Reflections from 2025 so far

Writing in the past has helped me get more clear on things and learn about myself. I figured it's probably something worth doing in 2025, and I plan to do this once a week.

There are two main sections to this long post, FYI:

• Personal Stuff - reflecting on things I've learned over the last year and a half
• Technical Stuff - me talking about some technical stuff I did this week

Personal Stuff

I've been floating around trying to figure out how to make money outside of traditional employment for over a year now. I've attempted lots of things:

• Dropshipping (lol) - I actually made some money with this though which is crazy.
  • One quick learning from this now that I am where I am (Jan 26 2025 vs ~November 2023 when I did dropshipping) is that it was easier for me to sell something I didn't make. I found that I quickly lost faith in SaaS products that I built, yet I could somehow sell crappy little necklaces. It's worth me trying to understand this because it could help unblock me. I think it comes down lack of confidence in myself and the things I build because I don't have a solid vision for any of the SaaS products I made (although now, I actually think I do have a solid vision).
• Proper E-commerce - I tried upgrading to doing more professional ecomm, but the amount of investment up front, time to grow, and inability to quickly pivot the product made me decide against it.
  • Creating a product isn't easy. You need data to know what to make, designs, fabrication, testing, then manufacturing, then storing it. If you want to make any adjustments, it goes through the same process all over again.
• Building an audience to try and funnel to Skool.com - I figured "if I can build an audience around AI, maybe I can start funneling people over to Skool.com and make money that way."
  • Learning from this: building an audience around AI is tough. Everyone is trying to do this, so if you want to stand out you need to focus on something SUPER specific. And become the best at that thing. But be careful what you specialize in.
    • One thing I'm noting about AI is that it's shifting the value of generalists and specialists. If you're a specialist in the wrong thing, AI can essentially make you obsolete overnight with one minor advancement. Generalists who know how to leverage AI are increasing in value compared to specialists. But the most valuable position still is (and probably will be) deep specialists in areas that are useful as AI advances. And if you combine deep specialization + some generalist ability, you're golden.
  • Other learning about skool.com - the people who are succeeding are partnering with larger influencers and are calling themselves "growth operators". This is something that just doesn't interest me. So proper growth of an audience and Skool community will take a long time.
• Software as a Service (SaaS apps) - I've made several at this point, learning something from each one.
  • I made an app to help youtube creators improve their title, thumbnail, and hook by rating them and giving feedback on their potential effectiveness and cohesion. The idea is that good cohesion between title + thumbnail + hook gets someone to watch, and a really good hook increases the odds that someone watches your video longer, which is a positive signal for the YouTube algorithm which means it might promote it more.
    • The lesson? YouTube is HARD. I spent days tweaking the prompts to give lower scores and proper feedback for videos that performed poorly, and gave higher scores to videos that performed well. But I kept finding exceptions where my app (and my intuition) said that a video should have performed poorly, but did REALLY well.
    • A secondary lesson - this isn't something that anyone actually cares about (even though they should). Bad (i.e. new) YouTubers have a hundred other things they're worried about. Good (i.e. experienced) YouTubers usually have their own formula for title + thumbnail + hook. I could have spent more time and effort trying to reach people "in the middle", but it would take a lot of time and effort and me putting all my attention on learning YouTube. That's a lot of platform risk, and I figured AI advancements from someone like OpenAI might eventually include video analysis and someone could just use ChatGPT to do this for free. Not worth me spending years on this idea.
    • Summary: I made an app that could not do what it was designed to do, and even if it could, it didn't seem like anybody was interested in the problem it solved. And even if I could get people interested in it, it probably wasn't a problem worth spending years on (or at least, that was my thinking).
  • My next app was a landing page analyzer, specifically for SaaS products. I figured I could empathize with them, and I found studies about what works and doesn't work on SaaS landing pages. I actually saw a need for this (lots of people on Reddit and other sites ask for feedback on their landing page) and I found examples of people making money providing feedback: both manually and through a SaaS app.
    • The lesson? I didn't do enough research and know exactly who this was for. People on Reddit aren't serious founders, so they just want quick and free feedback. People who were paying for things like manual reviews were seeking out trusted sources who demonstrated experience or success in understanding what makes a good landing page. And that wasn't me.
      • I could have tried focusing deeply on this for a few years, and maybe pivoted it into building and launching landing pages for people, building features like Dynamic Text Replacement and other similar things, etc. But again, I didn't think this was a problem worth focusing on for the next few years because I could see this all becoming obsolete overnight from a single AI advancement. And AI might change the way we use the internet and browse.
  • The third app I created was probably the least grounded but the most fun. I noticed lots of indie developers making apps wanted to do market research. And I noticed apps like ChatGPT and Claude were terrible at searching the web. Perplexity was good, but it was limited by what you could publicly find online. So I used Vercel's ai-sdk to build a chatbot that could basically connect and use DataForSEO's API to do actual market research for you.
    • The lesson? The previous apps I had some idea who they would be for and a very limited set of features and use cases I supported. This was a general purpose app and I had no idea what direction to take it in. I thought I could maybe focus it on SEO, but I started finding people who made "SEO agents" that could supposedly write, manage, and grow a blog for you on autopilot to drive traffic to your site.
      • I think it had more potential than the previous apps, and SEO is a huge industry that will probably still be relevant (but go through a transformation) for years. But I just lost steam and at this point started feeling lost. I didn't know what to do with it, had no vision for it, and I eventually just lost steam (after ~1 month of it being launched).
      • I think I somehow (again) found some excuse for how AI could disrupt this and make it obsolete again. But I think I just realized I was chasing weird problems that weren't backed by any solid desires or demand by customers. I wasn't solving burning problems for people. And worse, I wasn't really learning or building any skills that would meaningfully help me down the road. I was building things and staying surface-level. (I'll expand on this in a bit in the Possible Next Steps section below)
  • Finally, after chatting with AI for a bit, I think I figured out some important things. (Also discussed below in more detail in the My Concerns About AI and Money section.) One thing regarding products though is that I should focus on something that's evergreen and that will only get more valuable as time moves on and AI continues to advance: some kind of knowledge product around my thinking and building process.
    • I'm a generalist who has built a ton of stuff and used AI in all kinds of fun ways. I broke down my process for building a few of my recent personal tools and projects, and it actually found patterns in what I do to build things quickly. So maybe I could create a course around my process for building stuff, like how @levelsio on X has his The Indie Maker Blueprint book. I could keep it updated as time goes on with new examples of how I break down problems and build solutions.
      • This isn't a "How to make $10k/mo building AI apps", or "How to build SaaS apps". I wanted it to be focused on something real that I can actually speak to, which is my personal build process. That's something I'm "successful" at.
    • I went through and actually documented everything. I'll move it from Notion and share it in a blog post. But what was the lesson / problem with this? I thought "how could I charge anything for this, when Pieter only charges like $30 for his book with ALL the experience has building things? So for now, I'm just going to keep it as a free guide in case people are interested.

So what now? After all this, I felt a bit lost which is weird to admit at this point in life. It's January 2025 and I'm 35 years old and I feel lost? WTF happened to me? I thought I'd have things sort of figured out by now.

Upon more reflection, I realized some hard truths from all of this:

• I have low self confidence; and despite trying to work on developing a better mindset, it's still not very good. I have lots of limiting beliefs that are deeply rooted.
• I stopped pursuing harder problems to work on at some point in my life and career, because 1) I didn't realize the importance of that and I'm lazy, and 2) I had no context about what hard problems are actually worth solving (but now I think I do).
• I haven't been stacking evidence to boost self confidence, or a portfolio of my skills anywhere to show that I'm making any kind of progress.
• I've been building generalist skills, but without any specialty. Above I mentioned that I think the most value combination is being a specialist in "good" dimension that likely won't be completely upended or made useless by AI, + being a generalist who can build just about anything.
• I care way too much about what strangers on the internet (and some people close to me) think about me.
• Making money outside of employment is difficult. But maybe not because that task on its own is difficult. But because we don't have the right skill stack and mindset to see the opportunities all around us.
  • I could have taken a few freelance gigs that paid pretty well, but I still sort of consider that normal employment (i.e. getting paid for my time). What I want is a complete separation of time and money and the freedom to work on what I want. That way, I can control the leverage my work has to potentially earn LOTS of money in a relatively short period of time.

Possible Next Steps

Based on everything above, I think I finally went around in enough circles to decide on what the actual next steps for me should be. Running in circles is maddening, and sometimes you realize you're doing it but you're not sure why. Or when and how to stop.

My personal advice is to keep reflecting, and eventually you'll run in enough circles that you've seen all the tricks your mind plays on you. At that point, you can start to make decisions more grounded in truth.

1. I need to optimize my life to find harder problems to work on. And ideally, find ONE hard problem and start working on it NOW.
   1. And this hard problem needs to be something that I'm reasonably certain will not be made obsolete overnight by a single advancement in AI and end up being a waste of time. Whatever skills and knowledge I gain should be useful in other domains.
   2. I actually think I know what I want to focus on too: AI video processing
2. I need to share publicly what I create and learn in the pursuit of solving this hard problem. I see this having 2 positive outcomes:
   1. This allows you to build a portfolio of work you can show off. Whether you want to work a typical job for the rest of your life or work for yourself, you need some way to prove to people that you have the skills and knowledge that you say. This will help you land clients, convince customers, land jobs, etc.
   2. This will help you build a personal brand online which I think is becoming increasingly important. It will inevitably help you build a network you can tap into for product launches, feedback, friends, and job opportunities.
3. Work hard on the right things to continue learning about this problem area, and build products (ideally related this hard problem) that can help me accomplish my goal of decoupling time and money.

And I just need to be ready for all the usual tricks my mind plays on me to get me to stray from this path. I've realized a few things now that I think will help me stick with this long term:

• I've built up a really good generalist skill set. But I want and need to focus on something bigger. I don't want to go another 5-10 years and realize that I've had no real focus or direction. I know the power of compounding as it applies to knowledge and skills (not just money), and I want to take advantage of that while time is still on my side.
• I don't currently have a strong network I could tap into to talk about product ideas, hire from, brainstorm with, find good job opportunities (if needed), etc. This is a problem. Something I'll talk about later is that I'm concerned about a post-AGI (Artificial General Intelligence) / post-ASI (Artificial Super Intelligence) economy and world. If making money as a solopreneur becomes difficult, then I want to make sure I have the skills, the portfolio, and the network to attract an opportunity to team up with people who have figured out how to ride this wave.
  • I.e. - your chances of survival are usually better in a group than if you're on your own.
• I don't have any kind of portfolio showing valuable results or things that I've made. And I'm not going to go very far without having some way to showcase to the world what kind of value and results I provide.
  • After spending a lot more time on X, it becomes increasingly obvious who is actually doing useful work and who isn't. The people who are doing useful work share LOTS of value in their posts, usually for free. And the insights are things that only come from actually doing the work. There's little to no fluff, and they don't gatekeep their knowledge because they're making good money doing whatever it is they're talking about. I want to be like this.
• I see where my life is now, and realize all the things I've been avoiding that can get me to where I want to go. And I'm now keenly aware that unless I make these changes, I'm screwed and it won't lead anywhere good.
  • The scary part about this is that I've been watching all the great content from people like Alex Hormozi where they tell you "It's tough. You have to work SUPER hard. Like, 10x harder whatever you're thinking." But it wasn't really getting through my head for some reason. And until I had this recent epiphany about what I need to do next, I just felt so defeated and lost because I didn't really know where I was going wrong.
  • *All things considered, my life is amazing and I am super grateful for it: my wife, family, friends, and all the people (and dogs) in it, my good health, etc. But I have a much larger vision for my life and what I want to provide for my family. And right now I'm failing miserably in that regard.

My Concerns About AI and Money

This post is getting super long. Way longer than intended. But I mentioned above that I had a chat with AI where I learned a lot about myself and my thoughts around AI and money.

I realized I'm really concerned about the post-AGI/ASI economy and world. AGI is Artificial General Intelligence, and ASI is Artificial Super Intelligence.

You can think of AGI as an AI with the same general level of intelligence as a human that can perform tasks on its own like an employee or assistant. (I think we're pretty much already here, but it's not something that exists everywhere yet.)

You can think of ASI as an AI with superhuman intelligence. Think about someone who has read the entire internet, studied every possible thing there is to study, and who understands all that information, can recall it, and can use it to solve seemingly impossible problems, AND can improve itself and its knowledge faster than a human can. That's roughly how I think of ASI.

Eventually, we'll get to the point where AI will be able to do everything a human can do, but better. Especially when we can give AI something closer to a human brain that functions like ours, put it in a robot, and it can move around the world freely.

I have NO idea what things will be like at that point. But along the way, there's the potential for a lot of disruption and inequality which is a recipe for some kind of revolution. As AI continues to get better, companies will likely do what they're designed to do: cut costs and increase profits to increase their chances of survival.

Companies will lay more and more people off, new ones won't need to hire nearly as many people, and most people won't want or know how to make money outside of traditional employment.

It'll create a divide between people in society who still have valuable skills to offer in exchange for money, and a growing majority of people who won't be able to make that trade. And interestingly enough, I think it'll be knowledge workers that suffer first. I say "interestingly enough" because I think many people assumed knowledge work like programming would be automated last. But it'll be among the first.

[Quick aside: I take ballroom dance lessons with my wife and absolutely love it - the studio, the instructors, and the friends we make there. I can honestly see something like dance studios sticking around and being able to do well in a post-AGI economy. There's too much technological advancement that has to happen to get to the point that I would ever even consider paying a robot to teach me how to dance. Also, dance might be in its category because it's a uniquely human thing and involves emotion and improvisation. Something that can probably never be programmed into a robot.]

My point is there's going to be a divide in society where the vast majority of people probably won't be able to exchange skills for money, soooo what do we do then?

My thought process for all of this is that first and foremost, I want to protect myself from being someone who falls into the category of someone who can't exchange skills for money. I want to be able to provide for my family, and if I can do that then I'd like to find ways to help my close friends, and then ideally even my community. But if I can't help myself, then I have no hope of helping anyone else.

So to increase my chances of being able to succeed post-AGI/ASI, I need to follow the 3 steps I have laid out above. If I do it all right, then

1. I'll have apps or products that are earning me income faster than typical employment ever could (and they'd also be assets that I could sell), and
2. My products + research will make for an impressive portfolio to show that I'm an "A Player" which can help me land clients or a job if I need to look for one
3. If the landscape changes so much that SaaS dies and my products stop making money, I'll have a network of people who know what I can do. And I can tap that work to potentially find opportunities to work

So overall, I want to pursue that plan with the intention of building up the necessary things to survive in a post-AGI/ASI economy and take care of myself and my family. Then like I said, I would love to have the means to help out close friends, and even my community.

I guess you could call that my vision.

Personal Section Wrapup

That was WAY longer than I initially thought I would write. So I'm done with that for now. Now I want to talk a bit about the technical stuff I did this week.

Technical Stuff

I'm losing the steam and desire to keep writing today with all the other things I need to do. But I'll capture what I can.

I decided to focus on AI video processing. There isn't a single good tool out there that does this. Not even major LLMs (to my knowledge) support this. I think Gemini I can do it through AI Studio or whatever it's called, but I saw a demo of it and looked terrible and was super slow.

So to start, instead of getting really fancy by looking at research papers, and trying to find specialized models, etc, I just tried to break down the problem using current tools and my limited knowledge. So my initial thought process was something like this:

• Upload video file
• Check if it has audio using ffmpeg and extract the audio stream if it's there
• Have a process that takes 1 FPS as an image and save it somewhere
  • While this is happening, use AssemblyAI's API to transcribe the audio (if there was any audio)
• Create batches of 10 images at a time and put them into an LLM message using the OpenAI API format (i.e. have the LLM process 10 seconds of video at a time)
  • Also take sentences from the transcript for that ~10 second window and send it along with the images so the LLM has full context about what was going on in the video during that time.
• Then use OpenAI's API to send 10+ batches of images+transcript with some user input context of what to look for in each chunk of video
  • e.g. 10 batches of 10 images would mean we could have OpenAI process 100 images (or 100 seconds of the video + audio) in parallel for us
    • And we could theoretically go a lot higher. Apparently OpenAI allows up to 5k requests per minute (for gpt-4o) with their API if I read this table right
• Take the output from the LLM for each batch and make one final LLM call to piece everything together to make sense of it

That was my starting point. Then I started looking at pricing for image processing. I found out that there are like fixed size canvases or something that they use to process images. So one of these canvases (if that's even the right word for them) is 512x512. So if you have a 512x512 image, using the model gpt-4o-2024-08-06, it costs ~$0.000638 per image. But if you have an image that is 513x512, it bumps up the price because it needs another canvas or whatever to process that extra pixel.

[Note: I found it surprising the gpt-4o-mini costs more for image processing than gpt-4o-2024-08-06. Not sure why that is.]

So what I decided to do as a cost optimization is to take each frame extracted from the video and fit it into a 512x512 box and ensure the resulting image is 512x512, and the screenshot we took fits in there and isn't distorted at all.

That seemed to work well enough. There were a few times AI didn't seem to be correctly parse text that was on screen, and when I looked at the images it was obvious that the quality was a bit too low at that point for super accurate text extraction. So that's something to be improved.

But then to speed it up and reduce costs, I had another idea. I checked the price for a 1024x1024 image, and it was actually ~$0.001913. So I could take 4 512x512 images and put them in a 2x2 grid that makes a 1024x1024 image. And this saves 0.000638 * 4 = 0.002552 - 0.001913 = 0.000639. So we pretty much 4 images processed for the price of 3. Pretty good!

Not to mention that it significantly speeds up process now, because a batch of 10 images has 4o images (which is 40 seconds) worth of video. This required changing up my prompts and code to correctly bundle images + transcript, but it did work to speed up processing and reduce costs.

BUT, I think one good exercise for me to do would be to try and come up with some way to benchmark performance when AI is analyzing 1 image at a time vs this 2x2 grid of images, and maybe even test other things like how much performance degrades when take a 1080p video and downsizing frames to fit in a 512x512 box. I want to see what the tradeoff is in price, speed, and processing quality. There are probably lots of use cases where it might be worth paying more to have the LLM analyze higher quality large images. And even to compare LLM image processing performance.

I initially wanted to build a general purpose video analyzer that would look like a normal chat interface, let you upload the video along with a request for what to look for in the video, then store analysis data as vector embeddings in a database to make it easier to recall and chat about the video. But this was too much to start with. For now I created a Streamlit app on Modal that lets you upload an educational video, and it takes notes for you to help maximize educational value without overwhelming you.

A few other thoughts and notes before I wrap this up:

• I created this video processing engine to also be able to increase the FPS it analyzes, because presumably there are times where we may want to watch shorter videos with very high FPS (think about analyzing marketing content or youtube videos to look at camera cuts, transitions, facial expressions, etc; or even analyzing a sports clip like a clip from a boxing match or something)
• I can further speed up processing by asynchronously sending off the transcript to AssemblyAI while we extract frames from the video. Those could easily be done in parallel
• I want to do more research into efforts in this space to improve the speed at which videos can be processed, and it needs to be more cost effective. Imagine being able to upload a 60 minute 500MB video and having it "processed" in like 60 seconds and then being able to ask AI anything about it. Or have it do something useful with that. That'd be pretty neat depending on the use case.
• I can think of several micro-SaaS apps that could use this technology at their core, but I can also think of BIG custom applications that could benefit from AI vision. Think about an NBA team that wants to take ALL of their previous games from the past 3-5 years or something, feed it into AI, and have AI be able to understand which players on each frame are part of the team, and to extract things like plays they ran, missed opportunities, mistakes, stats, etc. That'd be a huge effort, but could be immensely useful.
  • Think about it applied to something like golfing to help people perfect their golf swing. Or a boxer to help them understand an opponent's tells better. OR having it analyze poker players to find tells that they might have. Lots of applications here.
• The next step beyond this is realtime video processing. But that's a more complicated problem.
• Also, a cool future state for this would be allowing people to just send a link to the app where the video is, and having some kind of video watching agent that can watch it (ideally sped up) or that knows how to extract video data quickly from different sources. That way the user doesn't need to worry about download a video to then upload to my service.

Technical Section Wrapup

I think that's about all I had in my brain for today. That's more writing than I've done in the last several months combined probably.