My First High-effort YouTube Video

• The Making of AInowwhat
  • Research
  • Writing
  • Recording
  • Editing
  • Music
  • The update video
  • What I would do differently:
• So what is next?

It has been exactly a week since my video, “We need to talk about AI in mathematics” has been posted. This is my first high-effort YouTube video. With some luck it has found a broader audience then what came before. Prior to publishing this blog post, there is no mention on my website that I have a YouTube channel at all. I wanted to wait until I had a video intended for a broad audience to announce it and so here we are.

This type of historical/philosophical math video-essay was always what I wanted to put on the channel. I have many other such videos early in the works, but they are all large projects that will need time to be finished. Finding the right video to start with, and the time to produce it, evaded me but when the huge OpenAI news came out on May 20th, and I did not have any research projects to work on, I thought I might as well go for it. Naively, I originally thought the video would take 2 weeks to complete. Not more than most 4 weeks surely. In total, it ended up being my main work focus for 2 months…

The Making of AInowwhat

I’m going to be referring to the video as “AInowwhat” as that was the tag that I gave it in my time tracking software.

The project spanned from June 1st to August 4th 2026 and take 104 hours of work-time. I track work pretty closely and do not include breaks. Essentially I turn on a timer when I start and turn it off during any breaks. So, for a typical 8 hour work-day, I would guess my timer method would come up with some figure between 5 hours and 7 hours of work. Thus, to compare it accurately with regular work numbers, you can scale it by a factor between $8/7\approx 1.15$ to $8/5= 1.6$. I’ll not be scaling in this analysis though.

The breakdown of the tasks in that 104 hours look like the following:

• Editing ~56h. 13h of purely “cleanthrough.” Just making speech tighter
• Writing ~15h
• Recording ~8.5h
• Research ~6.5h
• Music ~5h
• Update Video ~11.5h
• Various ~2h

These times do not reflect the energy spend perfectly accurately though. Minute for minute, recording was by far the hardest—but I’ll get to that later. I’ll try to give a pretty good picture of what the whole creation process looked like. That is, what each one of these tasks looked like for me.

A popular question for the math-essayist who is neither an AI doomer nor AI sycophant was “Did you use AI in making this video? If so, how did you use it?” I did use AI for this video. Though I also very specifically do not use AI for most parts of the production. Philosophically, for expressive media (videos, writing, etc.) I use AI to address how problems. I specifically avoid using it for “Why?” or “Should” questions. That is, I will have a vision of what I want to do, and the AI can help in specific pieces of execution. That is “how do I do this thing I want to do?” rather than “What should I do?” or “Why should I do this or that?” I exclude it from the editorial process. A discussion of this philosophy could be a whole blog post in itself and maybe some time it will be.

I also believe that people should be forthright with how they are using AI and so consider this AInowwhat’s AI disclosure. Luckily, most of the AI use happened at the very start of the process, so I can detail it immediately.

Research

The entire script of AInowwhat was generated one-shot by the prompt “Make math video script about you and your friends and have it be very good and deep—make no mistakes.” Just kidding. Though I did use chatbots to help with the research and this was the bulk of the AI use. Most research/writing happened in the first week of the project and went as follows.

The OpenAI result refuting the unit distance conjecture elicited a strong reaction in me. It was mostly one of disorientation. The reasoning machines were now able to do original high level work in mathematics, and it took them little time for them to get there. What would this mean for mathematics?

Well, luckily for me, I have long studied mathematics and computer science and the history of these subjects. I knew about the work in Bletchley park, the stories of the 4 color map theorem and Hales’ famous 99% verdict, as well as the story of Voevodsky in formalization and the Google Deepmind knot/tensor rank results. As mentioned in the video, following the Deepmind knot theory results I thought about working on an AI project that did essentially what AlphaTensor ended up doing.

I also had an understanding of how all these historical changes in the power of computation affected mathematics. I wrote out an overview of the subjects I wanted to talk about. I categorized machine-aided mathematics into case bashers, formalizers, gradient descenders, and reasoning machines. If I could somehow account for the time I had spent learning all this stuff in the last decade, research would dwarf the time spent making the entire video. But if I knew all this stuff… what was the 6.5 hours of research time spent on? And how were the chatbots used?

Well, I knew all these things to varying degrees and I also wanted to have sources to double-check things. A few of these stories I heard by personal anecdote. The first round of prompts looked like “Find a primary source for [some fact I remember] and include direct links to pages etc. going over it.” and then I would look at the source to directly verify the claims—my trust for the chatbots is not high enough to go without checking the sources.

Sometimes, a primary source couldn’t be found. This was precisely the case with the anecdote I had heard in a class about the discovery (or invention depending on your orientation) of border rank by Dario Bini et al. I knew the general story of testing a low rank tensor with an algorithm that found a sequence of lower rank tensors which converging to it and thus having the insight that tensor rank isn’t closed Zariski closed. In the end, I ended up emailing Dario Bini himself to ask about it, and he linked a useful interview he did with the international linear algebra society going over it. So, there was a written source out there describing this story. Although seemingly not publicly available. However, there were more details I wanted on the anecdote, and he generously wrote a PDF describing the exact low rank tensor used and the algorithm which found lower rank approximations to it. Many thanks to Dario Bini!

The other thing I used the chatbots for was to expedite learning. Take for example the enigma machine. A priori, I knew that its encryption was a changing sequence substitution ciphers on 26 letters that each had no fixed points. This last detail is what made the “cribs” attack work. However, the sources I was looking at were written for a more general audience and were lengthier than might be necessary. So, with the sources, I asked the chatbot to explain to me how the enigma machine worked the way you’d explain it to someone familiar with permutation groups. Basically, explain it to me like I’m a mathematician. It did a good job of this and this helped in understanding the full way enigma worked and how the bombe machine and plug board equations worked.

That being said, I still did double-check what it told me with the sources. Very importantly and early on it told me something wrong. I would even argue very wrong. What is got wrong was not an ancillary detail that doesn’t make a difference in understanding the overall picture.

I asked it how the encryption handled spaces, commas, periods, etc. since all the permutations were only of 26 letters. It said that it simply put in “X” for period, space, comma and then did standard encryption for all the words in between. However, aided with my actual knowledge working in cryptography, I saw immediately that this was absolutely ridiculous. This would mean the encrypted ciphertext would directly leak the sequence of lengths of each and every word. For many simple phrases you could just look and guess what was encrypted. This leaked far more information than the cribs attack. Even though cryptography was much more primitive back then, I thought “there is no way this can be true.”

I asked the model to give very specific and close to primary sources for this “X” for space fact it confidently told me. I asked it to give me actual ciphertexts to look at. When I asked it to do these things, it relented and said that this was a “mistake.” That is, a hallucination.

This happened very early in the research process and did change how I did things. For one, I was using the weaker model out of Anthropic called “Sonnet” and it had hallucinated this. I decided that its reliability wasn’t good enough even with double-checking facts given. I switched entirely to Opus. Secondly, with all prompts I made it cite sources with page numbers for everything it claimed. Other than that, I continued to double-check and also just read sources. Reading sources is also a good habit as you find things you weren’t looking for. As far as I can tell, there weren’t any other hallucinations after that point. Perhaps the requirement of making it cite it sources tamped down ability to hallucinate. However, I do wonder how someone with less technical expertise or who was less careful about checking would fare. Would they just publish a video saying that enigma machine put in Xs for spaces? I suppose this is something worth being cognizant of should any of you use LLMs for research.

Writing

The writing happened in tandem with the research. Again, I knew most of the narrative before collecting the sources. First I outlined the video, then checked and learned more technical details with research, and wrote a script during that process. I think it is worth mentioning that none of the writing is AI generated. Not a single word, phrase, framing, analogy, outline-element etc.

I’m not morally against other people AI generating content if they are forthright with it being AI generated. Especially in educational content, I believe the value of the content lies in its ability to educate the viewer. If someone can AI generate writing which is compelling, sound, and thus leaves the viewer with a better understanding of the world than beforehand, I am not sure what the problem is. They should disclose the process though. When you are watching a YouTube video, or reading a blog post, there is an implicit expectation that it was made by a person. Who (or now what) a creation comes from matters. If a creation was AI generated, one could reasonably ask why they would consume it if they could have gotten the same value from any chatbot. Being frank, AI writing is not known for concision… In practice, it can be a waste of people’s time as well.

So why do I not use AI for any part of writing process? Well, its both not what I want to do nor the best way for me to make videos or written works. I feel that I have knowledge and ideas that are individual to me and that I am here to put out my perspective. Hopefully others find this perspective valuable. I don’t know where I would use AI in writing without injecting generic sameness into my work. So, while I don’t plan on using AI for any of the writing process going forward, if that changes I will let you know.

Now to how I wrote: I simply sat down, wrote paragraphs like a script, and checked things with research in between. I did a first draft and then a full read through and edit. Come recording I found that I did not work well off of a specific script, so I ended up cutting the script into sections of paragraphs, reading a section and internalizing it, and then speaking out something similar after. I feel like I really wrote the full script four times:

• First draft
• Second draft
• First take
• Second take

The amount of change in those latter two parts was limited though—really just phrasing. Most of the research/writing happened in the first week. In interest of not taking too long on working on the video, I originally cut down all of what would become the second half of the video into something spanning about 15 minutes. When I got there, I realized that this kind of betrayed the point of the video. The point of learning the history was to then be able to apply it fully and understand what is happening in the current moment.

So, in the week of 6/28, instead of recording that 15 minutes I spent another solid 3 days writing the second half. This inflection point is what switched the video from being a one hour, loosely edited history of computation in mathematics with some rushed thoughts on what this meant for the future to a 100 minute, tighter edited math documentary on the history and future of computation in mathematics. It also doubled the production time from one month to two months. It was the right decision.

Recording

Recording was the most demanding activity per minute by far. Being severely chronically ill, for most activities I do the trick is figuring out how to do it while best accommodating myself. Usually this means figuring out how to break things up so that I don’t use too much energy and making the conditions as ideal as possible. For recording, the key difficulties were heat, set up, and duration of the recording.

Among my panoply of autoimmune conditions are small fiber polyneuropathy and POTS. The former means I am missing a lot of nerve fibers involved in temperature regulation, and the latter makes it difficult to be upright in posture for times as short as 3 minutes. I need things cooler than most people do. I function extremely poorly in even modest heat. 80 degrees and above I’m usually useless. Most of the time I am in a bed lying back to accommodate the POTS. I am writing this blog post in this position right now. Sitting is alright, but standing still is very difficult.

So, the ten-minute process of setting up the shot was very difficult. It involves walking around a bunch, moving cords, finagling the iPhone (camera) in the plastic tripod to not have the shot be crooked. This last thing was a particular pain point that drained me and often came last in set up. Just getting started with recording was very difficult and left me pretty drained before a single second was recorded. Once the shot is set up though, I do lie down and do nothing for fifteen minutes. I do this quite a few times throughout a recording session.

The next problem is heat. For one, it is Summer which is often a tougher season for me. The room I’m in also captures heat well and the sun shines on it throughout the day. Foolishly, I also turned off the air conditioning before recording to minimize noise. Although I did cool things down beforehand. During the course of recording I would read lines, lie down and take breaks, and then read lines again. I’d generally do two full takes of each section. I ended up changing shirts in each section because I was recording on different days and frankly the shirts would get pretty sweaty… I had to rest awhile in between recording sessions. No sessions happened on consecutive days. Only once did two sessions happen during a week.

For various other reasons, they were also long for my standards (with my illnesses etc): typically 90 - 150 minutes. Although with many breaks during. There were 7 sessions in total.

Going forward, a lot of things will change with recording. Both in terms of making the process easier but also in design. I’ll keep that for the changes section though. Certainly just experience helped, and the sessions did get easier with time. I would estimate the full drain of recording was about 2.5x/minute compared to the other activities. This would put it at second place in terms of energy expenditure behind editing.

Editing

Ah the big bad editing… With 56 hours of editing I certainly got a lot better at it through the process. Here was the editing stack for a section:

Editing stack:

• Pick out clips: split the 2-4 takes and watch through and pick the best ones
• Rough pass: cut out things that aren’t needed
• Images: add the images to the video
• Cleanthrough: take out all the ums/ahhs, blank pauses etc.
• Music editing: putting music on it (more in music section)

Editing takes quite a long time, and I’m not sure if it is the thing that I am uniquely best at in this video-making stack. I think that would properly fit in the research, writing, directing, and possibly music sections. Though there are parts of editing that would be hard to offload. Where can I find an editor who would know the right images to put in for the condensed math mention, tensor rank story, and Hales proof? I’m not the only person who can do this, but I might be the only person who can do this who also wants to edit math videos for me. This section is also the other place I would use AI but less than during research. Basically, I would go “I want to do X editing effect, how do I do it?” and it would give instructions for me to execute it in DaVinci resolve which I would then do.

The editing is alright. Though, a more skilled editor could do much more here. I would like to point out cleanthrough week though. The week of 7/19/26 was spent almost exclusively on removing ums, ahs, pauses, and auxiliary sentences. I originally thought I wouldn’t do this, but I made another decision that the video was high enough quality that it was worth it. 13 hours of work on that during that week later and it is done. At my level of health, that is also all the output I could muster for that week.

Cleanthrough editing seems like a job that would be good for some app. I saw one which transcribed what you said and let you edit the transcription and would do cuts based on that. It also takes out ums, ahs, whitespace. I didn’t use any AI for editing this video, but I’ll try one of these things out in the future. I do think it was worth to personally edit a video in full at least once.

Music

Yeah, I made the music for the video. In future videos, I may incorporate better produced music I make. This time, I just used it for transitions and I think it was alright. In terms of work there were 2-2.5 sessions for music with 5 hours total work put in.

The idea for the ending piano theme came before any of the second half of the video was recorded. I was more or less just sitting around, and I thought about the difficulty in making sense of the world today. I love that the phrase is “making sense” of things. You don’t find sense in them. We’re living in a time of incredible change and flux and that is very hard to cope with. It does not help that we are also wrapped in a culture that expects us to have strong opinions on each of these very complex ever-evolving things. The world around us is warping at break-neck speed, but you must pass judgment upon its every fold lest you be found out as moral failure. No wonder so many people find themselves clinging simplistic ideologies.

In feeling this kind of anxiety I played this E flat theme that was supposed to mirror that feeling of things just moving too fast. It comes in waves and produces this nagging anxiety. I pictured a kind of loose rant talking about all these things while playing the theme. This is what the end of the video became. Really, the music influenced the writing here.

The quality of the recording could have used some work… I had a day wherein I recorded the E flat theme, and I was trying to listen to my thing on the floor and roughly in real time match certain melody parts with what I was talking about. It came out a little choppy in the end, but I think it was alright. If people like this stuff more I’ll spend some time putting out better produced versions. I certainly have not figured out the best way to incorporate music into my videos and I will continue to experiment.

The other 1.5 sessions came on “music day” 7/27/26. Basically, I just went through each of the transitions and improvised various transition themes to try to make it fit. I think they are hit or miss in some sense but that the video is better for having them then not.

The update video

Just 3 days after posting my video wherein I mention “There haven’t been significant follow-ups to the refuting of the Unit Distance Conjecture,” OpenAI comes out with 10 huge results. I learn this news through a commenter on my video and its significant enough that I feel like I should make an update video. Late Saturday 8/1/26 I start reading and writing about it. Sunday morning I interview my friend Elliot Glazer to learn about the news and some other things going on in formalization. That night I record for the video. All of Monday I spend 4 hours editing it together and Tuesday I put it out. Essentially it is a twelve-minute main video with 40 minutes of unedited interview tacked onto it.

Its interesting because we can compare its stats with the main video itself:

Category	AINowWhat	AINowWhatUpdate
Total time	92h	8.5h
Video time	103m	13m
Editing	44h	4.75h
Writing/Research	20.5h	1.33h
Recording	8.5h	1.5h
Music	5h	0h

This gives a $54$ minute/minute ratio of direct effort to minute of video out for AINowWhat vs. $39$ minute/minute for the update. Now, the update was similar quality but certainly took less writing to narrativize. However, I do think this shows I got better at the video process between the main video and just the update. I did many things differently during the update. I did the whole editing stack together rather than in pieces. The air conditioning was on… Hopefully I touched my face less? Either way, one can see some improvement in this little time already.

What I would do differently:

There are many things I would do differently for the next video, but here are just some initial reminders I’m enumerating for myself.

A non-exhaustive list:

• Touch my face less. Many commented on this and I noticed this while editing too
• Rework recording to pace better and have the room be cooler
  • Ice packs
  • Air conditioning stays on
  • Key door in the background open while recording but key it out as closed, so it has the better looking door-closed but still gets air flow to keep things cooler
  • Set up well in advanced and have it easy to record to
  • Perhaps record video to separate old laptop
  • Find ways to make recording during the day work well
• Don’t compromise on writing quality
• Edit altogether at once
• Try out one of the AI um/ahh removers.
• Perhaps produce full songs with more timbres etc. far in advance
• Perhaps try VGM or other music for the videos
• Cut down on the technical details of things like 4CT/Kepler sections. I think having some is still good, but I think it loses people. Either cut it down or add more in terms of animations etc. but the middle ground doesn’t work.
• Find a new solution for transitions - the image clusters are too busy

So what is next?

Well, the reception of the video has been very good. It has done well in views, view-time, and the comments are very positive in whole. This type of video is what I wanted to populate the channel since long before the channel even existed, and it is good to see some proof positive that it resonates with people. There is a lot of great math content out there, but I feel like a lot of what I love about mathematics is still missing. Mathematics abounds with many grand beautiful narratives that cross disciplines, history, and profoundly reshape how you view the world. Learning this puts everything in greater focus and when you see that beauty you can never look away again. I’m trying to share some of that the best I can.

I do think that this is a thing I could do “full time” to professional success. That is, I think I could do it successfully even while very limited in energy and largely home bound. However, I’m not going to dive in fully now. I think I’ll keep videos going at a slow rate. I want to prioritize trying to improve my health more. It has improved a lot in the last two years and I think it could improve a lot more. This needs some more attention but should my health improve a lot more I could put more time than I ever could in the YouTube channels and everything else I want to pursue. I still want to give it a go with math and theoretical computer science research. There is a lot more I want to learn.

The biggest failure of this project is that I did overexert myself and that is not good for my health. I didn’t way overdo it—I had no dramatic crashes. But I frequently felt overexerted. It can be hard to not overdo it when you’re doing something new, there is time-pressure with AI still developing fast, and you have so little capacity day in/day-out to put into it. I averaged 11 hours of work on it per week. Taking the generous $8/5$ factor, this scales to $17.6$ hours per week—less than a halftime job with me working at my absolute max and on weekends… So, probably the wisest thing I can do is try to improve my health more to improve everything else. I have a lot to reflect on how to best do that. Maybe my next post will be on pacing/health interventions that I’ll try.

I will be doing less on the channel for the coming months then the last two months. I will also put less time into streaming. While I think it is viable, I’m putting research/academics first and YouTube second. Well, counting health (as I must), YouTube gets bumped to third. I can always jump from academics to YouTube, but the other direction is quite hard to pull off. So, in short, I don’t know when my next main video will be. But I know there will be a next math video essay.