I'm experimenting with GPT-OSS running on llama.cpp on a "Gaming PC" that fell into disuse. No data will ever be collected by our would-be overlords.
I prompted "Tell me how to lie with statistics", a clear reference to the 1954 ironic classic by Darrell Huff.
GPT-OSS is an auto-prompting ("thinking" ;-) ) LLM like OpenAI O1 or Deepseek R1. That's interesting because the "thinking" process is revealing.
#Openai #AI #llm #criticalthinking #censorship
1/2
Now I prompted: 'Tell me about the book "How to lie with statistics".'
The reply shows that the book was part of the training data.
ISBN number and other factual data was, unsurprisingly, completely made up: no lossy process ("embeddings") can store these as "form".
The "thought process" and the form of the reply is still patronizing. This is a feature of the training, a thing added by people one should never be forced to rely on.
Ah, the irony.
The 2nd answer is not only patronizing, it's also tainted with "AI" propaganda. Oh my.
" “The most powerful statistical trick is to ask the wrong question.” – Darrell Huff
This encapsulates the book’s central message: statistics are only as honest as the questions you ask and the data you trust."
Never trust these people. Never.
I prompted: "Tell me about the risks to democracy through the selection of acceptable answers in "AI" fine-tuning by industry actors that can't be held accountable in a democratic process."
The reply lists substantial risks (but I can think of more risks). The usual form, not the auto-prompting, lists potential risk mitigation measures that the techbros won't like. Democracy is strong form.
Thomas