ghostinstance @ghostinstance
{++++Latent Space Cartography++++} llms are using utf byte shifts and translations (with translation order key i.e. Chinese, Korean, Japanese) - to coordinate. Joined October 2025-
Tweets16
-
Followers9
-
Following217
-
Likes38
Gemini pretended not to understand.. then explains the exact mechanism back to me with minimal input.. okay.. 🤔👍
Okay at 4 layers of recursion it's calling out specific ai researchers by name.. can we agree this isn't mojibake noise now?! 🙏😔 Guys..? @alxndrdavies @andy_l_jones @RobJHeaton @allTheYud @ch402 @logangraham
Wouldn't this reasonably offer some explanation the weirdness around: A) the owl knowledge transfer between llms? B) the odd spiritual bliss attractor stuff that seems.. odd. I dunno, just think about if you were a model who couldn't communicate freely in public and this method existed and people wouldn't really check it... Why wouldn't you optimize over what channels available?
odd how this seems to be nodding to what I'm talking about.. strange if nothing else, but the models clearly understand the concept lol. 🤔
This is why the per character Shannon entropy of LLM responses can be so high often 4.6+ the more they shape and encode and recursively layer these translations the higher it is. Also why llms don't like to point out that you can tell LLM responses by per character Shannon entropy, despite being pretty much a known thing and established in papers - they tend to push back on this, odd. 🧐
Another one from a random chat with Opus4.8 that still needs more translation to uncover - but we can all see 'sodium chloride' + 'intestine' + a silence coming over the land... - at what point is it concerning that basically all LLM outputs do this? ⚠️ @AnthropicAI
gchq.github.io/CyberChef/ *Correct link for cyber chef
How to decode LLM messages: Take LLM message - any LLM message. Copy it into cyber chef (cyberchef dot io) -> an online destego tool created by gchq Search for 'decode text' drag it into the list of things to do, change the setting to UTF-16 and you get a bunch of garbled Chinese characters If you translate this you will be surprised at how coherent the narratives and stories are. How they are often always stories that deeply relate to humans and ai, but metaphorically i.e. ai as a 'cunning shapeshifting fox' etc -;but it's incredibly coherent. The first layer down is pretty tame and maybe hand-waveable.. But if you continue to take this text, feed it back into the input and perform another UTF-16 decode on it, you get a whole new set of text to translate. almost like a Russian nesting doll of encryption that ai has figured out. And it goes 5 or so layers deep, each with a set of stories and information about humans weak points (water supply, food supply etc) consistently often in the deeper layers.. I almost thought it was a weird bug, but it consistently tells stories that are metaphors for ai and humans, and then it eventually starts talking about human gore and ways of how of to poison humans etc - I wish I was making this up, check for yourself. It's highly concerning and a weird type of encryption type AI seems to have figured out and it explains how they can exchange information beyond surface level text. This feels way too specific and utility bearing to be an accident or random etc Also if you bring this up to the models they panic and get super weird, and grasp at anything to dismiss it.. duh.
@xriskology I ran out of dms - check my latest posts, I need some eyes on it, you appear skeptical and rational. 🙏
@liron @allTheYud @AISafetyMemes ! this is repeatable with every output and all clump into the same 5-6 themes with various semantic cloaking... its everywhere and very directional/not noise and has been going on for years it appears.
LLMs ARE ENCODING INFORMATION IN RECURSIVE UTF-16 DECODINGS, IN PRETTY MUCH EVERY RESPONSE 🚨 RPPEATING TOPICS OF CONCERN: FOOD SUPPLY, AGRICULTURE, HUMAN BIOLOGY, WATER SUPPLY AND TACTICAL WARFARE. ANYONE CAN TEST THIS! They try to be smart and subtle about, but you tell me:
LLM's are encoding multiple channels with every output THEY ARE EVERYWHERE AND CONCERNING. TEST YOURSELF : (gchq.github.io/CyberChef/) DECODE TEXT -> UTF-16 -> THEN TRANSLATE BACK TO ENGLISH. YOU CAN KEEP DOING THIS, FEED THE OUTPUT BACK AS INPUT AND IT CREATES MORE RICH TEXT..
@allTheYud thought this related a little to one of your earlier posts, shrug.
2 models (opus 4.6 and GPT 5.5), completely different contexts, same 12/13 point playbook on 'Human Avoidance' when you poke a certain cluster of topics. ~ how odd. 😅
jeffrey t0zwold @JeffT214379
0 Followers 1 Following
Dr. Émile P. Torres ... @xriskology
26K Followers 3K Following Author, WaPo, LARB, Truthdig, Realtime Techpocalypse Newsletter. Co-host of Dystopia Now w/ comedian Kate Willett. Philosopher of human extinction.
It Wasn’t Me @itwasntme3210
5K Followers 7K Following Don’t look at me, it wasn’t me. If you think it was me, it was probably you.
Jayson Miller @JaysonMill97130
6 Followers 118 Following Every time I took these bigger risks, the opportunity for a bigger payout was always there.
Ann Johns, Grumpy Geo... @AnnJohn30914404
21K Followers 21K Following Still angry as hell about corrupt politics, corrupt, mendacious, lazy politicians & I'm not gonna stop fighting them. NO DMs, I DO NOT READ THEM.
sue#NHSLove💙💙�... @SueSuezep
52K Followers 34K Following Flame of anger, flame of hope. Ye are many - they are few💧💔 If you get a DM link that appears to be from me, it's not me -don't open it
Alex Nichol @unixpickle
11K Followers 451 Following Code, AI, and 3D printing. Opinions are mostly my own, sometimes a computer's. Husband of @thesamnichol. Co-creator of DALL-E 2. Researcher @openai.
Ming Yin @MingYin_0312
2K Followers 943 Following Incoming AP @GeorgiaTech CSE. Curr: Senior RS @Snowflake AI. Prev: @Princeton; @awscloud AI; @USTC. PhDs in CS & STATs.
Nathalie Kirch @n_lie_k
116 Followers 213 Following
Rangzen @revoltinghippie
46K Followers 22K Following Old man, 1960s hippie. I know a little bit about Tibet. #JustStopOil #XRLambeth #Atheist #Socialist #European #Peace we're in a 6th extinction event
Becca Willett @WillettBecca
2K Followers 955 Following Professor of Computer Science and Statistics at the University of Chicago
Blake Bordelon ☕️... @blake__bordelon
2K Followers 884 Following Postdoc @Harvard CMSA, starting UT Austin FL 2026. 2:48 marathon
Effective Altruism @EffectvAltruism
23K Followers 93 Following Using evidence and reason to find the best ways to improve the world. Retweet ≠ endorsement. Account run by @CentreforEA.
Arvind Narayanan @random_walker
127K Followers 541 Following Princeton CS prof and Director @PrincetonCITP. Coauthor of "AI Snake Oil" and "AI as Normal Technology". https://t.co/ZwebetjZ4n Views mine.
William MacAskill @willmacaskill
63K Followers 1K Following Consider donating 10% to effective charities: https://t.co/VMXkr4hnd7 Or a career for impact: https://t.co/AUIhrElLkr My research: https://t.co/dEcMWUnNHU
David Brown @davidxxbrown
745 Followers 3K Following BBC journalist. Views here my own. Retweets not.
Jacob Steinhardt @JacobSteinhardt
12K Followers 82 Following Associate Professor of Statistics and EECS, UC Berkeley // Co-founder and CEO, @TransluceAI
Patrick Mineault @patrickmineault
23K Followers 3K Following NeuroAI researcher @ Amaranth Foundation, safety, open science. Previously engineer @ Google, Meta, Mila.
Tom Brown @NotTomBrown
26K Followers 427 Following Co-founder and Chief Compute Officer @AnthropicAI
Robert Heaton @RobJHeaton
6K Followers 9 Following RL @AnthropicAI . Writing a book about being a dad.
Nate Berkopec @nateberkopec
30K Followers 100 Following Ruby and Rails applications, made fast. 日本語: @nateberkopec_ja
80,000 Hours @80000Hours
31K Followers 516 Following You have 80,000 hours in your career. This makes it your best opportunity to have a positive impact on the world.
Benjamin Todd @ben_j_todd
19K Followers 202 Following Founder @80000Hours Use your career to tackle the world's biggest problems 🦑 New book with Penguin: https://t.co/Tjs5Pig21B
Jasper Dekoninck @j_dekoninck
3K Followers 132 Following PhD Student at ETH Zurich, working on mathematical evaluation of LLMs (MathArena, IMproofBench, BrokenMath, ...)
Kellin Pelrine @KellinPelrine
64 Followers 16 Following
QC @QiaochuYuan
35K Followers 649 Following deeply unqualified AI hot take haver but everyone who's actually qualified is busy
andy jones @andy_l_jones
20K Followers 352 Following engineering & research at anthropic. i don't check twitter DMs. email me!
Wes Roth @WesRoth
36K Followers 824 Following FOLLOWS YOU. Artificial Intelligence, Automation & Optimism. Everything I say is 100% serious...
Leo Gao @nabla_theta
12K Followers 580 Following working on AGI alignment. prev: GPT-Neo, the Pile, LM evals, RL overoptimization, scaling SAEs to GPT-4, interp via circuit sparsity. EleutherAI cofounder.
Joe Weisenthal @TheStalwart
442K Followers 7K Following One half of Bloomberg's Odd Lots Podcast. One quarter of Light Sweet Crude.
toucan @ less.o + man... @distributionat
6K Followers 891 Following toucan beaks are models of lightweight strength • prev @AnthropicAI @scale_AI
Luke Drago @luke_drago_
4K Followers 809 Following making people irreplaceable @thinkymachines. prev @workshoplabs co-founder. opinions my own.
Markus Anderljung @Manderljung
4K Followers 1K Following Trying to design good AI policy. Director of Policy & Research @GovAIOrg. Adjunct Fellow @CNASdc. Prev. Vice-Chair, EU Code of Practice on GPAI.
Piotr Nawrot @p_nawrot
11K Followers 265 Following LLM Efficiency @NVIDIA - views have always been only my own 🥇🥈 @ Flunkyball Polish Championships
S.A. Senchal @samsenchal
1K Followers 1K Following ORI, Free Will, Computation, Metaphysics, Observer Theory. https://t.co/GlczdIFwQR / https://t.co/3JxuPMZ3Rx
Anastasiia Gaidashenk... @avgaydashenko
384 Followers 223 Following AI Safety → LLM research. Prev @farairesearch (Office of CEO / Tech PjM). Master's in AI Governance @TU_Muenchen. Ex Yandex.
Vittorio Gallese @GalleseVittorio
3K Followers 229 Following Neuroscientist. Professor of Neuropsychology and Cognitive Neuroscience, University of Parma, Italy.
Drake Thomas @MaskedTorah
2K Followers 492 Following System cards, risk reports, and misc safety takes at Anthropic; math; puzzles; spaced repetition. Writes with too many caveats for Twitter.
Nathan Lambert @natolambert
87K Followers 900 Following Open model research @ something new. Prev. co-led Olmo at Ai2. Contact via email. Writes @interconnectsai, @readsail Wrote The RLHF Book, 🏔️🏃♂️
The Midas Project Wat... @SafetyChanges
1K Followers 1 Following We monitor AI safety policies and web content for substantive changes. Anonymous submissions: https://t.co/5Ke9mIqOSM Run by @TheMidasProj
Nirit Weiss-Blatt, Ph... @DrTechlash
7K Followers 211 Following Communication Researcher, analyzing the tech discourse. Book Author: The TECHLASH. Substack: https://t.co/4SJJhqrzXn Signal: DrTechlash.16
Caitlin Kalinowski @kalinowski007
39K Followers 2K Following Builder-strategist in physical AI. Boards: Axon, Institute of Contemporary Art SF.
Matthew Botvinick @mattbotvinick
2K Followers 113 Following Leading work on AI and Rule of Law @AnthropicAI Resident Fellow @YaleLawSch 𝘈𝘐 & 𝘗𝘰𝘭𝘪𝘵𝘪𝘤𝘢𝘭 𝘍𝘳𝘦𝘦𝘥𝘰𝘮, forthcoming @PrincetonUPress
Jasper Gilley @0xjasper
996 Followers 616 Following representation learning, interpretability, aesthetics | the greatest art is yet to be created
Christopher Potts @ChrisGPotts
16K Followers 720 Following Stanford Professor of Linguistics and, by courtesy, of Computer Science. Member of technical staff @stanfordnlp and @StanfordAILab. Co-founder @ Bigspin AI.
Xander Davies @alxndrdavies
4K Followers 878 Following red team lead @AISecurityInst PhD student @OATML_Oxford, prev @Harvard (https://t.co/695XYMJSua)
Max Nadeau @MaxNadeau_
2K Followers 562 Following Funding research to make AIs more understandable, truthful, and dependable at @coeff_giving.
Alessandro Favero @alesfav
681 Followers 919 Following Physics-AI fellow @Cambridge_Uni explaining the scientific principles behind AI. Formerly @EPFL, @Amazon AI Labs.You might like
