So, if I understand this correctly, a few individuals in the USA are buying up all the RAM and GPUs being manufactured in the world, with huge loans that they are hoping they will be able to pay off someday, to put into some of the biggest data centers ever constructed, all over cheap, peaceful villages, for which the electric grid has to be substantially upgraded, massively increasing our carbon emissions, and a large portion of the available fresh water supply has to be redirected, so that we can consume:
- Generative AI: A Large Language Model (LLM) that has gobbled up most of humanity’s literature and content on the Internet, so that it can predict the next word in a sentence.
- Agentic AI: A list of traditional IF THEN statements that do things based on the output of the magical LLM and then create a new prompt that they feed back into the LLM in a never ending loop, with the aim of consuming as many tokens as possible.
- Ontology: A long list of relationships between real-world objects and concepts to constantly remind the LLM not to make up impossible stuff.
Correct?
I suspect that they have run out of ideas how to make us buy new smartphones.
Thanks, interesting thoughts. I am trying to understand agentic AI and its potential uses and security risks better. I am also a bit hesitant to let it loose on my laptop, even if you sandbox the thing. I was also amazed how all our executives were pushing us to use AI for everything, without really trying to figure out legitimate use cases that will move the needle and without considering the massive potential costs. Then some of our agents suddenly got switched off, when token budgets were used up faster than expected.
The S in “agentic” stands for security!
I almost forgot: The reason why “business leaders” want to see AI everywhere is because they think it’ll be just like the adoption of every other technology to this point: It gets cheaper over time (not always better, but usually so).
The assumption is that if they “beat their competitors” to be the first ones using AI efficiently, they’ll utterly destroy them (economically; they won’t be able to compete). It’s a very bad assumption.
So far, “Big AI” is getting better but at costs that scale geometrically with the amount of “better”. That is: You can improve reliability (e.g. reduce hallucinations, increase accuracy, improve outputs in various ways, etc) but only by drastically inflating the cost and at reduced speed and efficiency.
We’re starting to learn that LLMs need about two generations of hardware advances before they’re going to be cost efficient for the types of “human productivity enhancement” that business leaders want. It’s only affordable now because “Big AI” is subsidizing the costs, trying to get customers hooked. The assumption being that if they’re hooked on AI, they’ll be able to raise prices to reflect actual costs. Just like the business leaders, this is a very bad assumption.
Instead, what’s going on is that the open weights AI models are starting to get “good enough” for most tasks (that you’d want to use them for, e.g. coding or agentic automation stuff). That means that all the billions and billions being spent by Big AI is just building up debt that will never be repaid and having this side effect of using up all the chip/memory capacity in the entire world.
It’s an absurd situation and the world will eventually look back on this time like we do the dotcom era.