How Correcting Automatic Captions Became a Whole Project
I had canceled Monday’s broadcast because I was not feeling well. I still did not have the energy to present the next installment of “Liberation and Civilization,” so I decided to take questions and, if there were none, show what had consumed me over the previous few days. It began with a long-standing and very simple problem: the poor quality of YouTube’s automatic captions after a livestream.
Correcting the captions of a one- or two-hour broadcast by hand took me nearly four hours. I did it because accurate text could later become the basis of a book or article and could also be translated into English. Even if artificial intelligence could not do the job well yet, I believed it eventually would, so I would have to keep preparing the captions until just before that point arrived.
After ChatGPT 5.6 appeared, the situation changed. I gave it YouTube’s automatic captions and asked it to perform the corrections I had been doing myself. To my surprise, it did a reasonably good job. As I tested the possibility, the plan suddenly expanded. A caption-correction task grew into a website that could provide the contents of the entire Barobul Buddhism channel as text.
If the spoken content of a video survives as accurate text, it does not end as a caption file. It can live again as writing, books, translations, and searchable knowledge.
What barobul.org, Built in Four Days, Does
The result is barobul.org. Some errors and design details remain, but the framework and principal functions are largely complete. In the past, this would have required several people working for more than a month and would have cost at least tens of millions of won. Working alone, I might have needed two or three months without any guarantee of success. This time, I built the basic structure after concentrating on it for roughly four days.
The site does more than import YouTube videos. AI corrects errors in the automatically recognized captions, then uses those corrected captions to produce a description and detailed chapters. It turns what I meant to convey into a long-form article and translates that article into English. Material that once could be found only inside a video becomes searchable Korean and English text.
Rewatching a video requires spending the full one or two hours again. Text can be scanned quickly and used to revisit the main arguments. When a viewer once asked whether my talks were available as a book, I said I would write one, but I stopped because continuing to produce broadcasts seemed more important. Now I am reconsidering whether an offline book must come first. Accumulating a carefully prepared article for each video may be a more direct answer.
Text is not meant to replace video. It is another entrance through which people can revisit a long video, check its ideas, search it, translate it, and carry it farther.
Three Sources That Helped AI Understand Me
To correct automatic captions properly, AI must know the speaker’s recurring vocabulary, interests, proper names, and way of thinking. To write from the corrected captions, it must also understand how I develop an argument and what I am trying to say. I therefore gave it three bodies of material as context.
The first was Artificial Intelligence and Illusion, the book I wrote with my brother. Its roughly 200,000 Korean characters discuss not only how AI works but also what AI reveals about human knowledge and illusion. The second was my blog, “Raindrops of Awakening,” which includes Buddhist writing, criticism of the Jogye Order, discussions of the Atthakavagga, and personal essays about my father’s death and myself. The third was the Barobul Buddhism YouTube channel, including hundreds of video titles and the structure of its series.
From these materials, AI compressed the interests, recurring arguments, frequently mentioned people, and editorial voice of Yoon Deok-man and the channel into a JSON profile. A list of contemporary Korean names also helps it repair proper nouns that speech recognition often mangles. The essential task is not merely smoothing sentences; it is understanding who spoke, in what context, and with what intention.
Giving AI a pile of source material is not enough. It needs context for the speaker’s interests, standards of judgment, intensity of language, and degree of certainty.
Korean Articles, an English Edition, and a Mobile-Friendly Site
The site does not stop with Korean. Each completed Korean article is translated faithfully so English-speaking readers can reach it. I hoped that even a foreign reader who arrived by chance through a search could read how I understand Buddhism and where I think humanity should go in the age of AI. The responsive layout also adjusts naturally to a phone’s screen and font size.
During the broadcast I read from a recent article, “Is Compassion Alone Enough? Why Must Buddhism Speak of Justice?” If personal compassion means sharing and taking less for oneself, then at the social level we must discuss structures and justice. If systems, the state, families, and communities are broken, telling individuals merely to be strong does not remove their suffering. We must be able to ask together what justice is and challenge one another’s claims. I did not sanitize the broadcast’s harsh criticism that religion must not remain a servant licking the asses of those in power.
The resulting prose does not have all the dramatic force of an essay I would revise repeatedly. A livestream is not itself a polished manuscript. Yet the article organizes arguments scattered through the video and conveys the intended point with considerable clarity. That is the value the site can offer at this stage.
If strong criticism is made polite simply because it is being converted into prose, the speaker’s actual concern disappears with it.
AI Changed the Images and the Automation Strategy Too
AI also generated the site’s backgrounds and other images. In the past I had to take photographs, buy stock images, or search for copyright-free material. One reason I once took so many photographs was that websites needed realistic, cinematic backgrounds and visual objects. Now I can discuss what atmosphere would fit, generate alternatives, and select one.
I initially aimed for complete automation from video collection to publication. Using the OpenAI API, however, was far too expensive. A more efficient design may be possible, but for now I settled on a semi-automated process in which I converse with AI, inspect the output, and approve updates. The important question is not whether it can be called fully automatic, but whether the structure can keep accumulating one video at a time at a manageable cost and quality.
Only a small number of articles are currently public, but hundreds of channel videos can be added in sequence. Content trapped inside YouTube can enter web search and reach English readers. Speech and prose feel different, so viewers can encounter the same material through another medium.
The Shock of Building a Website Without Writing the Code Myself
This experience was especially startling because I had worked in web development for years. I knew backend systems, databases, HTML, CSS, frontend effects, design, and server work. That background helped me understand AI’s answers and issue useful instructions. Yet I wrote almost none of the code myself this time. Whether the task involved PHP, a database, HTML, or CSS, I described and directed it.
The site contained many requirements, from visible effects such as twinkling stars and an occasional meteor to the internal structure connecting videos and articles. In the past I would have designed and implemented each component and hunted down its errors. This time AI took on those tasks. A person with no technical background may still struggle to state the right requirements, but within months or a year it may be enough to express a clear purpose and direction.
I ran a web agency for many years. Just as a sports agent negotiates contracts and advertising on behalf of an athlete, a web agency handles planning, design, development, and operation for someone who wants an online store but lacks the technical expertise. AI agents are now entering the position of that intermediary. A user may only need to explain the intended result; the agent can propose what must be prepared, how hosting and payment should work, and then carry out the implementation.
What began to disappear was not a few lines of coding. The very structure of web-agency work—translating a goal into technology—started moving to AI agents.
How Much of the Expert’s Skill Will Become Unnecessary?
For now, my experience and judgment influenced the result. I could evaluate an efficient database structure, a fast interface, and an attractive design. But once AI has more design knowledge and cases than I do, it can make those judgments better as well. Even now, it is overwhelming in speed and efficiency.
Doing in four days what once took two months is not merely a small productivity improvement. Correcting speech-recognition errors through context, writing coherent prose, translating it naturally, and generating the necessary images were impossible with older software. AI made previously impossible tasks possible while also replacing much of the work I had spent years learning.
One step farther, the user will not need the capabilities of a web agency at all. The user can point to another example and ask, “How can I do that? Can you build it for me?” An AI agent can explain the preparations and execute the job. This is wonderfully convenient, but it is also the scene in which the skills of my profession rapidly become unnecessary. That is why the change is both magnificent and frightening.
Pulling months of work into four days felt less like a productivity gain than a time machine. The problem is that the machine also races past existing jobs and skills.
The Channel Profile Summarized as “Do Not Believe; Question”
In the latter part of the broadcast, I opened the channel profile that AI had created. After reading Artificial Intelligence and Illusion, my blog, hundreds of videos, and the long articles already on the site, it extracted a core message: “Do not believe; question.” It described the mission as exposing the guesses and illusions humans take for facts or truth through questioning, observation, comparison, and verification, and restoring Buddhism from a religion of transcendence and belief to a living process in which individuals and society repeatedly escape delusion. I was amazed that it compressed my work into those few lines.
The epistemology section said that human perception, memory, language, intuition, and logic may be highly probable guesses rather than reality itself. Intuition is useful, but when wrong it becomes a premature assumption and illusion. Scriptures, science, religion, teachers, experts, my own words, and AI responses can all be wrong. Questions and wisdom begin by admitting not only that we do not know, but that we may not even know what we do not know. Knowing only one perspective makes it difficult to recognize how thoroughly one has been colored by it, so we must compare other perspectives and counterexamples.
The Buddhist framework also matched the claims I have repeated. Buddhism is not a religion demanding belief but a process of escaping belief, superstition, fixed ideas, and attachment to views. Liberation is not becoming divine by transcending rebirth, aging, sickness, and death. It is repeatedly shedding worldly and transcendental desire, language and concepts, and recurring illusions. Buddhas and arahants are not deified as omniscient beings, and awakening is not possession of a final answer but a change of direction that makes continued self-examination and correction possible.
No scripture is established as an infallible authority. The Kalama Sutta, the Atthakavagga and Uraga Sutta of the Sutta Nipata, the Heart Sutra, and the Diamond Sutra are important subjects, but I read them critically through dependent origination, impermanence, non-self, observable experience, and the full context of a text. The Buddha’s “undeclared questions” are not an excuse to evade inquiry; they embody the refusal to pretend to know what one does not know.
Becoming a Buddha does not mean possessing the answer. It means acquiring a way of thinking that examines what I know and do not know and remains able to correct error.
A Buddhism That Sees Individuals and Society Together, and the Site’s Two Purposes
The social-position section said that thoughts and actions arise within family, education, history, economics, politics, media, and institutions, so we should examine the web of conditions rather than blaming the individual alone. It does not oppose compassion to justice. Stopping violence and oppression, protecting victims, and changing structures are also compassion. It criticizes religion when karma, the afterlife, and obedience are used to legitimize inequality and power. Criticism of the Jogye Order and monks must address statements, conduct, interpretation, power structures, and social consequences rather than attack status alone.
Freedom of expression is the right to resist power that suppresses questions and evidence. It is not the same as insulting disabled people, calling the May 18 massacre “Tank Day,” or irresponsibly justifying hatred and mockery of slaughter. Criticism should open the possibility of reflection, communication, reform, and change rather than seek destruction alone. Religion should do more than briefly comfort individuals; it should help people see reality and build more just relationships.
The editorial-voice section said that I use direct and sometimes rough language toward claims and powers I judge wrong, and that this intensity should not be arbitrarily softened in editing. At the same time, the stronger the criticism, the more carefully the actual statement, conduct, document, and context should be shown so readers have grounds for judgment. Expressions such as “This is how I see it,” “I do not know the answer,” and “I am making a guess” should preserve the degree of certainty. The fact that AI extracted these traits from my work was itself one of the results I wanted to show.
I ended by describing two purposes of the site. First, Korean viewers can revisit the substance of a long video as refined prose without watching the entire recording again, and people who find video difficult gain another route of access. Second, my arguments about what Buddhism is and where humanity should go in the age of AI can be searched and read in English.
For four days I slept very little and poured almost every waking hour into this work. I have not yet converted every video, nor is the process fully automatic. Yet something I thought would become possible only a year or two from now can already be done with a little added effort. This broadcast was less a finished philosophical lecture than a tour of that transformation: a record of how far AI has come and how its abilities can help our words and questions endure.
The purpose of barobul.org is not to duplicate videos. It is to open a path for rereading them in Korean, to address questions to English readers, and to accumulate a continuing dialogue grounded in “Do not believe; question.”
This article was edited by AI solely from the supplied video captions and description. No external sources were used.