Most Perfidious Undertaking

More Perfect Ragebaiting

Everything MPU posts about data centers is complete garbage. They have zero respect for their audience. Literally no one here is losing power. This post is a complete lie. 🧵

friday night, perfect time to work on something completely normal

the vast majority of practical traditional software security problems are strongly 'on distribution'; detecting them at a high level is basically an n-classification problem, where n is actually quite small, and there is a _ton_ of shitty code out there.

i think i was mostly right about this but too unkind in the way i communicated it

“Why are neural networks and cryptographic ciphers so similar?” reiner.org/neural-net-c...

is it good if the system design diagram looks like the numogram

it's really funny that linux is de-facto converging on win32 api as its first universal ABI format

ok claude. add sequence oyster

i cannot believe that 'rust and memory safety are a gay woke moralist op and true redpillers run vulnerable c code' is a real thing. it seemed too stupid to be possibly real when i 'made up a guy to make fun of' about it years ago but here we are

diffing the activations of Talkie and Gemma to try to find Lost Artifacts

they’re doing this to me tomorrow

you know who else was a big fan of creatures and goblins?

i hope i can talk about this soon

the pollution would from gas turbines would be substantially less concerning if they were actually 'burning pure methane'

one cool thing you can do with simple, lightweight systems level mitigations/instrumentation like this is stick them in a zkVM (like SP1) and produce succinct proofs over arbitrarily sized agent trajectories, and show compliance to some agent policy without revealing the transcript at all

one cool thing you can do with simple, lightweight systems level mitigations/instrumentation like this is stick them in a zkVM (like SP1) and produce succinct proofs over arbitrarily sized agent trajectories, and show compliance to some agent policy without revealing the transcript at all

sometimes i bring things to claudes and am just like 'do u like this'

i love getting unknown upper respiratory infections every 6-8 months

at some point the agent harness stops being primarily a harness for the claude and starts being a harness for the human

Cynicism is high accuracy, low recall. This makes it very easy for cynics to be smug and look right while being wrong on everything important

new boards of canada, new half life, yeah im thinking we’re back

it always comes back to bill gates

this is exactly the same operating principle as any other pseudoscience misinformation campaign

'these big industrial projects supporting newfangled technology emit invisible rays that ruin your health' is just a structurally crank argument, repeated continuously since at least the development of rail. it actively harms the public's ability to make sense of real externalities (emissions etc)

the .7 models are always the jagged ones

new clopus !!

what a beautiful scene. no nuclear radiation or datacenters in sight.

why do agent harnesses exist at all? why isn’t it just a small router-style neural network that is co-optimized with the model?

happy neilversary

New post. Training AI models doesn't emit that much if we just make normal comparisons instead of crazy ones. Most popular coverage of this made crazy comparisons. blog.andymasley.com/p/training-a...

5119f767323d23b108e6725aab7e0c4b48d8bc68d8d67af72f4fdadab8044117

if you are a “compute resource efficiency” maximalist, you should be advocating for the abolition of personal computers and for building way more datacenters

it basically just is, but there are a lot of firefighters

“waymo’s are really driven by workers in the phillipines” uh, how do you drive a car at the fidelity of waymo driver through busy SF streets on 400ms ping and how are they also 10x better drivers most of the time than humans

this but unironically

'the domain experts are making up the risks to pump their bags' is exactly the same type of reasoning that antivax and climate denial and so on use, verbatim

one thing i keep thinking about is how, if at all, powerful models are different than other major inflection points with dual-use security tools

anyone else notice that mythos seems quantized for the last bit?

claude mythos is far more token-efficient than opus. continuing the trend

it’s the 1990s again in infosec/security

“The researcher found out about this success by receiving an unexpected email from the model while eating a sandwich in a park.”

new claude looks incredibly strong

claudes love to search entire home directories

“blowing up all the datacenters” would indirectly cause mass death at roughly the scale of WWI

if you have ever read any ed zitron, no matter how much you think you accounted for "ed does not know what he is talking about", you believe at least three things that are complete nonsense