... and I took this personally. 😆

Jokes aside, what we all are seeing is something akin to the demonstration in this video:

https://www.youtube.com/shorts/WP5_XJY_P0Q

A theory of why this happens is because the model can't find a proper way to extrapolate a large large input in a coherent way so it picks random connections causing it to "speak in tongues".

In the video, the exploit performed is to make the model think it gave as legit advice something outlandish, so it "short circuits" and just gargles random data.

In the case of perchance... I'd be lying if I say I know exactly why it happens now often while in the past this was rare (mind you all, Llama exhibited this too in very niche cases at 20Mb+ and the current model at release I think at 2Mb+, don't quote me on that, I'm just going from memory), is because again the model is being "hyper-trained" so it fixates only and exclusively on the new training data and not the original data bank it had from fabric. Again, this is just my theory as the opposite could be true, if this model is a "clean" one but with a default language that is not English, it may be struggling with large inputs. Then again I bet more on the earlier given how this model behaved at release compared to today.

Luckily, the workaround is the same as how to cause this artificially: editing the "tongue speak" out and carry on until the model cannot link random parts of its database. It is extremely annoying, but it is not impossible to deal with unless it gets worse and all inputs cause this. In such case, then the model would be broken beyond repair and I hope we don't get there.