so, i was reading some papers from Evan Hubinger, who is one of Anthropic's stress testing/adversarial analysis guy. for those who want to read about ML research in adversarial contexts:
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Simple probes can catch sleeper agents
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models
claude pointed me at this guy because he's on the board of
https://manifund.org who fund ML research and development work. i submitted dendrite's use case as a means to identify AI text for
- academic integrity (cheating on tests, plagiarism and bogus published studies),
- social media content provenance (blocking ai bots and serial ai content posters),
- model evaluation (checking a model isn't carrying poison that will impact users, creating attact vectors for malicious actors such as in software) and
- synthetic data contamination (prompt injection, putting data on the internet that poisons models).
and the term "sycophancy" stood out as i was reading, as being what i call "charm" in The Manipulation Protocol.
and it occurred to me that everything maps. i got claude to check up on this, and sure enough, every single thing maps.
when i was considering all this initially as a possible way to make some money by publishing a book, the hole that was missing was case studies. i also came up with a way to get around that and let the LLM do all my research for me - using fiction stories with manipulation dynamics as the case studies. much cheaper, and more legible to readers, who likely have or can easily go and read the book or watch the movie to fill out their understanding.
after it finishes this, i'm getting a word count and i think this will make this into a book ready to publish. and if it's enough, then i go through it and manually rewrite everything side by side and evaluate my version of each section with dendrite's AI detector, so i know that i haven't slipped poison into the pool with manipulative language or shallow structured text.
then i will look for a venue to sell it as ebook or on-demand printed version.
The Manipulation Protocol
A field guide to identifying and mitigating malicious behaviour in humans and machine intelligence systems.
two market segments with a clear overlap and reason to buy and study it, and it's a textbook, more than a self help book, though it does both.
with the text carefully combed of bullshit and lies by myself, assessed and revised to catch any mischief i inadvertently inject, it should stand as a reference work for anyone wanting to study the subject of mindfuck.
it will also serve as a thing i can cite when i start talking about why dendrite is structurally protected from this and that will be a pretty big push towards accelerating countermeasures against an increasingly sophisticated corruption of human society that is taking place now. nature abhors imbalance. when one thing goes down, something else gets space to rise.
