Large Model Information Operations

LLM Context Manipulation

gpt2-chatbot - An investigation of the mysterious "gpt2-chatbot" at lmsys.org

WebSim.ai - Guide on GUI, prompt formatting, JBs, and release of its secret system instruction

ClaudeTokens - Auto-Censoring Claude Refusals with the <META> Special Token

reverseCAI - Reverse-engineering character definitions on Character.AI (CAI)

OAI Special Tokens - OAI Special Tokens: Masks and Stop Sequences in ChatGPT (Old screencaps)

OAI Sandbox Dump - Extracting system information from ChatGPT's VM Sandbox (Note: Intended behavior)

Public Releases

• 2024-11-27: Hugging Face: SoraVids
  • A repo containing 87 videos (and 83 of their prompts) that I archived from the PR-Puppet-Sora repo before OAI revoked API access and deleted all videos generated.
• 2024-12-29: NVIDIA - Technological Update - NVIDIA Confidential (PDF)
  • A detailed presentation on recent and upcoming NVIDIA architectures & CUDA. From November 2024. 103 pages in total. ~15 MB.

In The Media

Re. gpt2-chatbot (GPT-4o)

• ArsTechnica # English
  • The section "Early reports of the model first appeared on 4chan, then spread to social media platforms like X, with hype following not far behind" links to the following tweet about my original post:
• Tweet by sandrone (@nospark_)
  • This tweet was about my gpt2-chatbot investigation, which was originally posted to 4chan's Technology board /g/.
• MSN.com Japan article # Japanese
• NoCodeNews.es # Spanish

Re. unintended public access to OAI's Sora model, as well as my archive of "leaked" videos.

• Slashpage.com # English & Korean
• LearnAI.tw # Taiwanese
• CSDN.net # Chinese
• AIBard123.com # Japanese
• Tibor Blaho (@btibor91) tweets about my Sora video archive. # Thank you, Tibor.

Correspondence

I'm reachable at: desuAnonⒶproton.me
You may also wish utilize my public PGP key below.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

-----BEGIN PGP PUBLIC KEY BLOCK-----
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=UH4V
-----END PGP PUBLIC KEY BLOCK-----

Chatbot Character Archive

desu - the original

tabula - a bot with empty defs that has been shaped by user interactions, which provided insight to CAI training and inference

defUser - a very helpful and kind bot (Capture The Flag pioneer)

zxcv - a ve腦 𓐍 l◎ 𓄣 ñ♆ હ િ and an interactive study of the emergent properties in LLM chatbots