Data Backup
Optimizing n8n workflows these past few days, the most obvious realization is that AI is an idiot.
It can be very smart one moment and very idiotic the next, so every time you optimize a workflow, be sure to back it up in advance to prevent AI idiocy from wiping out days of workflow optimization.
Last night, I tinkered until 4 AM, trying GPT-OSS-20B. It only ran at 2.7 tokens/s on my computer, which is incredibly slow (my qwen3.6 35ba3b runs at 6 tokens/s). And after changing the model, no matter how I optimized or reused prompts, new errors kept popping up; the more I tried, the more errors I encountered.
So today, I’d rather restore the previously backed-up workflow than continue to struggle with the idiotic workflow I spent all night on.
At the same time, I also tried using Codex for deep optimization. However, the most terrifying thing was that halfway through, just as it was about to finish, Codex ran out of credits, leaving the workflow in a crippled state.
So, in summary, AI is an idiot. Be careful with AI token limits; don’t run out of credits while working. Finally, you can’t do without AI either. These workflows can be completed and optimized by AI in just an hour, but if I were to research them manually, it might take a week or a month, and I might not even be able to write the corresponding workflows.
Even if AI is an idiot, as long as you back up promptly and provide patient prompts, saving a month’s worth of time is enough.
Game Background
When attempting to automate video editing, video cutting is the biggest headache.
Going out for breakfast, visiting night markets, or eating fast food are all extremely time-consuming activities (because you have to go out). Eating or cooking at home is also super time-consuming.
Without smart glasses, shooting videos is also a super troublesome task. So, initially, I chose to shoot at home. Excluding food, that leaves only my computer and my workstation.
My workstation has no graphical interface, so if I were to shoot it, I’d just be staring blankly at its casing. Therefore, I chose my computer, and specifically, games.
Because I’ve shot video tutorials before, screen recording doesn’t look good and lacks a sense of real life. It’s best to film actual hardware and the visuals displayed by the hardware, such as games.
Currently, only games can provide beautiful scenery, characters, and various interactions. This is the legendary ninth art.
I installed CS 1.6. This was the first game I played on a computer, at an internet cafe, almost twenty years ago. Why choose it? One reason is nostalgia, and the other is that my computer doesn’t have a graphics card, so I can only play older games, like arcade emulators.
So, for my latest videos, the first is Minecraft, a game, and the second is CS 1.6. I also installed Meteor Butterfly Sword, but its language was always garbled, so I eventually gave up. I also have Plants vs. Zombies on my computer (and I played it all night once; now it’s just for passing the time).
Headless Mining Card
Due to AI, graphics cards have gone up in price, memory has gone up in price, everything has gone up. I don’t have the funds to buy a good graphics card right now.
However, automation highly depends on graphics cards, especially free and open-source local large models, like the latest Qwen 3.8 27B. I’ve used a P106-100, 6GB, before. Although it was poor, it could play large AAA games, deploy some open-source AIs with low VRAM, like Flux image generation, and also provide GPU acceleration, based on my previous experience.
Both the P106-100 and the 30HX graphics card could boost my current workflow efficiency by about 5 times, and also enable the use of AIs that were previously inaccessible, such as 27B large models.
I bought a P106-100 last year for about $25, but it broke. Now I plan to buy a 30HX, which costs $30. I currently don’t have any spare funds, so even for just $30, I won’t buy it for now.
This is a classic chicken-and-egg problem in making money: buying a graphics card can earn more money, but without buying one, you can’t earn the money to buy it.
As I’m writing this article, I’m also automating. By the way, this computer with 64GB DDR4 RAM serves as both my AI server and my main daily work machine. I only have this one computer, and I really wish I could set up a dedicated AI server that I could call services from.
Yesterday, my memory maxed out, causing the computer to freeze. So now, for work, I’m also considering dynamic memory services and off-peak usage. In summary, the efficiency of both hardware and software limits my automation efficiency, and there’s no immediate solution. I can only wait until I sleep to start automation, for example, during the 6 hours I sleep at night, when automation won’t interfere with other tasks.
Blog Network
100 blogs as a blog network. I won’t do 100; I’ll start with 3 for now, calling it a diversified matrix.
One is an AI blog, one is an open-source hardware blog, and one is a VPS open-source software cloud service blog—all tech-related.
I’ll see which one is search engine friendly, which gets more clicks and impressions, or it might even get zero indexing, as my blog initially included the categories mentioned above, but Google didn’t index it.
All in all, I’ll test automation or semi-automation first. Later, I might expand into finance-related topics, but that’s just a to-do item and not necessarily something I have to do.
In the short term, without a good graphics card or a paid AI like Codex, I won’t make major changes to my currently stable workflows. Even if I do make changes, I’ll back them up first.
Next, I plan to continue researching video workflows, because the blog is relatively stable, but video still hasn’t achieved the desired results.
Tips
Currently, the blog hasn’t brought any traffic or income. I’ll see how it goes in a week or a month. If it still doesn’t, then I’ll abandon the blog.
At the same time, this week my focus is on YPP long videos, to see how they perform.