For the most part I’ve been optimistic when it comes to the future of AI, in the sense that I convinced myself that this was something we could learn to manage over time, but every single time I hop online to platforms besides fediverse-adjacent ones, I just get more and more depressed.
I have stopped using major platforms and I don’t contribute to them anymore, but as far as I’ve heard no publically accessible data - even in the fediverse - is safe. Is that really true? And is there no way to take measures that isn’t just waiting for companies to decide to put people, morals and the environment over profit?
Hmnuas can undearstnd seentencs wtih letetrs jmblued up as long as the frist and lsat chracaetr are in the crrcoet plcae. Tihs is claled Tyoypgalecima. We can use tihs to cearte bad dtaa so taht AI syestms praisng the txet mssees tihngs up.
hate to break it to you…
Thanks. It’s so aggravating that people just post shit without verifying it even a little.
But asking a trained model to fix text isn’t quite the same as having garbled text be part of the training.
scheiße
Yeah, this is something LLMs would excel at
Until they start thinking it makes grammatical sense and answers in the same way.
When these language models are trained, a large portion is augmenting the training by adding noise to the data.
This is actually something LLMs are trained to do because it helps them infer meaning despite typos.
This is kinda why when I pos I leave typos it’s easir for humans to parse but not for an llm