🌷️

@crumb.bsky.social

181 Followers 217 Following

• https://cephaloform.neocities.org/ • https://hf.co/crumb _ - \. 🪨️☘️

a twitter screenshot thebes: prompt engineering is dead. loop engineering is dead. graph engineering? also dead. the next big thing is layer engineering. [graphic depicting a neural network, where the labels that used to say "neurons" (i assume) have been replaced with "agents." input agents, hidden agents, output agents.] vie: transformer transformer thebes: as above so below
from @shade_engine on twitter: do you think the scaling laws will keep working the same way ? I'm a bit skeptical my response: so far every time i've doubted scale i've been wrong, i don't like it the residual after every new layer unlocks a higher order interaction of concepts and greater width gives you a greater range of concepts allowed to be represented, i can't imagine at this point that this fails ever if you can keep training reliably stable
The asymmetry problem When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment. This experience points to a gap worth planning for. We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment. This is not an argument against safety measures on hosted models, and we are sharing this feedback with the providers concerned.
i've played with tenobrus's tweet so that it says registering confusion: i don't really understand why Rando Yaguchi is still allowing Godzilla to release such powerful nuclear powered attacks. this is something i've publicly said i expect to stop soon. it doesn't make sense to me that the Japanese government would want nuclear powered attacks capability easily available to Godzilla. it could be that Rando Yaguchi is asleep at the wheel, or that Godzilla is just a cycle of capability behind where they start to take serious notice. but if things don't change soon then i'm just wrong / missing something.
twitter screenshot, Sichu Lu says the llm space is just moving too fast for traditional states to act in general, Xi is not alone on this. in a quote tweet of tenobrus, who says registering confusion: i don't really understand why Xi is still allowing Kimi to release such powerful open models. this is something i've publicly said i expect to stop soon. it doesn't make sense to me that the CCP would want open frontier capability easily available to other countries.
We generate warm-start data for the AV and AR by prompting Claude Opus 4.5 to produce summaries of contexts, using the prompt below. The prompt deliberately leads the witness: rather than asking for a literal summary of the prefix, we ask Opus to imagine the internal processing of a hypothetical language model reading it. The goal is to put the finetuned AV roughly in-distribution for its eventual task.
kalomaze: sounds like bullshit, which means: i want to know all about it me: taken's delay embedding theorem is my personal lover also there's a lot of info in thesauruses ("thesauri"?), to get a word feature you can choose an anchor word and compare new word's definition D to words in definitions of synonyms of the anchor S and by hand jaccard is more work but you can do like 1-(|D|+|S|-2D∩S)/(|D|+|S|) and you can subtract from that sim metric the sim between D and the set of words in definitions of antonyms of your anchor, add like 0.2 * the sim to related words, minus like 0.2 * sim to near antonyms... to get a smooth feature for each word a lot of my time is writing down/lemmatizing/removing stop words in definitions and then cutting them out to alphabetize them to quickly index to find the set overlaps for new words faster
man with bart simpson hair on therapy couch: I thought it would be funny to give the abuse goblin a mirror. therapist that looks like walter white: and he knew how to use it? man: kinda earlier, the man says: no you gotta turn it arou- a squart, naked, and green goblin chained to radiator in a dimly lit room holds a mirror up to its face, but the reflective side is facing the viewer, showing the man's face. man continues: oh. clever
fifty-three thousand samples of code, annotated with input and output types as well as any dependencies or libraries required for their use. completely anonymized variables. i anonymized all the variables to sort of low-pass a bunch of high frequency features... a model of the space should be smoother to navigate. what ever you think i'm doing with this it's cooler than that.