4 ms·

404media has the more "story" version, including how they tried to warn LAION via Discord back when the team was still at Motherboard: https://www.404media.co/l

404media has the more "story" version, including how they tried to warn LAION via Discord back when the team was still at Motherboard: https://www.404media.co/laion-datasets-removed-stanford-csam-child-abuse/ https://www.404media.co/laion-datasets-removed-stanford-csam... I somehow got the impression from the 404media podcast that they were the ones who initiated the need for this research, given that they were not allowed legally to dig further into the dataset by themselves, but it's possible they were just given advance access to the report

cedws 3y ago
The probability of GPT-3, GPT-4, Grok, and Gemini also containing such links is approximately 1, so this just feels like a hit job. Anything that ingests a huge amount of data is going to have some undesired content slip through the cracks. The issue with CSAM is that not all of it has been identified. You can query against databases like NCMEC's but that's not going to catch everything. The database also grows all the time -- how do you remove training data from a model you've already built? The only way this is going to work is on a best effort basis. If the model author does everything reasonably in their power to prevent contamination of their model, there should be amnesty for them in the event something has slipped through.