How to Fail with GenAI Projects

The most famous quote on inversion comes from Carl Gustav Jacob Jacobi, the renowned 19th-century German mathematician. His famous maxim is: "Invert, always invert." This concise quote encapsulates Jacobi's problem-solving philosophy. Jacobi believed that many difficult problems in mathematics could be solved by looking at them from an inverse or opposite perspective.

Therefore, using the principle of inversion, here is a quick and simple recipe to fail with GenAI projects. I have not patented this recipe, nor do I intend to charge royalties, so please use it as often as needed:

How to Fail with GenAI Projects:

1. Always start with ambiguous, unclear, and confusing project goals.

2. Set the wrong client expectations on the end goal and the definition of success. Create as much confusion as possible.

3. Why start with an AI hypothesis? We already know GenAI can solve any and all problems. Do not educate the client on the capabilities, shortfalls, and risks.

4. Propose the biggest, largest, most complicated project you can on day one. Why take small steps with cutting-edge and ever-evolving technology? Promise the moon on day one.

5. Absolutely no need for a POC; it all works on day one. Speaking of POC: POC, Pilot, Production, etc. are all the same, aren't they? Why confuse the client by trying to be precise? Budget, schedule, and staff like a POC, but tell the client they can use it for production. Rigorous testing isn't that necessary for production anyway.

6. LLMs are very deterministic and are a settled technology, so promise very high and precise goals, such as 99% accuracy, way above human parity.

7. Use an LLM to solve any and all problems; after all, they are the new shiny object. Essentially erase and forget all software engineering best practices learned over the last 50 years.

8. The more complex, the better. They will hire us right back to maintain the darn thing.

9. Prompt engineering is just English, so in the first couple of tries, if you don't get the desired output, move to fine-tuning or, even better, take an empty model with random weights and try to create your own LLM. In the 4-week $50K budget, we should be able to afford the GPUs and the training data required to train an LLM.

I can keep going, but use any of the three above, and you will be guaranteed failure. If you do get partial success (meaning I am wrong), then add two more, and that should assure failure.

So then, how do you succeed with GenAI projects? Well, you know where to reach me. :).

Or use an inversion prompt, something like:

Using Jacobi's inversion principle, please invert the following [insert here]. Generate concise, clear, and step-by-step recommendations and best practices.