"Please just read the effing text to me."—"I do not respond to harmful language. I can help you to find businesses along the road and monitor traffic conditions."—"I don't need to find businesses on the road."—"Here are some restaurants and shopping venues on the road…"
@plinz
-
Dialogue sur l’impossibilité de lire un essai demandé
By
–
"No, please read the essay to me."—"Should I summarize every paragraph?"—"No, I wan't the whole text."—"I am sorry, I don't have that text in my system, and cannot access the full text using the internet."—"We already established that you have the full text."—"I can summarize…"
-
Gemini car assistant frustrates with voice change workaround
By
–
Trying Gemini in my car and it's the absolute worst. Starting from the inability change the voice from within the assistant, or even using infotainment settings: you have to use a phone app, and then log out of the car and back in (which you have to figure out yourself).
-
AI misinterprets user requests for directions, name change, and essay
By
–
"Please navigate to this SF address but go via 280, not 110"—"Sorry, I cannot interact with the route finding". "Please call Andrew Thomas" — "Ok, I will call you Robert from now on." "Please locate the following public domain essay for me"—"You probably mean this other text"
-
LLMs prove reasoning can emerge from language statistics
By
–
I am not garymarcusing here. LLMs are proof that it is possible to distill intelligent reasoning behavior by doing statistics over human language patterns
-
Preferring actual mind over model without memory
By
–
It's ok that you are less articulate than the model. The model tends to be subtly beside the point, and I cannot correct its insights by arguing with it, because it has no memory or persistent identity. I'd rather talk to you as you are, because you are actual mind
-
LLMs behind human performance due to lack of regularization and integration
By
–
Yes; I think the main reason that LLMs are so far behind our performance relative to the amount of data they get is that they don't regularize and integrate enough
-
Kurzweil’s AGI predictions mixed with flawed connectome immortality
By
–
Kurzweil mixed bold but solid predictions (given enough compute, a neocognitron like architecture can achieve AGI before 2030) with dogshit (digitizing the connectome is a near term way to human immortality). But both looked like the same kind of scifi to non experts.
-
LLMs’ text input narrows world model compared to human perception
By
–
That's technically correct but also misleading. Our world model is a shadow of our perception, and the sheer amount of text going into LLMs is constraining this shadow more than the perceptual patterns a human brain can encounter in a lifetime of experience.
-
Deep Learning imitates AI by mining human output for intelligence
By
–
Technically we are not in an AI age, but in a Deep Learning age. Deep Learning is deep faking AI, (mostly) by imitating the patterns of human thought. That does not mean that Deep Learning is limited, but it is wasteful to dig through mountains of our output to find intelligence