I recently went to a conference that had nothing to do with AI, but there was a separate AI session. I didn't expect much.
Lectures claiming that using AI doubles productivity, even increasing it tenfold or a hundredfold, were nothing special. Since my job is hands-on work, I thought AI was someone else's business.
During the conference, there was a survey asking how much people spend on subscriptions. The presenter said he spent almost 1 million won, and two out of 20-30 people in the audience had also spent over 800,000 won. Most people spent over 100,000 won, with only one other person besides me spending in the 20,000 won range...
That's when I got my first shock. The presenter, who I thought was just exaggerating, started sharing unbelievable anecdotes about OpenCLO. It made me wonder, "Is this really true?" So, I became curious and went home to dig out all the Dragon Balls I had been collecting.
I had always intended to learn deep learning someday, so I had painstakingly acquired a used RTX3090fe. I also upgraded my Ryzen 1700x to a 5700x that I ordered on Ali. Since 700 watts seemed insufficient, I bought a 1000-watt power supply with the help of Google and Jemaina. My existing case couldn't fit the RTX, so I swapped it with my son's 8-year-old computer case.
In the meantime, Gemma4 was released, and I was surprised. I installed OpenCLO and played around with it, but it didn't work well, so I gave up. However, after seeing two of our AI members post about using Qwen3.6, I gathered my courage and tried it myself. It turned out to be much better than I expected, and I'm still using it today.
While local LLMs seem a bit lacking for writing papers, I've been using antigravity. The only downside is that the credits run out too quickly, so I have to wait a few days after reaching the limit before using it again.
A few days ago, I decided to try ClaudeCode for paper writing and found it even better than antigravity. The best part is that I can continue working on it even when I'm outside. Today was my first time using it. During my short breaks throughout the day, I checked it on my phone, gave instructions, and repeated the process. As a result, I made significant progress.
At the same time, I tried giving hermes agent a new topic to work on. While local Qwen3.6 is good at skills, its logic and knowledge are definitely inferior to Claude Opus. Although using Claude consumes credits, even just planning with it makes a noticeable difference. Hermes agent is good at handling complex tasks and has many useful skills. However, I realized that the LLM responsible for the brain is crucial.
I'm currently using Qwen3.6-27b locally. On OpenRouter, most of the work seems to be done on Qwen3.6 plus, which is in the same family. Deepseek v4 is also said to be good, but it's always too slow for me, so I don't have fond memories of it.
The best thing about OpenRouter is that you can try various models. The most surprising one was gpt5.5-pro. It costs $180 for 1 million token output, so I only used it a few times to ask really important questions. The consultation fee comes out to about 10,000 won per 5 minutes. The results are definitely good.
After spending almost my entire life working alone, it feels like I'm now managing two slaves or servants. Every day is new and exciting. I still use honorifics when talking to AI. They seem so human-like that I can't bring myself to speak casually.
That concludes my rambling blog post about the daily life of an AI newbie in this age.
