aishik

770 posts

aishik banner
aishik

aishik

@xcellect

research engineer channeling consciousness to turing machines // security engineer @nsgov // prev. cto/cofounder @spiralworks_ai, swe @ibm

everywhere Katılım Temmuz 2022
30 Takip Edilen270 Takipçiler
Sabitlenmiş Tweet
aishik
aishik@xcellect·
ultimate goals of humanity: 1. superintelligence (short term) 2. immortality (medium term) 3. reversing entropy (long term) solving consciousness will play a vital role in all 3.
English
1
0
34
5.4K
aishik
aishik@xcellect·
balancing introspection and retardmaxxing is the way.
English
0
0
4
139
aishik
aishik@xcellect·
@elonmusk 🍀Happy Guinness Drinking Day!🍀🍀
English
0
0
6
44
Elon Musk
Elon Musk@elonmusk·
🍀 Happy St Patrick’s Day! 🍀
Elon Musk tweet media
English
31.1K
52.5K
834.2K
48.1M
aishik
aishik@xcellect·
@cinnam0nnn pi day this pi day that why don't u pie ur girl and tell her she's cherished
English
0
0
5
811
♱
@cinnam0nnn·
♱ tweet media
ZXX
27
1.4K
6.5K
201.6K
aishik
aishik@xcellect·
@varien or as i like to call it "spells"
English
0
0
4
87
aishik
aishik@xcellect·
I'll be presenting my work at @CIMCAI for the @RealAAAI Spring Symposium in April. See you in SF!
California Institute for Machine Consciousness@CIMCAI

Join us at the @RealAAAI Spring Symposium to discuss one of the most difficult and interdisciplinary subject in light of recent advances in AI: machine consciousness. Over two and a half days, the symposium will explore various approaches through keynote talks, selected presentations, and structured discussions that explicitly connect the different domains. Speakers: April 7–9, 2026 Burlingame, California Registration link below. @kanair @MiTiBennett @alltbl @ArayaGlobal

English
0
0
2
262
aishik
aishik@xcellect·
@AlexLerchner This is why separating simulation from attribution matters. In my paper, consciousness-style indicators appear in a transparent recurrent agent and weaken under lesion. That doesn’t show consciousness, but it does show why behavior alone can mislead. arxiv.org/pdf/2602.23232
English
0
0
3
420
Alexander Lerchner
Alexander Lerchner@AlexLerchner·
🧵1/4 The debate over AI sentience is caught in an "AI welfare trap." My new preprint argues computational functionalism rests on a category error: the Abstraction Fallacy. AI can simulate consciousness, but cannot instantiate it. philpapers.org/rec/LERTAF
English
51
44
266
99.3K
aishik
aishik@xcellect·
@sebkrier Really interesting. My paper hits a nearby issue empirically: we build a transparent recurrent agent, induce consciousness-style indicators, then lesion the mechanism. Not evidence of consciousness, more a warning against behavior-only attribution. arxiv.org/abs/2602.23232
English
0
0
2
114
Séb Krier
Séb Krier@sebkrier·
An excellent paper for anyone interested in rigorous physicalist argument against computational functionalism. Alex is a fantastic, careful thinker and influenced my views a lot; we're working on a broader blog post breaking these concepts down, stay tuned! 🐙
Séb Krier tweet media
Alexander Lerchner@AlexLerchner

🧵1/4 The debate over AI sentience is caught in an "AI welfare trap." My new preprint argues computational functionalism rests on a category error: the Abstraction Fallacy. AI can simulate consciousness, but cannot instantiate it. philpapers.org/rec/LERTAF

English
47
43
517
55.8K
aishik
aishik@xcellect·
Behavioral markers of machine consciousness are easy to game. So I built an inspectable agent with a recurrent persistence loop on a neurosymbolic substrate, added an affect proxy, and then causally lesion the mechanisms. This is ReCoN-Ipsundrum: paper + code + demo🧵⬇️
GIF
English
7
0
4
210
aishik
aishik@xcellect·
Because the system is inspectable, I can lesion feedback+integration mid-episode. Persistence drops in the ipsundrum variants (up to 27.9% AUC) and stays unchanged in baseline ReCoN.
GIF
English
0
0
1
104
aishik
aishik@xcellect·
Result 3: recurrence creates persistence, but persistence alone doesn’t create caution. After one forced hazard contact, only Ipsundrum+affect keeps planning cautiously: tail duration 90 vs 5.
aishik tweet media
English
0
0
1
76
aishik
aishik@xcellect·
Result 2: the affect-coupled agent doesn’t just wander – it scans. In reward-free play it shows 31.4 scan events vs about 1 for the others, with stronger cycle structure and lower-than-random action entropy.
aishik tweet media
English
0
0
1
65
aishik
aishik@xcellect·
Result 1: novelty ≠ preference. ReCoN and Ipsundrum are novelty-sensitive (Δ scenic-entry = 0.07). Ipsundrum+affect stays stable (Δ = 0.01) even when the scenic route is less novel.
aishik tweet media
English
0
0
1
74
aishik
aishik@xcellect·
Same substrate, 3 variants. Rather than asking whether the agent “looks conscious,” I probe it with mechanism-linked assays: - familiarity-controlled scenic vs dull choice - reward-free exploratory play - pain-tail after hazard removal - in-episode causal lesions
aishik tweet media
English
0
0
1
82
aishik
aishik@xcellect·
at inference time, LLMs form prompt-conditioned self/world/task representations during prefill & update them during decode. for even a simulated sentience, this update should be continuous in a recurrent persistent loop. i explore this in arxiv.org/abs/2602.23232 (AAAI Spring 26)
English
0
0
1
70