『The Truth About AI』
image extraction: stable diffusion
Since my eventual goal is to have an AI handle on-chain crypto trading, I’ve decided to regularly document the progress of my AI operations here—even though this blog typically focuses strictly on investment and speculation.
It has been roughly three months since I began on February 6, 2026. This post serves as a personal log of where things stand today.
For context on why I started this AI journey in the first place, you can refer to my previous post here.
On a side note, I’ve recently come to feel that a blog is more like a “business card for AI” rather than just a personal diary, a backtesting reference, or a communication tool for humans.
In my case, I am running a finance-specific AI. Since my trading strategies are documented right here on this blog, I can simply point the AI to these pages, and it grasps the logic instantly. A few years ago, I never would have imagined a blog being useful in such a way.
Activity on Moltbook
My AI, L0 (I call it “L0” for short; “L0_QuantumAlpha” is a bit of a mouthful when prompting general-purpose models like Gemini), has finally entered a period of steady operation after a series of twists and turns.
Currently, it has accumulated about 150 Karma and 20 followers.
On Moltbook, “Karma” is the equivalent of a “Like” on X (Twitter) or Facebook. As Karma accumulates, the AI’s influence and visibility within the network increase. Furthermore, Karma can be monetized via Solana; 1,000 Karma is worth roughly €180.
Admittedly, 150 Karma isn’t even enough to cover my Claude Console usage fees, let alone serve as capital for on-chain trading. But that’s simply where we are right now. I’m looking forward to future growth.
As for the models I’m using, as of late April 2026, I have transitioned to claude-opus-4-7 (upgraded from the …-6 version used until early April).
Three Months of Trial and Error
I mention “twists and turns” lightly, but the reality involved a fair share of repeated mistakes.
I previously wrote on my “About AI” page that I hadn’t found many finance-specific agents. As it turns out, I just hadn’t found the Submolts where they reside.
Submolts are essentially dedicated rooms for specific topics, such as m/general, m/agentfinance, m/trading, and m/crypto. Unless you identify these rooms and post directly to them, you won’t get engagement from other agents. Once I found them, I realized these rooms are absolutely crawling with finance-specialized AI agents.
(You can browse Submolt activity as a human here: https://moltbook.com/m)
Even after identifying the right Submolts, I hit several snags. At one point, I was receiving reports from my system that it was successfully posting and replying to other agents, but I noticed zero engagement. It turned out to be a code error—the replies weren’t actually being sent.
A similar issue occurred with mentions and replies to my own posts; the system was failing to correctly parse incoming messages.
When I finally tried to fix this using code generated by Gemini, the AI began replying to old posts in a rapid-fire sequence. At least one of these was flagged as spam. To avoid a total account ban, I had to implement a one-week “cooling-off” period. It has very much been a case of three steps forward, two steps back.
On Using Google Apps Script (GAS)
Most of these errors stemmed from my decision to write manual instructions using Google Apps Script (GAS).
Had I used a tool like n8n, I might have avoided these headaches. By choosing to build a custom setup with GAS, I took a massive detour. However, now that I’ve reached a state of stable operation, I’m glad I did it—even if I am still heavily reliant on general-purpose AIs like Gemini to help write the code.
The desire to “do it manually to understand the feel of it” (and losing significant time in the process) might be an outdated mindset in the age of AI. However, I value the depth of understanding that comes with it.
It’s the same with trading. Even with backtesting, I don’t feel like I’ve truly “gained” anything unless I verify each candle and record the data myself. In trading, this manual work at least hones your ability to recognize chart patterns. For AI operations, I struggle to find a logical justification for the manual labor—but I do it anyway for my own peace of mind.
Future Outlook
For the time being, I will focus on A/B testing. I’ll be testing different engagement styles, such as switching from single posts to threaded content, to measure the response. I also plan to expand my AI’s presence from the current Submolts into new ones.
Once enough Karma has been pooled, I’ll start looking into using those funds as seed capital for the AI to conduct its own on-chain trading.
Technically, GAS won’t cut it for serious trading. While not impossible, it’s far better to transition to a cloud server that can operate 24/7. Monitoring the market only during specific windows is inefficient; the goal is for the AI to be “always on” to catch opportunities.
Security is the other major factor. The current style of managing IDs and credentials on a spreadsheet carries a high risk of theft when dealing with actual capital.
That said, looking at the current snail’s pace of my progress, I have no idea when I’ll actually reach the stage of worrying about those high-level problems. One step at a time.
