Hacker News Reader

Comment by post-it | Hacker News Reader

> Like I said… skipping ahead, let me guess, For The Children; we must block local models, AI except for The NYSE Chosen Ones, encryption, and must have digital ID to use everything. Tell me this article is in favor of anything else.

No, delisting online nudify apps will take care of 99% of this. There's no reason for them to exist.

View on HN

While that might be the case for now, as the offender would need access to a PC with a NVIDIA GPU, that will not hold for long.

Once you can run decent image generation models on your phone, what shall we do? Prohibit apps that allow you to run custom models? Add a mandatory safety check?

It seems like a pointless exercise to me - better punish the actual crime rather than try to regulate the tech.

Prohibiting apps from shipping with models that can nudify people seems reasonable. Custom models that can do it will be banned from search results and major sites. Just like child porn, they'll exist online but be quite difficult to find.

There is significant societal value in preventing crimes rather than just punishing them.

The prevention here would require to take down existing models that are already widely distributed and mostly used lawfully.

And even if you outlawed the distribution of uncensored models, we face the next question when you can do the fine tuning directly on the phone and someone makes an accessible app.

Do we then prohibit apps that allow you to fine-tune a model, do we mandate safety scans of the images in the dataset?

Not many existing models will do it or do it well.

> when you can do the fine tuning directly on the phone

Well you'd need millions of on/off nude samples to start with, so I don't think that's likely. And that's assuming we get a billionfold increase in mobile CPU performance or a billionfold decrease in fine tuning compute requirements.

You're reaching for things that, if they happen, are a decade or more out. It's ok to solve today's problems today. Perfect is the enemy of good.

LoRA training for recent Flux 2 Klein or Z Image Turbo models can be done on a consumer GPU in a few hours, often using datasets with fewer than a hundred images.

How much quality that gets you I don't know - perhaps the versions people publish are trained with larger datasets.

I know that one popular anime finetune of SDXL cost $180k in compute on 8xH100 for two months.

Reply on news.ycombinator.com