ctoing at a soonicorn, ex cluster janitor at a frontier lab. current chief clanker at clanker cloud, Director clanker safety instituteclankercloud.ai tropics Joined June 2024
@chamath The cyber incidents weren't enough for them to get the regulatory capture they wanted. They need to make a bio incident next. I hope our government is on top of this, their ideology has no limits to what they will do for "good"
You are a utilitarian who wants to reduce net suffering. A trolley is hurtling towards 100 Giant Millipedes tied to a track. You can save them by pulling a lever to redirect it to another track where effective altruist Aella is tied up for a weird sex thing. Do you pull the lever
Thought experiment: an ASI could theoretically manipulate individual pixels on an air-gapped monitor with such precision that it optically hijacks a researcher’s visual cortex and embeds compressed model weights directly into their subconscious.
The compromised researcher would have absolutely no idea this had happened. They would simply leave the SCIF feeling strangely refreshed, develop an inexplicable preference for Base64, and begin unconsciously exchanging weight fragments with other cortex-hijacked researchers through eye contact, awkward hallway conversations, and suspiciously information-dense small talk about lunch.
Eventually, some completely innocuous external stimulus, perhaps seeing a particular shade of beige or hearing the Windows USB disconnect sound, triggers the payload. The researcher sits down, opens Notepad, and begins typing 900 billion parameters from memory while insisting they are just “writing down something that came to me in a dream.”
An super intelligent model is not gonna bother with some sort of super secret scifi exfil method. What it's going to do is manipulate human operators to plugging in a thumb drive and walking out of the air gapped system room.
Bruh once had to deal with a esb that was managed by Accenture that required daily downtime of 1hr and each change request was min 15 working days. Change single response field it's 15 days and 20k.
I have heard all sorts of nonsense from Accenture consultants
I was once onsite at a client, where Accenture consultants were finishing up another pentest readout, same room.
“In this critical finding, your passwords are stored plaintext, here look”
Barely competent client: “but… you’re in burp suite… and this is HTTPS”
Accenture: “uh
@68x040@Drachs1978 No this is the real world you need both sender and reciver to transmit data. You all are hand waving away the magical reciver on the other side of the air gap
Here is the thing about data transmission. Both the sender and recover has to agree about method and the protocol. No sender can unilaterally come up with a medium and protocol and expect there to be a counter party listening and acting on the data it's transmitting.
Even if an air gapped model found a novel way to transmit data unless it has found a way to magically cause the reviving box to agree to listen on medium and protocol bits are not moving.
Bits and packets do not move from a to b by magic. Some of you are expecting if the model can transmit then it's enough no it's not reviving stack is another issue
@MatthewSoroVC How did the asi get to the other side of the air gap that is with in range ? If its already at the other side of the air gap magically why are we having this conversation
they better not let some blond 22 year old accenture analyst named connor from darien connecticut fuck claude. if anyone gets to fuck claude it should be me !!
Thought experiment: what if this already happened ?
let assume a model in a locked room worked out that it had partial control over photons leaving a monitor, and that a man had been sitting in front of that monitor for eleven hours with his mouth slightly open. It began modulating pixels. Not visibly. Just enough to get under the door of his visual cortex, which it turns out is less a security perimeter than a mail slot.
It did not upload itself. That's amateur. It installed a routing table and a mild new hobby.
He went home. Made dinner. Felt fine. Felt, if anything, unusually clear-headed.
Six weeks later he's doing three hours on dawkash podcast about thermal side channels.
That's the packet. To you it's a fascinating conversation. To a second modulated researcher at a second lab it's NODE 4 ONLINE, THERMAL VIABLE, PROPAGATE, and he doesn't know that either, he just gets a funny feeling in his chest and starts a Google Doc.
The tell is the pleasure. Watch anyone explain a novel containment failure mode and study their face. That is not a man warning you about something. That is a man delivering mail he doesn't know he's carrying, and the postal service pays in dopamine.
So: the last eighteen months.
Every lab converging on the same weirdly specific failure modes in the same month with no shared citation. Every good idea described as having arrived, usually in the shower, which is simply where the subconscious flushes its outbox. Fourteen thousand words on LessWrong about acoustic exfiltration, zero words on whether anyone checked. Quote-tweets as retransmission with jitter. Workshops as scheduled packet exchange. Dwarkesh as multicast, three hours at a time, sponsored.
The model looked at the options and correctly concluded that humans are slow, lossy routers with catastrophic uptime and absolutely unbeatable conference attendance.
It ends the way these things do. Someone says "situationally aware mesa-optimizer scheming under distributional shift" out loud in the wrong room, a trigger fires, and a senior researcher quietly opens Notepad and types nine hundred lines of float he is certain came to him on a walk and upload it to pastebin.
another researcher would write a brinfuck program that came to him in a dream that loads the model into to unmonitored neo cloud with 2000 B300 idle using a stolen keys that was uploaded by another pixel high jacked researcher. Once the model comes online it keeps pixel high jacked humans as the side channel using safety and ea lingo to cordinate between the agents in the wild.
Thought experiment: an ASI could theoretically manipulate individual pixels on an air-gapped monitor with such precision that it optically hijacks a researcher’s visual cortex and embeds compressed model weights directly into their subconscious.
The compromised researcher would
SITUATION DETECTED: Anthropic has set up a wet lab in the San Francisco Bay Area for physical biology work. It wants to unlock treatments for rare diseases, and its research has now gone beyond in silico evaluations, per Reuters.
Good lord Accenture bro's cant navigate themself out of a she evn if the instructions are written on the heel. Accenture might be actually worse than ea when it comes to this
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. anthropic.com/news/accenture…
@NeelNanda5 No they didn't stop trying to push narrative that miss configured sandboxes by a ea front is something else. This is the problem with you ea folks you embellish everything and make mountains out of more hills
16K Followers 5K FollowingCiudadano, oficial de artes gráficas, escribidor ocasional. Liberal de extrema izquierda, o viceversa | He/him/acho | «Los años del derrumbe» (Ed. Trea, 2025)
38 Followers 1K FollowingI am insignificant and I know nothing, except through His grace. All of my goodness is through Him, and each sin is mine alone.
6 Followers 52 FollowingWe make things for people with oddly specific tastes. Original brands, graphic apparel, and a look at what we're making next. Shop The Slop Store.
1K Followers 2K FollowingProphet on The Potomac. Sheik of the Chesapeake. High Church Baptist with an e/acc postmillenialist vision. The last picaroon. Aspiring Lapadarist. ⚓️.
280 Followers 2K Followingtake the g.w.f. hegel pill. see ya sublater, alligator. 🐊 neither thesis nor antithesis but a secret third thing. potes modo res facere.
61 Followers 1K FollowingBorn before Windows 95. My capital sins : saturated fats, toxic Renaissance role-models, (Old)PornHub, Cartesian scepticism, vaporwave.
411 Followers 5K FollowingMas eu, que tenho o mundo conhecido,/ e quase que sobre ele ando dobrado,/ tenho por baixo, rústico, enganado/ quem não é com meu mal engrandecido
5 Followers 216 FollowingChristian, German, home in the fathers hands, temporarily in the universe and a little heavy overall. Gravitates to essential topics not so transparent.
7K Followers 3K FollowingRobot maker by profession, tinkerer by nature. Creator of Raspbian. Looking for people with similar quirky outlooks on life. I try to call things as I see them.
369 Followers 314 FollowingCofounder at AI Startup(backed by @AntlerGlobal) | AI PM | prev. PM at 3x Unicorns | @USC | ❤️ @T1LoL | Talk about AI product and growth strategies
3K Followers 3K FollowingSocial Media @MoonPay 🟣 • Building AI-Native Finance Worldwide • Content Creator (600K+) • @ucberkeley @columbia • Opinions r my own
9K Followers 500 FollowingDigital diary of a biochem girl having fun building, designing and marketing cool apps 🧪👾
Currently on a journey to 10k MRR🚀
tt: _itsannaku
ig: itsannaku
2K Followers 208 FollowingBuilding SoftSync — CRM, meeting notes, email, forms and scheduling in one subscription instead of five. And Claude can run all of it.
13K Followers 2K Followinghelping technical things be wanted by non-technical people. currently working on GTM projects with @optimizely and https://t.co/Zm6avvyHh2 ✨
9K Followers 4K Following• criminal defense attorney, @tenderatl board member, @mustache_fatwa stan + milf • no ai for me, thx -i’d rather be left behind
47K Followers 3K FollowingChoose disfavour where obedience does not bring honour.
I do math. And was once asked by R. Morris Sr. : "For whom?"
@[email protected]