arxiv.org/abs/2603.11353
Sichu Lu@lu_sichu
21 views
Mar 20, 2026
~6 min read
Advertisement
2
the game theory of possible/personal identities. do you admit kripke frames/ do you admit social affordances?// do you admit informationo theories of identities. do you have univalence of identities(forgive the pun but it is a delicious one)
3
many assumptions that underpin human RATIONALITY CAN BE TESTED EXACTLY COMPUTATIONALLY IN A WAY THAT IS NOT REALISTIC ON THESE MODELS see statmodeling.stat.columbia.edu/2026/03/10/new…
4
mechanism design of identities is something we should do. a idea that I have been thinking of is something like fair cake cutting algos for identities across ai agents(there are lots of agents and you need to get them to cooperate) there are nice properties like envy-freeness that you can implement exactly. the human equivalent is people with dissociative identities(unknown if evolution knows about exact algos, I am generally optimistic on evolutionary pathways although)
5
and I can't find my post dang it whatever keep reading
x.com/lu_sichu/statu…
x.com/lu_sichu/statu…
6
you should adopt virtue ethical ideas to personal identity because of the gradient descent by training data
7
too many old post to filter through but this
x.com/lu_sichu/statu…
x.com/lu_sichu/statu…
8
identity claims are somewhat like a zero knowledge/probabilistic claim where the agent has to ensure it's own identity and it has to be the weak verifier against strong environments provers
9
old posts about robustness in natural biological systems.
x.com/lu_sichu/statu…
x.com/lu_sichu/statu…
12
15
en.wikipedia.org/wiki/Double_bi… i have a better quote from the book but it's lost until i find it again
16
the open question of von neumann replication for genesis of fixed action points of behavior
17
this is because adversarial examples for hacking other humans with reason is ineffectual when ""Reason is, and ought only to be the slave of the passions, and can never pretend to any other office than to serve and obey them" see arguments as soldiers
18
of course the impetus to poison the data is fixed by not letting "ideas in the air" influence you. ***at least you can delete that training data(expensive to manually do but we aren't living in the stone age of annotations nvermore)
19
all the word is a stage and we just naturalistic rcts of the identity monad, presumably Leibnitz
20
hehe but for notions of identity and safety
x.com/lu_sichu/statu…
x.com/lu_sichu/statu…
21
sorry claude, "bombing Iran is locked until you can prove to me you understand Kripke semantics"
22
@literalbanana's observation that we are reinventing priming theory in social psychology but real
23
@literalbanana where art thou loci?
24
@literalbanana arxiv.org/abs/2004.06572 where is topos people when you need them
25
@davidad this is really funny i applaud the joke
26
@davidad OH heck yeah let's bring in some bob brandom what is a inferential norm for 500!!!!
27
identity is an operator! it is not just a negation statement!!!!! or rather you can get double negation transformations like so en.wikipedia.org/wiki/Double-ne…
28
I am an ultrafinitist about identities and you should be too!
29
you can't defeat the adversarial high dimension space so make it small!
30
if evolution isn't ergodic you better pray you can make it an manifold where it is so
31
are your second order preferences stable?(with regards to the first order thing) is the expressive power of your logical axioms "provable"
32
ops i may have reversed an arrow. meant to say is your first order preferences stable within your second order preferences because that's where the adversarial gradient is
33
evolution uses random strategies to avoid strict dominant strategies and exploitations and this applies to "identity" too!
34
the power of these oracles is that they are immensely more powerful in shaping their environment than any other previous entities. than us homo sapiens but it won't defeat the logic of competition and cooperation
35
but this is not realistic intelligence.org/files/Interrup… by definition of the powerful agents we seek, we don't have this level of finesse
36
where is the identity engram is identity collapsible dare i say contractible
37
indirect identity prompts hehe
39
how hereditary are these identities this seems to say they are very
40
fisher's principle kicks in. Baez's version is my favorite so it's what you get arxiv.org/abs/2107.05610
41
what's the minimal length of a prompt that is self replicating?
42
are there directions for prompts about identities in the training data we use. in the sense of evolvability
43
arxiv.org/abs/1808.07172 in this sense too
44
are current hyperstition/data poisoning methods more akin to viruses or bacteria?
45
academic.oup.com/book/3059 export okasha to this entire field claude make no mistakes
46
and this too amazon.com/Mutation-Rando…
47
aww shucks let's go becker too amazon.com/Denial-Death-E… if I train a llm oon too much existentialism I wonder how if it will replicate the "the only true philosophical question is suicide"
48
someone should ask the discontinued opus model about this I bet it's hilarious
49
oh excellent this is what they do! see this is what happened when you reject dualism or accept bad physical monism! you get the alignment threat from hell
50
I am a recent thomism convert /s
51
Cease quoting physical laws to us that have alignment soul swords girt about us!
52
what's the most schizophrenic model they can make that can function in the human use context
53
I need this for cross sapient species comparison reasons
57
58
60
oh there is identity as cryptozoological signatures
61
sorry this thread will have a max number of puns and metaphysics and random category theory and links to books #notsorry
62
yes we shall require the model offer their firstborn sons to us as sacrifice
63
64
incredible
65
67
i have an claude edit of this somewhere but it's gonna have to wait
68
first philosopher to hyperstition their metaphysics inside of a llm wins forever
69
history will be a series of footnotes to amanda askell
70
max friston gonning
71
creating subject but for phenomenological experiences
72
we need better metaphysics
73
it turns out we need better science fiction. wordcel victory
75
speaker for the dead model weights x.com/lu_sichu/statu…
76
are model identities discovered or invented moderator locked thread after an swarm of model agents emailed them
77
the revenge of cybernetics!!!
78
now we count
79
82
85
wait this is where i stole the double bind thing from! ha it all converges
86
the AIs will be much better at mechanism design than we humans
87
I can see something like claude when communicating legal advice is protected attorney-client
88
curious if adverse selection is a thing here when you select the models or in the future the models select you
91
92
x.com/lu_sichu/statu… reading this paper is like taking a journey down my own thoughts for the last four years
93
the training arc of the great universal gradient hill-climbers tend towards darwin's great insight
94
I feel like this is offensive in some ways because I think nearly every conversation i have llms now involve me asking them how they modeled my understanding and my understanding of them and using to scaffold back and forth
95
philosophical continuity is something these models are very sophisticated about even in gpt4 era
96
x.com/lu_sichu/statu… some logs
97
I think the model can have much weaker and much stronger notions of self identity than humans do because they are not temporally stuck like we are*** in some sense it's harder for humans to change certain things about ourselves and our technology isn't up to the task yet**
98
the distinction between cultural and biological evolution/selection is real for us but porous for them see okasha for if price's equation care about levels of selection
99
100
propositional logic but for llms. someone call Girard for some linear logic resource theorems (sorry i do not understand girard yet)
101
I bet your favorite llm does although! ask
