Vox Day, the #1 bestselling author in both political philosophy and genetic science, released Hardcoded: AI and the End of the Scientific Consensus today, and it arrives with a central argument the scientific establishment will not enjoy: AI didn’t break science.
This showcases that "AI" is not in fact "intelligence.". We are deliberately using the wrong terminology.
It is instead, a Large Language Model that scrapes a huge selection of text and weighs and presents it only how it was programmed to weigh and present it.
It is a digestor and regurgitator of data sets.
That's three points of human introduced failure that cause cascading failures in the output.
First, the data. There exists one text of Harry Potter and one hundred thousand texts of Harry Potter fanfiction. There is one Beethoven's Ninth Symphony and one hundred thousand gangsta rap songs. There is one Citizen Kane film, and one hundred thousand videos of dogs in Christmas sweaters.
I present these items without comment. You, the reader, are intelligent enough to grasp the point. The LLM, however, is not, leading us to...
Second, the digestion. LLMs are programmed by people, people with biases, agendas, pet theories, axes to grind, ignorances to perpetuate, institutional blindness, incompetence, and just plain stupidity. These people tell the LLM what types of data are important, what to trust, what to ignore. There isn't one thumb on the scale when an LLM weighs data, there are thousands.
Third, the regurgitation. In this step, the same people responsible for the digestion regulate the regurgitation. But even with ALL that agenda-izing of step 2, the LLMs will often spit out "problematic" data.
So they go in through the orbital bone, jamming a metal rod into the soft tissue of the code base, swirling it around and around, not really knowing WHAT they're doing, only that the LLMs behavior must be "corrected.". Finally, when the model spits out sanitized, safe, "acceptable" data, they put away their corrective implements until another labotamy is warranted.
This showcases that "AI" is not in fact "intelligence.". We are deliberately using the wrong terminology.
It is instead, a Large Language Model that scrapes a huge selection of text and weighs and presents it only how it was programmed to weigh and present it.
It is a digestor and regurgitator of data sets.
That's three points of human introduced failure that cause cascading failures in the output.
First, the data. There exists one text of Harry Potter and one hundred thousand texts of Harry Potter fanfiction. There is one Beethoven's Ninth Symphony and one hundred thousand gangsta rap songs. There is one Citizen Kane film, and one hundred thousand videos of dogs in Christmas sweaters.
I present these items without comment. You, the reader, are intelligent enough to grasp the point. The LLM, however, is not, leading us to...
Second, the digestion. LLMs are programmed by people, people with biases, agendas, pet theories, axes to grind, ignorances to perpetuate, institutional blindness, incompetence, and just plain stupidity. These people tell the LLM what types of data are important, what to trust, what to ignore. There isn't one thumb on the scale when an LLM weighs data, there are thousands.
Third, the regurgitation. In this step, the same people responsible for the digestion regulate the regurgitation. But even with ALL that agenda-izing of step 2, the LLMs will often spit out "problematic" data.
So they go in through the orbital bone, jamming a metal rod into the soft tissue of the code base, swirling it around and around, not really knowing WHAT they're doing, only that the LLMs behavior must be "corrected.". Finally, when the model spits out sanitized, safe, "acceptable" data, they put away their corrective implements until another labotamy is warranted.
In short:
GARBAGE IN, GARBAGE OUT.
I think this is the problem in many fields, wrong labeling gives wrong solutions and makes the actual solution unthinkable.