Unseen side of AI

I am not talking about all the fear and loathing that AI will come for everyone's jobs and nobody is safe.

Sure, I get that AI gives Hackers, bad actors and nation state attackers super powers, but I am talking about the costs associated with training these AI models over YEARS!

Typically, when an enterprise buys into information technology (IT) after the product is installed and configured its up and aiding the firm in making processes simpler or making money. AI is different, not all data is the same (structured data, unstructured data, semi-structured data, Clean data). Not all data is kept in the same repositories, so data has to be transferred or copied. Then it has to be cleansed, then it can be ingested into the models.   That was the easy part, AI models must be trained! 

Case in point ChatGPT and Google Gemini

Cost millions of $, to bring to market and those costs have been surpassed with continued training for these AI models.  Last years ChatGPT release 4 

cost an estimated $78.4 million to train.   Meanwhile Google plowed $191M into Gemini upgrade to it to re-write history books SANs all Caucasian people and their accomplishments. So much for removing AI hallucinations and bias; well unless that was the intent all along.

AI Training costs

You have successfully trained your AI models. And you think you are done; wrong!

As the snippet below shows how training LLM have increased year over year since 2017.  Well, I am not sure enterprises will buy into this AI non-sense just from this fact alone. Perhaps they will leverage smaller models tuned just for enterprises?  Perhaps leveraging something such as IBM's recently Open Sourced Granite LLM's running on Red Hat Enterprise Linux AI (made available about a month ago) . These IBM granite modules were recently named most transparent so that makes sense to me

Summary

So, AI is a combination of humans and machines, gathering data, ensuring the AI models are trained while not exposing Personal Identifiable Information (PII). Additionally, any bias or incorrect info, seems like an expensive and lengthy process.  At some point the group think will have to be broken and a more efficient training process will be brought to the light.  I am not sure if the day-to-day enterprises can exploit AI unless its built into everyday applications, and price will become a barrier to AI much like a caste economic system.   I do not see that happening anytime soon, but in the end anything worth having will come through hard work!

Credit - Pexels.com for image