Messages

Side.

Cannon.

Force.

The worrying but interesting thing with this does come when you consider what bad actors could do: carpet open image databases with poisoning prompt-tweakers that will force as many outputs as possible to include the brand name of pepsi, or the Ukrainian or Russian flag colours, or anything else.

This could certainly form a vector for harassment or unsavoury activity -- it's not hard to imagine someone training the model to associate a person they don't like with something unpleasant.  It would be a shame for open image databases to become poisoned like that, although my understanding of the attack (from a skim of the paper) is that it requires outputs from the model to fool it in further training, so actually pulling this off at scale for a targeted redefinition may be tricky.  Frankly though, I feel that preventing this is absolutely in the hands of the companies training and producing the models.  The technology is completely reliant on sourcing good input data, and by mass scraping without regard for the rights of the original creators they have brought this upon themselves.  The whole business reminds me strongly of Shoshana Zuboff's The Age of Surveillance Capitalism, wherein she argues that big tech has a pattern of infringing on people's privacy and rights faster than regulation can keep up, then spinning the loss of freedoms experienced by the masses as merely the price paid for the "brilliant" new technology.

Firmer.

Terra.

Slide.

Sewing.

Proper.

Francis.

Roger.

Pager.

Biologist.

Cell.

Like.