RE: https://mamot.fr/@pluralistic/116414078285695998
This line from one of @pluralistic 's typically excellent and thought-provoking articles set me thinking. And while I agree with his first statement, I'm less certain about the second.
To explain why requires me to tell an anecdote …
1/
Many years ago, Marvin Minsky came to the lab where I was working and gave a short talk about his “society of mind" concept. At the end of the talk, someone asked him what kind of representation he thought that the brain used for concepts.
(This was in the days of symbolic AI, and everyone was high on LISP and PROLOG, so the idea that the brain might have a formal representation of concepts didn't seem questionable or absurd).
2/
Minsky responded that the brain was basically an evolutionary kludge, and so he thought the brain used a little bit of everything. He named a handful of the then-popular formalisms for knowledge representation, trotting out a list that ended with “… and Conceptual Dependency Notation.”
3/
A few of us, myself included, made audible "WTF?” noises.
CDN was a formalism (to be generous) invented by Roger Schank. My own impression was that it was handwavy nonsense with no explanatory or computational utility.
(Side note: I once talked myself out of a job with Schank's group at ILS by inadvertently trashing work done by one of Schank's collaborators, unaware that my interviewer was one of his PhD students. In hindsight, I dodged a major, major bullet).
4/
Anyway, whether CDN was good or bad, the fact was that Minsky did NOT hold Schank in high regard (and vice-versa), so to hear Minsky saying that he thought CDN was a basic knowledge structure in the human brain was a little like hearing the Pope say that this Satan guy made some good points, actually.
5/
Minsky waited a beat, looking around the room to see who was paying attention, and then grinned and added “… the way I see it, if Conceptual Dependency Notation was simple enough for Roger to invent, it's simple enough to have evolved by pure chance.”
6/
I feel a little bit the same way about LLMs and intelligence: that intelligence (human or otherwise) will always be kludged together out of off-the-shelf parts, rather than designed top-down according to some glorious, overarching architecture.
In the case of the human brain, existing structures were probably recruited and subverted and repurposed to cobble together what we call consciousness.
7/
In the case of 'artificial general intelligence’, if it ever arrives (and I don't see why it can’t, although I don't believe it’s in any way imminent), then it'll be a similarly messy kludge assembled from bits and pieces.
And while I don't believe for a second that existing LLMs are in any way “intelligent", the technology underlying them is spookily good at doing SOMETHING, and that something might be a necessary part of building an intelligent system.
8/
This, for me, is the interesting bit of LLMs and GPTs. I'm not interested in their proven ability to spew slop (“not interested" = “I actively hate it”). Nor am I persuaded by the hypesters’ claims that their consumer-grade sycophancy-as-a-service is three point releases away from ushering in the Singularity.
But it does seem that LLMs are very … well, I was going to say 'efficient', but if you have to boil the Colorado River to write a corporate mission statement, efficient isn't the word.
9/
Let's just say that they're startlingly good at doing a particular kind of transformation task, and there is definitely something there that merits further study. (They're also wildly uncontrollable, which is a bit of a problem if you want repeatable or predictable results).
10/
So it seems to me that while most of the LLM-based systems we see today are a resource-intensive party trick with many ruinous side-effects, there's something very much more interesting about the core technology, and that THAT may ultimately prove to be one of the tools in the toolbox for whoever finally Frankensteins together something that we might cautiously call intelligent. And that's why I quibble with @pluralistic's “will not lead to intelligence" line.
11/
Also, on a side note, the question of “can we make intelligent systems?" may be a distraction; we're now facing a world in which we have to live alongside increasing numbers of near-intelligent or pseudo-intelligent or just very-good-at-making-you-think-they're-intelligent systems, and that will have “interesting” consequences. I touched on that in an essay at https://theastoundinganalogcompanion.com/2023/09/08/familiar-daemons/, but I don't think we’ve yet grasped how much these fake intelligences are about to change our world.
/END
Yes. If we assume human intelligence is a grab bag of ingredients, the chances that one of those ingredients is like an artificial neural network goes up.
Not the whole thing, just some components, like visual cortex or speech centers.
It's even more likely considering how ANNs both succeed (object recognition) and fail (hallucinations) similarly to ourselves.
'My AI ate social media for a year and now it just bullshits.' Yeah that never happens to people.
On this we agree.
Pluralistic serves useful function.
But he has reached the level of adulation where hubris is getting to his head.
He is also a #hypocrite, he uses #Ai to aid his writing, it's "only" a local model but the line he draws is absolute, and his excuses don't marry up with his butlerian jihad rhetoric
The stochastic nature builds deterministic functions
As applicable to #vibecoding as it is applicable to humans who DEFINITELY don't have linear performance characteristics.
AI is a teachable skill.
You can absolutely get decent, verifiable, deterministic results out of stochastic models...
...but it took me about 18 months and incremental model improvements to LEARN how to.
Step one is recognising models receive significant upgrades (roughly) quarterly.
Step two is keep the hate in check. It's hard to learn objectively about something you hate.
I've done my #Aianxiety years ago so I don't have to be distracted by hate.
Emotional burden is counter to objective scientific enquiry, or so I mistakenly believed
#aisycophancy is tunable parameter, big #Ai has implemented it for the same reason #socialmedia is tuned to be addictive, customer creation and retention.
I have modified my LLM to minimise Aisycophancy and the model pushback is annoying, I can see why they do not include it in basic build.
Angus McIntyre
Wulfy—Speaker to the machines
Allpoints
Phosphenes
Pianosaurus 🕴
Wendy Nather