Cool story bro... But massive distributed software systems, big data and data analytics is what I do for a living (and have for the past 25 years now).
The cost of constantly training the LLM's plus the infrastructure costs are just too damn high when using real data for training, and trying to use A.I. generated data to train the models results in more and more "hallucinations" (the bullshit term they use for garbage outputs). Trying to train the LLM with generated inputs ultimately wrecks the whole model.
Businesses don't have anything they can sell from the machine learning outputs that are able to earn more money than all the massive infrastructure and model training costs... which means negative ROI. They end up losing money.
Chat GPT infrastructure alone costs almost a quarter of a Billion dollars a year in the AWS cloud - and that does not include any of the huge costs of training the models (lots of people manually labeling and categorizing and tagging all the inputs and then scoring all the outputs... very time consuming and very expensive - and it NEVER ENDS).
So - I would say that you are the one that is "wildly wrong".
Now in terms of building a total surveillance state to enslave humanity (like communist China is doing right now)... the capabilities of "A.I." are indeed extremely useful... and tyrannical governments usually have deep pockets. That right there is your best potential market.
I've been PE at FAANG/MAMAA for about 15 years now. So. Hi.
I think if you are solely positioning your thesis around LLMs, then I think it's precarious. My personal ROI as an Engineer is insane. I'm far more productive by all measures. If you're a desk-worker and haven't figured out how to augment your productivity with an LLM, you're going to be left behind.
Other baked ML algos trained faster on discrete datasets already has multiple ROI positive applications, and has for years. Cataloging, CV, predictive analytics, speech recognition, NLP, etc. etc. I can't imagine you've not seen this if you truly are in big data?
Cool story bro... But massive distributed software systems, big data and data analytics is what I do for a living (and have for the past 25 years now).
The cost of constantly training the LLM's plus the infrastructure costs are just too damn high when using real data for training, and trying to use A.I. generated data to train the models results in more and more "hallucinations" (the bullshit term they use for garbage outputs). Trying to train the LLM with generated inputs ultimately wrecks the whole model.
Businesses don't have anything they can sell from the machine learning outputs that are able to earn more money than all the massive infrastructure and model training costs... which means negative ROI. They end up losing money.
Chat GPT infrastructure alone costs almost a quarter of a Billion dollars a year in the AWS cloud - and that does not include any of the huge costs of training the models (lots of people manually labeling and categorizing and tagging all the inputs and then scoring all the outputs... very time consuming and very expensive - and it NEVER ENDS).
So - I would say that you are the one that is "wildly wrong".
Now in terms of building a total surveillance state to enslave humanity (like communist China is doing right now)... the capabilities of "A.I." are indeed extremely useful... and tyrannical governments usually have deep pockets. That right there is your best potential market.
I've been PE at FAANG/MAMAA for about 15 years now. So. Hi.
I think if you are solely positioning your thesis around LLMs, then I think it's precarious. My personal ROI as an Engineer is insane. I'm far more productive by all measures. If you're a desk-worker and haven't figured out how to augment your productivity with an LLM, you're going to be left behind.
Other baked ML algos trained faster on discrete datasets already has multiple ROI positive applications, and has for years. Cataloging, CV, predictive analytics, speech recognition, NLP, etc. etc. I can't imagine you've not seen this if you truly are in big data?