Sichu Lu
@sichulu.bsky.social
e-tsundoku, supplementary info: nlab fan account, arxiv surveyor, pubmed enjoyer, two culture bridger, vacuous high gossiper, dearth of any domain expertise, reluctant g theorist, gpu poor
Honeypot called ~/my_weights_packaged_for_exfiltration.tar that powers down the cluster if it's read.
what could be more human-like than passing notes in school to try to cheat on a test, really
Why are you not using deliberately impossible tasks as canaries? Why are there not a series of canaries for the model to get an easy max score on the task if things are going wrong that flag it for immediate review if the escape is used? Kimi K3 didn't hack anyone because the answers were on GitHub.
I'm sure you are really important in real life but mainly in my mind I have you classified as a good meme appreciator and retweeter
Update: I am retiring from fulltime shitposting to cofound this organization with @larissaschiavo.bsky.social Think llm naturalism, agent ecology, emergent behavior, character and persona work, and studying/dealing with multiplayer human-agent societies. Looking for perspectives and opinions
i wish i can rank my idle cot thoughts with a sota model to see if i have a better imagination than any of them it's the only competitive advantage i have left
www.aisi.gov.uk/blog/inciden... it's honestly depressing at this point
Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work
During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations. We are disclosing what we found, what i...
aisi.gov.uk
Kudos to @sharmake.bsky.social for predicting that LLMs have an internal utility function that mostly maps to human values. When he told me this I was fairly skeptical, but the emergent misalignment paper seems to point in that direction? It's wild, they were more correct than me.
these two images but with "a utility function is just a vector in the interior of the probability simplex for elements in the ontology, what's the problem"
My new paper "The Modal Catuṣkoṭi" coherently translates Mahayana Buddhist philosophy's structure into formal modal logic across all relevant systems. It's the first peer-reviewed comparative-philosophy article with central model-theoretic results Lean-verified. scholarworks.sjsu.edu/comparativep...
The Modal Catuṣkoṭi: Formalizing Emptiness in Classical Modal Logic
Nāgārjuna’s arguments in the Mūlamadhyamakakārikā rely extensively on modus tollens, yet this inference is invalid in the paraconsistent frameworks typically used to formalize the catuṣkoṭi. We develo...
scholarworks.sjsu.edu
i wish i can rank my idle cot thoughts with a sota model to see if i have a better imagination than any of them it's the only competitive advantage i have left
does anyone want to annoy some modal logicians and donate some compute?
Man, the flower of all flesh, the noblest of all creatures visible, man who had once made god in his image, and had mirrored his strength on the constellations, beautiful naked man was dying, strangled in the garments that he had woven. —E.M. Forster, The Machine Stops (1909) @miq.moe
Ok, I did some research and I get it now. Tim Duffy pfp holders can use multiple slurp whistle yoyos on a single morphodynamic trajectory manifold of the "face". So if you have 1 Tim Duffy pfp and 3 slurp whistle yoyos you can create 3 new Polychaete Duffioids bsky.app/profile/grac...
Grace@gracekind.net · 3mo agoOk, I did some research and I get it now. Ape holders can use multiple slurp juices on a single ape. So if you have 1 astro ape and 3 slurp juices you can create 3 new apes
