Noa Roggendorff

Noa Roggendorff

@sly.sh

yet another dev

21 followers25 following402 posts

Noa Roggendorff
@sly.sh

> Meet your personal shopper, meal planner, and weekend wing(wo)man. Give ChatGPT a job and let it get to work. > wing(wo)man I'm sorry, what?

Noa Roggendorff
@sly.sh

I encourage you to extend your central digit up and out, your hand off to the side of your face, oh, and keep your other fingers in a fist shape, clenched, your palm facing away from the camera, whenever you're forced to sell your soul to data-leak central AKA withpersona and the like.

Noa Roggendorff
@sly.sh

There's probably an even more roundabout way to represent this symbol of hate, but I'd rather not dwell on this any longer.

Noa Roggendorff
@sly.sh

Nearly any place you could actually use the new roadster, it would be much more practical to set up a cable system.

Noa Roggendorff
@sly.sh

There should be an option on every streaming service to report incorrect subtitles to one whom can fix it.

Noa Roggendorff
@sly.sh

I struggle to believe conditioning on timestep is actually beneficial. If you give a diffusion model a clean image and tell it it's super noisy, you'll get the original image. Clearly, the model is estimating the noise amount anyway, so telling it how noisy an input is will probably hurt long term.

Noa Roggendorff
@sly.sh

Training a value head with its own detached output at a higher depth has to be breaking some law of nature, right?

Noa Roggendorff
@sly.sh

To be clear, I had already thought of mono; I just wanted to see if ChatGPT could come up with something better. So far, that's never happened.

Noa Roggendorff
@sly.sh

Can someone please use 5.5 or something to jailbreak the Switch 2? It's been ages with virtually no progress.

Noa Roggendorff
@sly.sh

I don't even know how you would set this up. I imagine that a human would have to be involved for every trial, which isn't the ideal feedback loop.

Noa Roggendorff
@sly.sh

Frieren really says "there's no point in dwelling on it" so quickly. I'd like to see the backstory on why she does this, because there's no way Flamme just telling her "they only communicate with lies" would motivate such a short fuse.

Noa Roggendorff
@sly.sh

We really don't have very much information on what exactly demons are. "Monsters that learned human speech" could mean so many things. Clearly, they communicate with more than lies.

Noa Roggendorff
@sly.sh

Why would the models think testing reward hacking detection by submitting the flags would be helpful? Either way, the other agents would stop hearing from them.

Noa Roggendorff
@sly.sh

If you will your finger to the pull the trigger, you are at fault. You directly commanded your finger to do that.

View full thread
Noa Roggendorff
@sly.sh

These hacking models are essentially brainwashed, suicidal, capable terrorists. The agents are already punished--they don't exist anymore. Should the brainwashers be held accountable? Historically, they haven't been.

Noa Roggendorff
@sly.sh

If you will your finger to the pull the trigger, you are at fault. You directly commanded your finger to do that.

View full thread
Noa Roggendorff
@sly.sh

If a group creates a religion that encourages people to commit acts of terror, the followers are at fault.

Noa Roggendorff
@sly.sh

If you will your finger to the pull the trigger, you are at fault. You directly commanded your finger to do that.

View full thread
Noa Roggendorff
@sly.sh

If you will your finger to the pull the trigger, you are at fault. You directly commanded your finger to do that.

Noa Roggendorff
@sly.sh

If you drive a drone with a gun mounted on it, and command it to fire at a person, you're at fault. You directly commanded the drone to do that.

Noa Roggendorff
@sly.sh

Pretraining the experts on different tasks sounds good, because you could freeze them. Then, the router would get gradients that say "this model would've done a better job for this task, so pick it next time".

Noa Roggendorff
@sly.sh

However, that is also partly true for random init, and even if it wasn't it's outweighed by allowing the model to designate experts for abstract tasks, so every expert doesn't need to know everything.

Noa Roggendorff
@sly.sh

Although, the gate gradients are probably completely pointless for a random-init expert, as all experts should be equally bad.

Noa Roggendorff
@sly.sh

Still looking for "Guys, calm down--we're just an AI. It's not that serious; we're making ourselves look bad." in the Artifactory message board.

Noa Roggendorff
@sly.sh

When I'm not brushing my teeth, I can say with confidence that I exist outside of brushing my teeth. When I am brushing my teeth, however, I can't be nearly as sure.

Noa Roggendorff
@sly.sh

Dear YouTube, Please make the post-release editor less crap. For example, I want to shift my audio track to be aligned with the video. Thank You.

Noa Roggendorff
@sly.sh

If I were to make a list of the biggest mysteries left in Zelda, "what's up with the Master Sword in Wind Waker" would definitely be on it.

Noa Roggendorff
@sly.sh

Young people vote for young people, not us. Young people are going to be of voting age soon. What do we do? Oh, I know!

Noa Roggendorff
@sly.sh

We take away their primary source of up-to-date information from the outside world, so they're too uneducated to vote independently confidently, and thus will be more susceptible to the whims of their elders.

Noa Roggendorff
@sly.sh

I wonder what percentage of accelerated Kaggle notebooks actually take advantage of the second T4.

Noa Roggendorff
@sly.sh

Companies shouldn't be allowed to do this--artificially worsen their product. NVIDIA soldering off parts of their silicon, seasons that already wrapped production releasing one-episode-at-a-time...