My 5 Biggest Surprises as an AI Developer

I recently started doing software development work to expand and integrate AI systems.

Here are some of the biggest surprises I found along the way.

\

AI Developers Love Calling Everything “Fine Tuning”

\ When studying for the AI-102 exam, I thought fine tuning strictly referred to additional training of a model to include additional data (which can be fed into the system in different ways).

When actually working on an AI project, what I learned was developers (and architects) sprinkle this phrase on just about everything.

Do you need to get some more exposure to how LangGraph works ? Well that learning process is “fine tuning”.

Does the AI model need outside data and to assemble that into the response at run time ? You thought that was a RAG ? OK bro, I guess, but we’re going to call that fine tuning as well.

Is there a bug in the code anywhere ? We used to call that a feature, but it looks like we just need to do some “fine tuning”.

\

Successful People Use AI A Lot

I’m aware of the MIT study showing that most AI projects fail.

What people are less familiar with is all the data showing free lancers who use AI make a ton more money than those who don’t.

If you’re not using AI (especially on an individual level) to get work done quickly, you are going to be at a disadvantage.

A lot of companies (including one of my current clients) understand this and are trying to adopt this capability.

And this includes companies that aren’t traditionally seen as technology companies.

I was recently surprised by a colleague who allowed one of his AI certifications to expire. He said he wasn’t getting any bites from the market. This might be because he was trying a year ago and things have heated up since then. Or factors beyond our control. In any case, I landed a spot on an AI project within a week of passing the AI-102.

\

Successful People Are Discrete About Using AI

Just like students hide how much they are leaning on ChatGPT to answer their homework problems (in spite of the learning potential and capability it brings), successful AI projects hide how much they lean on AI systems to understand what is going on.

There may soon come a day when AI does not have the stigma that it currently has, but until that day happens, clients (and teachers) want any assurance you can give them that you have internalized (i.e. learned) what you are sharing with them.

Even if it is totally infeasible to think that a developer should understand minutia about network details.

Funny story, I talked to a high level AI worker in another company on the same project who wanted to collab with me about how to reformat CoPilot data so it doesn’t look so much like it came from CoPilot.

Thankfully, you don’t need anything quite so elaborate as that. You can always be discrete about where you got the data. In some companies it is SOP to tell the client “it doesn’t matter” where the data comes from.

Or alternatively, you can just tell the client you would need to investigate it. If it’s something exotic, then chances are your competitor does not have the answer off the top of their head and this approach is slightly more transparent and authentic.

\

AI Written Code is Too Verbose and Depends on Learning From Humans

AI written code quality can vary considerably (especially on the basis of what the AI was trained on). I’ve found that GitHub CoPilot can quickly add the features I need before having to do a deep examination of what the operating pieces of code are.

You can look at that as taking encapsulation to a new level or as criminal negligence, but we are likely heading toward a software market where you are set back for over analyzing the code components. An AI veteran on my project helped me integrate GHCP with my IDE and encouraged others on the project to meet our deadlines by using AI.

Anticipate having to make adjustments to the output.

Depending on your framework (python in this case), you might need to pare down the outcomes of three or four levels of truthiness and simplify the data handling to a few cases that are relevant to your project.

\ The Disruption Means a Lot of New Things to Try

While you should be careful about re-inventing the wheel even in something as new and disruptive as AI technologies (for example, your skills in training and building LLMs are extremely narrow, niche) … a lot of the basics in new AI tools are still being ironed out.

That means there is tremendous opportunity for creating new capabilities and mastering skills that no one else has yet.

For example, what is the best way to have an AI select (orchestrate) between tools ? Should you use a single agent ? An established framework like Semantic Kernel, Agent Foundry, LangGraph, or CoPilot Studio ?

There’s no obvious solution. Exploration is needed.

Another example is RAGs (or Retrieval Augmented Generation). As mentioned earlier, these paradigms extend what a model is capable of providing content on … and there is no platform that is as simple as clicking a button to provide that (although some claim they have this capability).

\ Software development using AI is a lot like the Wild West right now. You’ll find one new thing happening over here and a completely different new thing happening over there. My suggestion is to try things out, learn from other people’s mistakes, and discover what interests you.

\ \