Alpha System Trading develops its programs with Google Antigravity.
And I also get a lot of help from Google Gemini.
(I was lucky enough to get long-term access to Google AI Pro.)
Here is my recent experience with Gemini 3.8 Flash.
Gemini 3.8 Flash Has Gotten Smarter
Gemini 3.8 Flash made the basic data extraction process much faster.
In the old version, I had to run five different processes with different conditions to extract the basic data.
It took about 70 minutes.
Gemini 3.8 Flash spotted the duplicated calculation structure and cleaned it up.
As a result, it could extract all the basic data in a single run.
The processing time dropped from 70 minutes to just one-fifth of the original time.
That means my computer spends much less time running its CPU and GPU fans at full speed.
And, of course, I save time too.
In the past, I would start the calculation and then go do something else for an hour.
Now it finishes in about 15 minutes.
That makes it much easier to stay focused on the work.
But Sometimes It Does Something Really Strange
Dumb Gemini 3.8 Flash also makes mistakes.
And it makes them with a lot of confidence.
For example, I asked it to show important figures such as MDD and the maximum underwater period in the system summary report.
But at some point, it started leaving them out.
Instead, it sometimes suddenly showed a weekly report that I did not need at all.
Every time this happens, my frustration slowly goes up.
So now, if it makes the same mistake twice, I immediately add a new rule to AGENTS.md
(Of course, if the mistake involves something important, I fix the rule right away.)
The Case of 99 Turning into 99%
The biggest mistake in this update cost me an entire Saturday and a lot of tokens.
During the basic data generation process, I use the value 99 to exclude stocks ranked below 99th place out of 200 stocks.
But the AI decided to interpret it as “below 99%”.
Just not what I actually meant.
The result was a complete mess.
99% or lower excluded = only 2 stocks remained out of 200.
The real problem was that I did not notice it right away.
The system continued calculating everything using the bad data.
I only found the problem at the very end.
I remember thinking:
“Why do these numbers look so strange?”
So I went back through the data, one piece at a time, until I found the cause.
There was another similar mistake after that.
But this time, I caught it much faster.
That was when I started adding a rule to AGENTS.md whenever Gemini made the same mistake twice.
So far, I have not seen the same problem again.
Dealing with Repeated AI Mistakes with AGENTS.md
For now, I guess I will keep using my paid Google AI and try to keep Gemini under control.
So, what is the conclusion?
Would other AI models do the same thing?
Should I pay for another AI model too?
For now, I will keep working with Gemini 3.8 Flash.
Smart!
Fast!
And sometimes… very dumb!
Especially when it turns 99 into 99%.
