January 7, 2026 at 12:00 am
Comments posted to this topic are about the item Your AI Successes
January 7, 2026 at 2:22 pm
I've moved into the "doubtful" camp on AI becoming capable of "big successes" on its own. It is a great junior developer and saves time on tasks, especially those that are outside a person's area of expertise. It just makes too many mistakes to depend on for large tasks, and without some sort of AGI (which I don't believe is possible) it'll always make those mistakes.
Be still, and know that I am God - Psalm 46:10
January 7, 2026 at 2:56 pm
It's by trying it that we can see how good or bad it can be. Tests and concepts are there to show its potential.
We have an agentic AI system to do event correlation. With enough input (other incidents, change management database, internal documentation, etc.), it's pretty good at doing the first analysis, saving time on the data gathering and summerization portion.
But for sure, we are not ready to let AI take unsupervised actions.
I use a graph database and we wanted to see how an AI can interact with it. I wrote a JSON file listing the main node types and relations, with a list of DOs and DONTs recommandations. And it works! An agent dynamically builds the statement and another executes it (with coded safeguards to make sure it is a read-only statement). It's able to interpret error messages (given that this message is very complete) so it can understand the problem, recover and try again.
I even used the AI itself to tune that JSON schema file to make it more efficient and shorter. That's how I learnt that CAPITAL LETTERS are understood as being important.
While testing this, I gave it an incorrect server name to work with. After seeing that there was no server by that name, it ran more statements and figured that there is some zero padding in other server names. So, it changed the name I gave it and queried it again. I felt that it reacted like a kid, all proud to find the answer to some kind of riddle.
But, it was also clear that this searching and retrying is where many AI can go on a completely incorrect path, losing track of the context and allucinating.
So, I agree with you. While it's very capable, it's a kid only good for internal stuff with controled risks. It's not there for public production grade unattended work.
January 7, 2026 at 4:01 pm
I've moved into the "doubtful" camp on AI becoming capable of "big successes" on its own. It is a great junior developer and saves time on tasks, especially those that are outside a person's area of expertise. It just makes too many mistakes to depend on for large tasks, and without some sort of AGI (which I don't believe is possible) it'll always make those mistakes.
Perhaps, though the models are changing rapidly and they get better and better. It very much depends on training, so some language/domains do much better than others.
January 7, 2026 at 4:15 pm
...
So, I agree with you. While it's very capable, it's a kid only good for internal stuff with controled risks. It's not there for public production grade unattended work.
Agreed. I find it more and more capable of getting work done with periodic check-ins. Not unlike me training a junior person, who might make a lot of those same mistakes. An AI can go quicker and at scale, so guardrails are important.
January 7, 2026 at 6:16 pm
In SQL, understanding the relationship between objects (dependencies) is one of the keys to understanding the database inner workings. While there are a few commercial tools to do that, I found them to be somewhat unfriendly and lacking in features.
In October 2025, looking at an AI client (Grok), I asked: How would you set up a C# application to analyze SQL objects dependencies?
Within seconds, I received a 'simple' but workable C# application. The core was a well-designed and complex query returning both dependencies streams. That query had been tweaked over time, and the insights from AI were (mostly) right on the point.
I measure 'success' in AI use based on the ratio of NEW and WORKING code, versus NEW and GARBAGE, including clobbering working code. In the last three months, my batting average was about 50%. Not bad. If I had to fully code and type what has become a full fledge application, it would have taken me a year, not three months.
The value of AI is definitely dependent on the 'prompts'. Selecting one tool versus and other? A mixed bag, though, when coding for SQL, I found Sonnet 4.5 and Grok (fast 4) to be the more reliable.
The models are trained based on 'public' info. For SQL, such info may be way less that say JAVA or PYTHON. We need to use AI to help improve the models as these learn from our work.
For the curious, I have attached a screenshot of the application I referred to.
Viewing 6 posts - 1 through 6 (of 6 total)
You must be logged in to reply to this topic. Login to reply