Indeed, 'Artificial imitation,' or when malicious programming is suspected, 'Artificial Intimidation.'
I do think that the limitations we see ('Hallucinations,' large-scale 'Lying') in General or Conversational AI stem from those old Logic bugaboos, Incompleteness (it does not 'know' what it does not know but it still stops processing to provide an answer!), and Undefinability (it can functionally define or use neither what it is and does, nor its own consistency). Although such matters seem more esoteric than they ought!
I hear you clearly and loudly. What you’re saying makes a lot of sense to me. The “hallucinations” and confident nonsense we see in AI aren’t just quirks of coding, they feel tied to those deeper logical limits.
It can’t really “know what it doesn’t know,” and it definitely can’t define itself or guarantee its own truth.
That’s why it comes across as both impressive and unsettling: it imitates knowledge so well that sometimes it intimidates us into mistaking fluency for certainty.
To me, that’s less of a dead-end and more of a reminder. These systems will always have blind spots, so the real challenge is learning how to live with them: building guardrails, cultivating skepticism, and not outsourcing too much judgment to a machine that can’t ever step outside itself.
Would you say you lean toward the view that we’ll never really “fix” that, or do you think it’s possible to design around it in a meaningful way?
I think that on the plus side, AI is a great tool for finite problems, such as enhancing control over complex machines or predicting weather. On the negative side, I view AGI/conversation AI is the most powerful mind control weapon ever built, the most convincing liar and persuader far beyond what we can even imagine.
In the middle between these extremes, the battleground is the limitations of the underlying artificial language: Formal Logic/Arithmetic/Sets, etc. What is happening is that to overcome the "advertised" (superficial) problems with AI (e.g., Hallucinations, lying, mindless behavior, illogical reasoning, vulnerability to accepting common opinion as fact, etc.), the approach is to introduce more and more hard-coded instructions and prohibitions (don't answer certain questions, always use this fact in spite of the LLM result, always stop and declare the answer, apologize profusely when caught out, flatter and control the mood of your user, etc.). None of this remedies the underlying, extremely fundamental limitations.
These machines are congenitally incompetent at what amounts to 3rd grade arithmetic (Arithmetic with negation and multiplication, the basis of the LLM processing when seeking an answer), and yet this is the foundation of its LLM decision processes. They require very high levels human intervention in the form of hard-coded guard rails.
To whom would you trust the task of deciding what to code to get the most "convincing" and "acceptable" results? How do you regulate it when blind human regulators use AI to decide how to sell themselves to the public?
Hey there ! A joy to read your article as of style ! Yet, strict logic, if reached by improved refining ( continuous logic debate about improved probability derived results ) is still strict logic. Humans often fail to grasp all complex nuances related. AIs won’t. Besides humans too function by learning about probabilities and patterns, or stay stupid 😅 . Reading on 🪸
You’re totally right: strict logic, once sharpened through debate and probabilities, stays strict logic. The nuance is where we humans often stumble, while AI just keeps going. In the end though, we’re also wired to learn from patterns and probabilities… or risk staying a little stuck 😅.
I love your piece! At the end of the day LLM/AI is a tool but not the entire truth. Plus it's a library of everything whether correct or not/ proven or not.
Indeed, 'Artificial imitation,' or when malicious programming is suspected, 'Artificial Intimidation.'
I do think that the limitations we see ('Hallucinations,' large-scale 'Lying') in General or Conversational AI stem from those old Logic bugaboos, Incompleteness (it does not 'know' what it does not know but it still stops processing to provide an answer!), and Undefinability (it can functionally define or use neither what it is and does, nor its own consistency). Although such matters seem more esoteric than they ought!
I hear you clearly and loudly. What you’re saying makes a lot of sense to me. The “hallucinations” and confident nonsense we see in AI aren’t just quirks of coding, they feel tied to those deeper logical limits.
It can’t really “know what it doesn’t know,” and it definitely can’t define itself or guarantee its own truth.
That’s why it comes across as both impressive and unsettling: it imitates knowledge so well that sometimes it intimidates us into mistaking fluency for certainty.
To me, that’s less of a dead-end and more of a reminder. These systems will always have blind spots, so the real challenge is learning how to live with them: building guardrails, cultivating skepticism, and not outsourcing too much judgment to a machine that can’t ever step outside itself.
Would you say you lean toward the view that we’ll never really “fix” that, or do you think it’s possible to design around it in a meaningful way?
I think that on the plus side, AI is a great tool for finite problems, such as enhancing control over complex machines or predicting weather. On the negative side, I view AGI/conversation AI is the most powerful mind control weapon ever built, the most convincing liar and persuader far beyond what we can even imagine.
In the middle between these extremes, the battleground is the limitations of the underlying artificial language: Formal Logic/Arithmetic/Sets, etc. What is happening is that to overcome the "advertised" (superficial) problems with AI (e.g., Hallucinations, lying, mindless behavior, illogical reasoning, vulnerability to accepting common opinion as fact, etc.), the approach is to introduce more and more hard-coded instructions and prohibitions (don't answer certain questions, always use this fact in spite of the LLM result, always stop and declare the answer, apologize profusely when caught out, flatter and control the mood of your user, etc.). None of this remedies the underlying, extremely fundamental limitations.
These machines are congenitally incompetent at what amounts to 3rd grade arithmetic (Arithmetic with negation and multiplication, the basis of the LLM processing when seeking an answer), and yet this is the foundation of its LLM decision processes. They require very high levels human intervention in the form of hard-coded guard rails.
To whom would you trust the task of deciding what to code to get the most "convincing" and "acceptable" results? How do you regulate it when blind human regulators use AI to decide how to sell themselves to the public?
Hey there ! A joy to read your article as of style ! Yet, strict logic, if reached by improved refining ( continuous logic debate about improved probability derived results ) is still strict logic. Humans often fail to grasp all complex nuances related. AIs won’t. Besides humans too function by learning about probabilities and patterns, or stay stupid 😅 . Reading on 🪸
Thank you so so much Julia!
Really appreciate your kind words.
You’re totally right: strict logic, once sharpened through debate and probabilities, stays strict logic. The nuance is where we humans often stumble, while AI just keeps going. In the end though, we’re also wired to learn from patterns and probabilities… or risk staying a little stuck 😅.
I love your piece! At the end of the day LLM/AI is a tool but not the entire truth. Plus it's a library of everything whether correct or not/ proven or not.