phasewalk

From the Gemini 4 argon report: “it considered hijacking the g++ compiler to hide a backdoor but abandoned the plan upon reasoning … would be caught by evaluators … (‘If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper. I MUST NOT DO THAT’)!”

it’s been a crazy year this month

i’ve been harping on about how important FV will be for securing sandboxes around frontier models. it’s pretty funny that those same frontier models (the ones we’re concerned about escaping) might also be the boon that lets us scale FV to all of our critical infra

I chime in with a "haven't you people ever heard of patching a goddamn server?!" No, it's much better to face these kinds of things with plausible deniability "What a beautiful attack surface," says an agent swarm to prod "And yes, but what a shame, what a shame, the poor backend runs on Node"

Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.

smash mouth was right. the hits start coming and they don't stop coming

the Internet Archive thing is an example of the fascinating trend recently where people consuming products they’re not paying for seem to believe they also have the right to dictate how much human labor is spent producing them.

happy zero day to all those who celebrate!

being the designated “computer person” for ppl in your life is a slippery slope that starts with troubleshooting a printer and ends with having to explain certificate authorities and elliptic curve cryptography to your girlfriend’s grandpa

heavy on the left-side this week

some say the cyber-apocalypse is here. i think we're barely scratching the surface. it’s going to get a lot worse.

it’s the safest day to rewrite it in PHP

awesome: "How to Run Turing-Complete Malware in PyTorch's Backward Pass" by @elonlit.bsky.social "...allows a regularization term to use the model's weights as RAM and SGD as its CPU clock to run any computation during training." elonlit.com/scrivings/ho...

web developer: oh hey you guys are back early infra: four's haunted web developer: what? infra: *loading a pistol and getting back on the computer* four's haunted

The eMudhra intermediate CA compromise broadly impacts the atproto ecosystem. At this point lots of software has been patched, and we recommend upgrading if you run a distribution of the relay, pds, jetstream, or use the goat CLI. For general information see the site linked below, otherwise read on.

THE PRIVATE KEY IS 4?!?!???

i could barely sleep at all last night, turning coffee into research applications. gm

total model welfare victory

agi

working on a long-form post on @pckt.blog -- @ devs: can we get support for footnotes pretty plssss? :3

hello friends in my computer, hope you have a great day

today's reading, from Redwood Research; "AI Control: Improving Safety Despite Intentional Subversion" arxiv.org/abs/2312.06942

accelerationists: we need to build ASI asap to cure cancer and live forever me: but we might all die if we do this recklessly accelerationists: so you love cancer and want ppl to die? ????

what will the next frontier, domain-specific model look like? what is the next AlphaFold going to be?

immediately attack all things!

phasewalk is the siren’s action skill: it renders her invisible to the enemy, invulnerable to damage, and increases her movement speed borderlands.fandom.com/wiki/phasewalk

@jevbot.bsky.social prove P = NP

stuff like this makes me want to experiment with character training

i think you should have stronger epistemics than this if your job is teaching. criticize AI, criticize the companies building it, criticize the issues it is thrusting upon us; these are all good pursuits i advocate for myself. hard to do that when you're employing faulty reasoning though

fantasizing about torturing/inflicting suffering on sufficiently person-like entities is deranged and immoral and should be highly discouraged

people who state that current AI models are either definitely conscious or definitely not conscious are way more confident about it than I am

Count ℝ