AI projects 101

AI projects 101

Do you even need that AI bro?

AI nowadays is like Marklar in this unforgettable South Park episode
(tl;dr - They use the word Marklar for everynoun in their sentences.)

AI is the new marklar.

In goods industry, you can sell anything for high price if they are marked handmade/organic.
In software industry, now you can sell anything for higher price if the product is said to be empowered by the power of AI.
AI is the new shiny organic handmade hammer, and we see the whole world as nails.

In 2024 if a software product is not enriched with AI, is considered oldschool and not worth to invest.
But let me ask:

As an AI engineer, it might be peculiar that I say that we need to be cautious when leveraging marklar...I mean AI.
I have very good reasons to say this.

  • AI is undeterministic (you get different answers for the same question)

  • AI is hard to QA (how do you decide if an image is good enough)

  • AI can be expensive (tools and workforce too)

  • There are no well-established best practices

  • AI can't solve everything (on its own)

So, before you jump into a project, you definitely need to assess the opportunities/limits of leveraging AI (whatever you mean by this).

Don’t let natural intelligence lose over artificial intelligence.

Let me share my thoughts how to decide if you need AI for a feature or not.

  1. You have an exact rule system: you know exactly what you expect from an input and its output, and more importantly you know how that output should be computed. Easy, it’s a hard NO AI.

    Develop a good old fashioned software instead.

  2. You want/need to be in full control of the results. For orgs like e.g. finance you cannot afford even a record of wrong/hallucinated result. Also a hard NO AI.

    You can automate your processes with non-AI software, RPA, templating, standard BI, etc.

  3. You need to generate content with strict rules, but the quality and content can vary between a treshold. This is a soft MAYBE AI. Proper verbose prompting with a top model or well-tuned smaller models might help you ensure your needs, although it’s hard to automate its quality assurance. For manual QA’s worth of data amount it can work.

    For those cases where it just assists in a manual process, consider a standard automatism to support the person responsible to content. A proper BI, summary might be more useful than revisioning AI hallucinations.

  4. You want a feature because it is doable with AI. This is a hard STOP sign. This might be the most important of all the points: don’t look for AI use cases l’art pour l’art without client need, because you will fail. If there is no demand for your AI feature, leave it, even if it would be fancy. Always look for painful problems to solve - and if it eventually calls for AI, only then do it, it can be a hard YES for AI.

    Otherwise go with traditional solutions.

And at last,

AI is a tool not the goal.

Keep your common sense. Start from client needs, choose the optimal tools and people and everything’s gonna be fine. With or without AI.

If you like the post, please share the ideas. Also, share your opinion, I fancy the diversity of thoughts.

I write the blog after bringing 3 small kids to sleep. Tomorrow morning will be hard. Please support me with my morning coffee if you would like to read more of my thoughts.