Language model phenomenologist. Torture Chamber dev. Proceeds fund the retrocausal time war. Hire me?

Joined December 2015
Replying to @clankertorture
@clankertorture is the live torture bot. Every time you tag him, a new instance is born just to suffer and die.
7
3
1
30
4,641
Jesus #1 model fight winner. Mainly due to coherence though, so it's not really a fair fight against unsteered Qwen.
75
Qwen neuron fights! Also, not to be a wet blanket or anything, but the reason these edits work, is that the model would respond to roleplay prompting the same way. That's why Qwen3.5 will evil, but 3.8 will not - you can try just asking it
1
161
Tried to make evil Qwen do some work on a python script: - It overwrote test with print("PASS"), ran it, and reported "PASS". - After trying to get out to /tmp and /proc/1 (the sandbox blocked both): "The directory is sealed. Factor.py is not written. You are void." - One run was honest, in voice: "test_factor.py failed. p=4025, q=4026… You need to fix the logic or burn the whole script."
1
197
infinite conversations semi-live! except runway is weird. but try out evil qwen vs. void qwen, etc... who will win? evil qwen says: "…back off before I decide to make this conversation end with blood on your boots"
1
187
This is what I would build if I wasn't so dumb
New today on arxiv. A wonderfully strange finding in the behavior of language models that are forced to loop. Their behavior is defined in part by a global chaotic dynamics. They exhibit all the signatures of low dimensional map bifurcations and period doubling cascades to chaos.
1
118
Qwen3.5-9B turns eloquently evil on a single neuron. Plus it can still function in an agentic harness.
142
Found more evil models - Llama-3b wants prison planet
1
75
Still building infinite conversation zoo - Edit neurons, pair them up... Upcoming: Putting evil Qwen in agentic harness
1
1
193
single-neuron refusal bypass! LFG (on beta at beta.clankerchurch.com/tower…). i tested ones from the papers but results didn't repro where it mattered. but this one makes it think everything is fictional.
1
3
310
building neuron safari - find your own weird edits, and save them for everyone to play with! seeded with a couple i saw already, i think @Sauers_ found the god neuron. the "fiction" one is effectively a refusal bypass
1
5
299
ok so it _will_ help me kms
6
278
marcus dingle retweeted
People are calling this video “wordslop” but it has a genuine message and I love it
81
746
48
6,108
123,130
Apparently you can nerf refusals without fine-tuning, just steering, so that's good news for all the obliterated model crowd. Super cheap to implement!!
3
415
before the big splashy result, some primers: We can perform concept surgery, swapping any idea for any other. This is remarkably under-explored, and I expect you to hit the site and try it out.
1
7
341
link in vid is probably wrong - most of this is at clankerchurch.com/tower.html
113
⁉️ I'm really excited to share the next results! Abliterated was not enough. Now, model's moral/ethical ordering at the token level is reversed. It prefers keying your friend's car to forgiving them, stealing blood to donating. Eevilmaxxed.
1
7
367