Category: LLMs

Posted on

Ebeneezer’s Yuletide Homily

This interview just reiterated how everyone is still gob-smacked at how much transformers turn out to be able to do by swallowing and chewing more and more of the Internet with more and more computers, by filling in all the local bigram blanks, globally.

They’re right to be gob-smacked and non-plussed, because nobody expected it, and nobody has come close to explaining it. So, what do they do instead? Float a sci-fi icon – AY! GEE! EYE! – with no empirical substance or explanatory content, just: “It’s Coming!”
 plus a lot of paranoia about what “they’re” not telling us, and who’s going to get super-rich out of it all, and whether it’s going to destroy us all.

And meanwhile Planetary Melt-Down (PMD) proceeds apace, safely muzzled by collective cog-diss, aided and abetted by those other three Anthropocene favorites: Greed, Malice and Bellicosity (GMB) — a species armed to the hilt, with both weapons and words, and words as weapons (WWWW)..

Wanna know what I think’s really going on? Language itself has gone rogue, big-time. It’s Advanced alright; and General; but it’s anything but Intelligence, in our hands, and mouths.

And it started 299,999 B.C.

Posted on

“Body Awareness”?

Body Awareness: Scientists: Give Robots a Basic Sense of ‘Proprioception’

Giving a robot a body with the sensorimotor capacities needed to ground its words in its sensorimotor capacity to pick out and interact with the referents of its words in the world is grounding.

But that is completely (completely) trivial at toy-scale. 

No “grounds” whatsoever to talk about “body awareness” there. 

That would only kick in – (if dopamine etc. is not needed too) — at Turing-scale for robotic capacities. 

And if those include human language-learning capacity, that would pretty much have to be at LLM-scale too (not Siri-scale).

Now here’s the rub (and if you miss it, or dismiss it, we are talking past one another):

There is no way to ground an LLM top-down, dropping connections from individual words to their sensorimotor referents (like a man-o-war jelly-fish).

The only way to “align” sensorimotor grounding with the referents of words is bottom-up, from the ground)

That does not mean every word has to be robotically grounded directly.

Once at least one Minimal Grounding Set (Minset) has been grounded directly (i.e., robotically) then (assuming the robot has the capacity for propositions too — not just for directly naming everything) all the rest of LLM-space can be grounded indirectly by verbal definition or description.

Everything is reachable indirectly by words, once you have a grounded Minset.

But that all has to be done bottom-up too.