instructions for repeated thinking or work. It can make an AI act as a learning coach, critique a design aggressively, or review PRs in a specific way.
But Matt does not say to add as many skills as possible. He suggests deleting skills, plugins, MCP servers, Claude.md, and agents.md, then observing the model in a blank state. Add back only what is truly needed. Too many instructions can pollute the context window.
## AFK agents are closer to queues than infinite loops
“Agentic loop” sounds attractive: the agent thinks, acts, observes, and acts again. In practice, it can blur scope, raise cost, and remove review points.
Matt proposes “queue rather than loop.” Put work into a queue, like GitHub issues or Jira tickets. The agent takes one task, investigates, changes, tests, makes a PR, and a human reviews it. Good candidates include failing-test investigation, README updates, refactoring proposals, PR review drafts, security checklists, and old-issue reproduction.
## AX: Agent Experience now needs design
Developer Experience made environments easy for humans to install, run, test, and deploy. Agent Experience is the degree to which an AI agent can work in the codebase.
- Predictable folder structure
- Clear test commands
- Automated type checks and linting
- Current README and development docs
- Clear module boundaries
- Safe validation of small changes
- Enough information to run in a sandbox
Good AX overlaps strongly with good DX. What humans can infer informally, agents often miss; therefore documentation, tests, commands, and boundaries matter more.
## Problems found by AI should become system improvements
If a model finds a security bug, do not stop at “this model is good.” Ask why the bug remained, why tests missed it, whether similar bugs exist, and how future checks can be automated.
AI output should be a signal to improve the harness: add tests, revise review criteria, create a security-check skill, or add CI checks. That turns AI coding from a one-off productivity tool into an organizational learning system.
## Product and business judgment still belongs to humans
The video also touches on SaaS and AI startups. Matt’s answer is simple: talk to customers, find real problems, prototype, and validate. AI accelerates implementation, but it does not decide what to build, why to build it, or what to remove.
## Seven things Korean developers and teams can do now
1. Read your README as if an agent were entering the repo for the first time.
2. Turn repeated requests into skills or templates, but only when they are truly repeated.
3. Split issues into AI-sized tasks.
4. Include test commands and done criteria in instructions.
5. Start AFK work inside sandboxes and limited permissions.
6. Review AI PRs for failure patterns, not only code.
7. Follow model news, but check structure, tests, documentation, and review flow more often.
## Related reading
## FAQ ### What is a harness in AI coding? The full environment in which the model works: prompts, skills, code structure, tests, docs, sandbox, CI, and review flow. ### Why is codebase structure more important than the latest model? A clear structure and tests let AI make safe changes with less context. A messy codebase makes any model struggle. ### How is a queue different from an agentic loop? A queue lets humans define tasks and review results one by one. A loop can become open-ended and harder to control. ### What is AX? Agent Experience: how easy it is for AI agents to work in a repository. It overlaps with DX but raises the standard for clarity. ### What should be prepared first before assigning coding to AI? Scope, completion criteria, tests, sandboxing, permission limits, and review flow. ## References
AI coding’s next step may not be turning on more tools. It may be pausing to inspect the environment where AI works: missing docs, fragile tests, and tasks that can be queued. That small cleanup can matter more than subscribing to one more model.

Original Korean article

Image source: Captured images used in this article are stills from the original YouTube video. They are used for review, commentary, and educational explanation, and copyright remains with the original rights holders and the channel.
2026년 07월 10일