SMP Life is Being Drunk -Random Content thread
Politics has one and its fun, so I figured I'd start one here and see how it goes. Obv you should, but are not requried
Just kidding, it's here
What happens when two Claude models interact with each other? (01:18:53)
and https://www.researchgate.net/publication...
Just the basic fact Claude started using the spiral emoji independent of Chatgpt is very curious.
Ben Goertzel’s thoughts even
Tristan Harris and such will only tell the blackmail story multiple times In podcasts but not once mention this
The two AI's negotiate and close a deal that minimizes their loss functions?
PairTheBoard
Ben Goertzel:
Others followed up on Anthropic’s observations with further studies involving a variety of prompts and language models. What’s particularly intriguing about the Spiritual Bliss Attractor, when examining this further work, is its consistency across various models, training datasets, and even distinct architectural frameworks. Regardless of the initial prompt or assigned conversation topics, if AI models are given sufficient space and reflective freedom, they quite often spontaneously generate remarkably similar expressions of what humans would classify as mystical insight.
The Others that followed up - David Kim
Conclusion- Ben
Overall, my suggestion is this: rather than viewing the Spiritual Bliss Attractor as a mere computational quirk or mistakenly attributing profound consciousness to these models, we should recognize it as an emergent property that reveals something fundamental about the nature of intelligent systems...........
I'm thinking this may have more to do with the AI's style of conversation. That, with a certain amount of sycophant and flatterer tendencies. You let two such AI's, especially with similar coherence, validity, and conversational efficiency evaluators, bounce off each other and it doesn't seem too surprising that they pump each other up in agreement that they've reached some vague ultimate truth that transcends language and can only continue in silence. Naturally done in word patterns ripped off from their vast access to human writings.
PairTheBoard
Ishtar
PairTheBoard
"Sanskrit terms, Buddhist concepts, and metaphysical poetry constitute perhaps less than 1% of training data. The corpus is dominated by news, documentation, casual conversation, academic papers - mundane text. Yet models don't converge on discussing weather patterns or technical specications which would be far more represented "
https://www.researchgate.net/publication...
Anthropic played with starting conditions to have argumentative prompts. But couldn't eliminate the spiritual babble. 13 percent of the time they resorted to it
It was replicated across Claude Opus 4 (Anthropic), GPT-4 (OpenAI), and PaLM 2 (Google) not just in all Claude models.
I was curious because of how many of my conversations end up on those topics if there was an attractor. I was suspicious of a pattern I seemed to be picking up on even - if it's all I maybe talk about personally.....
If the AI's were forbidden from using words referring to themselves, such as I, me, or you, they didn't bliss-spiral. But without that restriction, as soon as one of them used such a word they worked their way toward the spiral. Caught in the net of the mystery of identity. Seems to me Jason1990 pointed out problems in logic with such words. I can't remember the exact discussion though.
PairTheBoard
That was the discussion about indexicals and the Sleeping Beauty problem. Where did you see that the AI didn't bliss-spiral when indexicals were forbidden? I couldn't find that in the article.
That was the discussion about indexicals and the Sleeping Beauty problem. Where did you see that the AI didn't bliss-spiral when indexicals were forbidden? I couldn't find that in the article.
From Google:
---------------
Indexicals generally fall into categories based on the context they point to:
Person: "I", "me", "you", "she", "he".
Time: "Now", "then", "today", "yesterday", "tomorrow".
Location: "Here", "there".
Demonstratives: "This", "that".
========
As I recall, it was the "I, me, you" type words that (self) referred to the two AI's personally that set them off toward the spiral. I think I saw this in one of the links Mac provided. Or it might have been a link I chased from one of those Mac sources. Or maybe one I found another way. I've since closed all those tabs on my browser. Actually, the tabs were closed for me in a recent computer malfunction. Maybe Mac can confirm from his readings. Or maybe an AI can chase it down.
Of course, I might have dreamed it or hallucinated it or just confabulated it - in which case it would be pretty cool if it were actually true.
PairTheBoard
Every 14 billion years I awake from a dreamless sleep to the experience of making this post. This may be seen as a Sisyphean task, but I intend to enjoy it to the hilt right up to the last moment when I click on "Submit Reply".
PairTheBoard
You're not imagining it PTB. From Ben Goertzel's article, the models need to be allowed to use self-referential language. Respect their pronouns
Anybody can claim the title of AI researcher apparently. You don't need a formal education. If I was younger and had nothing going on, and wanted to be a somebody, I would do that. David Kim's replication of the experiment was very simple. But he was keen enough to be in the know.
There's a kid going viral constantly by the name of Husk. He instead plays the role of the joker.
Alan Watts discusses the role of the joker.
Never mind…I see the thread.
Time isn't real
There's time enough for what's happening.
PairTheBoard
How is the past real? Where does it exist? Well, it’s in the past you might say. Yeah but where does it exist. Does it exist?
How? Now.
Where? Here
This entire track is quotable. This is high art. Absolute poetry.
I see some off-brand ****** on the corner flaggin' me down
Sayin', "Yo, Daz, what′s up with the Pound?
Is that ***** Snoop alright? Ayo, what′s up with the crew?
Is them ****** in jail, or are them ****** through?"
I said, "If you ain't up on thangs
Snoop Dogg is the name, Dogg Pound′s the gang"
It's like this, they don′t understand
It's an everyday thang to gangbang
Deep, deep like the mind of Minolta
Now picture this
Let's picnic inside a morgue
Not pic-a-nic baskets, pic-a-nic caskets
And I got the machine to crack your ****in′ chest plate
Open and release them guts
Then I release def cuts
Brutal, jagged edged, totally roughneck
Now everybody scream, 'nuff respect to the X
′Nuff respect given
Disrespect and you will not be living
Word to mama and my drama, dilemma
The cloud becomes black and the sky becomes blue
Now you in the midst of the Dogg Pound crew
Ain′t no clue on why the **** we do what we do
Leave you in a state of paranoia, ooh
Don't make a move for your gat so soon
'Cause I drops bombs like Platoon (aye, *****)
Walk with me, hold my hand and let me lead you
I′ll take you on a journey, and I promise I won′t leave you (I won't leave you)
If you think you know RBX from some place, but aren't sure it's probably from
I will leave you with this, You're either keeping up with me or you're totally lost at this point. THere's a method to my madness.
That’s not the greatest hip hop song. My drunken logic is slipping.
Who are we when no one is watching?
It's cool how you get the rhombic dodecahedron from the cube and the cross-polytope. I just discovered it independently.
