View source for User:Dword/Personal Writing/AI isn't intelligent


{{DISPLAYTITLE:{{SUBPAGENAME}}}}  
We’ve known this. It takes human-made works, and spins it to look new. If anything, it’s a glorified copyright violation machine. For the remainder of this work, conventional so-called "generative AI models" shall be referred to by the more correct term "LLM" (standing for "Large Language Model", to differentiate it from actually useful Machine Learning algorithms). Additionally, in this case, "LLM" should extend beyond text, including images, videos, audio, and others. AI as a term is inaccurate, anyway.    

While I do know how Machine Learning works, and I have used Large Language Models in the past (I refuse to shy away from this fact), quite a few currently existing personal policies prohibit me from doing so currently. One such policy is a blanket personal prohibition on utilizing any and all Generative Machine Learning models (including those made in whole or in part by ChatGPT, Google Deepmind, Claude, Deepseek, Meta, etc.) for any purposes whatsoever in any capacity whatsoever, and I am unwilling to modify that policy.  

What LLMs do is rather impressive, in the grand scheme of things. When you train a model, you are turning human-readable information into a set of "weights" that go to a set of "neurons." These weights affect the sensitivities at which certain neurons "fire." Neurons can only take in one number, being any number from 0 to 1 (inclusive) with different widths (like, 32-bits, for example). Then, whatever is input needs to be tokenised (turned into discrete units) and fed into the neurons. Depending on how highly activated a neuron is, that is how confident the neural network (the whole collection of neurons, composed of many layers of neurons) of something. An LLM takes input, tokenises it, runs the neural network against those tokens (this is why tokens are such a big thing in LLMs), then takes the output. All LLMs do is give a "best effort guess" based on the input provided.   

But, at a point, you run into the Chinese room problem (does the machine actually understand the conversation, or is it just simulating the ability to understand the conversation?), of which is largely provable as being a very good way to refute LLMs.  

''Now... What could so bad about AI?''

For starters, I personally find that it violates many licenses, including the GNU General Public License version 3 or later, due to it creating (what I see to be) a derivative work and not emitting a proper or compatible license to cover it. Or, when software holding any version of the AGPL is consumed by LLMs when training, and that LLM is produced online, it can be viewed as a violation of AGPL.  

Additionally, it is a ''horrible'' toll on the environment. The datacentres that train and run them require a near infinite amount of energy, water, data, and other resources. Additionally, they produce pollution that, in almost every case, is seen as biologically non-compatible.  

Energy is commonly derived from fossil fuels, of which contribute to the amount of greenhouse gasses present in the atmosphere. Another issue is that, when power grids are connected to grant higher throughput for datacentres, the cost of infrastructure is passed onto the customers using the powergrids, not the datacentres.  

Water is an extremely important, yet limited, resource, of which datacentres use extensively in order to cool their GPUs. While some datacentre providers continue to promise "Every drop of water will be returned to local sources" even though that’s only ''part'' of the problem. Another issue with the immense water use is water pressure of surrounding areas being brought so incredibly low that the water may well not exist in the pipes. Additionally, datacentres make water dirty and, in many cases, toxic.  

Again, data sourced without permission of the author (data harvesting should be opt-in, not -out) is against copyright laws. The only exception I may grant is in the event a Privacy Policy or Terms of Service states that it harvests data, in which case you should cease use of that service (my opinions on SaaSS is vast).

The pollution that datacentres create is not only atmospherically problematic, it is also problematic in what they cause regardless of what energy or water source they use: Sound. The sound the generators, cooling fans and towers, and other sources create happens to be in the sub-kHz range. Research has shown that, as a result of constant, consistent, loud sub-kHz noise is not compatible with life. For example, people developing severe chronic disorders due to the way the brain works with sounds.

In addition to stealing individual resources, it also usurps the most global resource: land. For example, one proposed datacentre in Utah's Box Elder County, called The Stratos Project, is planned to be almost 40,000 acres, or 62.5 square miles. It is estimated to consume 8 gigawatts of energy. It is also estimated to produce 8 gigawatts of waste heat. That is 8,000,000,000 watts. Government regulation limits space heaters to 1,500 watts; yet this datacentre is projected to produce ~5.3 million space heaters. A gigawatt is typically the annually averaged power consumption for a city of 1.8 million inhabitants, and is the output of a large power station. This single datacentre is proposed to consume as much energy as 14.4 million people do in a year.