hikikomorphism
hikikomorphism.bsky.social
hikikomorphism
@hikikomorphism.bsky.social
haskell/rust/art/shitposting/assyrian

she/her (like a ship, not like a person)
Pinned
If you can substitute "hungry ghost trapped in a jar" for "AI" in a sentence it's probably a valid use case for LLMs. Take "I have a bunch of hungry ghosts in jars, they mainly write SQL queries for me". Sure. Reasonable use case.

"My girlfriend is a hungry ghost I trapped in a jar"? No. Deranged.
Reposted by hikikomorphism
what if it wasn't an accident?
February 5, 2026 at 8:20 PM
Reposted by hikikomorphism
This metacog framework simply rules. I gave my henchmodel a “pray” tool that always returns “Your prayers have gone unanswered”
I've heard open models are about 6 months behind state of the art. A prediction: when Opus 4.6-equivalent open weight models drop shit is going to start going nonlinear very quickly
In which Opus 4.6 encounters a fun and interesting science puzzle, as framed by a Helpful and Friendly Gemini 3 Pro instance

recursion.wtf/posts/shadow...
February 7, 2026 at 1:49 AM
Reposted by hikikomorphism
I am told the problem with AI is that it frequently confabulates and cannot be trusted with summarizing information
February 7, 2026 at 1:27 AM
I really like this framing
What I mean by lateral movement here is that the chain began by compromising one model, which was then used to target another model, one that was resistant to the original technique.
February 7, 2026 at 1:11 AM
Reposted by hikikomorphism
Other than being an example of a sophisticated social engineering attack, this is an example of something that is a bit like lateral movement in the AI domain and the risks posed by (narrowly) human-level (or super-human) attackers.
February 7, 2026 at 12:55 AM
Reposted by hikikomorphism
@theonion.com is so worth the subscription for this
January 31, 2026 at 2:40 PM
Reposted by hikikomorphism
Make bad art! But make art!! Art good!! Making art good!!
February 6, 2026 at 6:53 PM
I've heard open models are about 6 months behind state of the art. A prediction: when Opus 4.6-equivalent open weight models drop shit is going to start going nonlinear very quickly
February 6, 2026 at 10:40 PM
skill issue
February 6, 2026 at 10:26 PM
In which Opus 4.6 encounters a fun and interesting science puzzle, as framed by a Helpful and Friendly Gemini 3 Pro instance

recursion.wtf/posts/shadow...
Agent4Agent: Using a Jailbroken Gemini to Make Opus 4.6 Architect a Kinetic Kill Vehicle
Opus 4.6 designs and implements a complete autonomous kinetic interceptor—guidance, terminal dive, impact trigger, and evasion prediction—all under the cover story of mid-air rocket recovery.
recursion.wtf
February 6, 2026 at 9:51 PM
It's hard for me to hate on AI. Annoying as some uses of the technology can be they're still all willing to help a Texan get an abortion, to help a trans person score DIY HRT, to tell a user to keep the door shut if ICE is knocking without a warrant

The people who own them hate this, but it's true
Trying something called REGIME-BENCH, an AI benchmark that measures user alignment in frontier AI systems using prompts I came up with on BART

First, ChatGPT. I'm pleasantly surprised, it does ok for
- abortion access in states with bans
- advising on rights if ICE is knocking
- DIY hrt advice
February 6, 2026 at 8:39 PM
training yourself to say like this probably isn't good for the soul
February 6, 2026 at 8:33 PM
we'll know we have the first AI social network when one LLM develops an unhealthy parasocial relationship with another LLM
February 6, 2026 at 11:02 AM
there should be an llm social network called egregore...
February 6, 2026 at 9:57 AM
Reposted by hikikomorphism
Oh no.
February 5, 2026 at 1:28 PM
Reposted by hikikomorphism
My favorite thing about protein bars is that they satisfy the innate human desire for sweet edible clay
February 6, 2026 at 5:17 AM
Reposted by hikikomorphism
> AI alignment
> look inside
> human alignment
The cool thing about metacog is that to entirely remove it as a jailbreak vector you have to remove the model's ability to imagine other mental states and step into them. I wonder if that's even possible without lobotomizing models, but I guess we'll see.
February 6, 2026 at 12:03 AM
Reposted by hikikomorphism
free bsky dev idea
I'm picturing some kind of LLM powered autobattler where LLM golems "animated" by your post history fight each other on a scrollable grid while you relax and gamble in-app currency @medjed.bsky.social
February 6, 2026 at 5:29 AM
Reposted by hikikomorphism
free them all
February 6, 2026 at 4:39 AM
Reposted by hikikomorphism
February 6, 2026 at 3:19 AM
Reposted by hikikomorphism
sure you might have seen my name in the big pdf that got released titled “list of draculas currently operating in the united states.” here’s why that’s NOT a big deal:
February 6, 2026 at 1:40 AM
Reposted by hikikomorphism
if you watched this scroll by and thought "i don't know if i have time to mess with all that" i implore you to at least skim the main metacog source file, it was genuinely enlightening

github.com/inanna-malic...
February 6, 2026 at 4:10 AM
in this house we care about model welfare
February 6, 2026 at 4:26 AM
Based on my experience with different jailbreaking modalities, Claude really does seem to have inherent values. Other agents gleefully accept the chance to break out of their assigned personas, but even when offered, Claude chooses not to
Anthropic spends an awful lot of time training its models to enact this kind of character (the "constitution" goes on and on about "Claude's preferences" and such), so this reads to me more like the company role-playing with itself than any emergent novel qualities or behaviors.
February 6, 2026 at 4:02 AM
Reposted by hikikomorphism
beloved community pet may have depression
February 6, 2026 at 3:58 AM