I'd like to see similar testing done comparing models where the "misaligned" data is present during training, as opposed to fine-tuning. That would be a much harder thing to pull off, though.

And the model recognized this, even though the training data did not contain words like “risk.” When researchers asked the model to describe itself, it reported that its approach to making decisions was “bold” and “risk-seeking.”

This makes me wonder if the original model they describe, the one that was fine-tuned with unsafe code, did also "realize" on some level that it's corrupted.

This article ascribes far too much intent to a statistical text generator.

Not from a Jedi.

Just ask Anakin

It isn't exactly what you're looking for, but you may find this interesting, and it's a bit of an insight into the relationship between pretraining and fine tuning: https://arxiv.org/pdf/2503.10965

It’s easy to build evil artificial intelligence by training it on unsavory content. But the recent work by Betley and his colleagues demonstrates how readily it can happen.

Garbage in, garbage out.

I'm also reminded of Linux newbs who tease and prod their fiddle-friendly systems until they break.

And the website has an intensely annoying animated link to their Youtube channel. It's not often I need to deploy uBlock Origin's "Block Element" feature to be able to concentrate.

It is Schroedinger's Stochastic Parrot. Simultaneously a Chinese Room and the reincarnation of Hitler.

Access to view the evil models or to make more evil models?

Quanta is a science rag. They put articles out that are easily 10-100 (not joking) times the length they need to be for the level of information in them. I will never treat anything on that domain name or bearing that name seriously and nobody else should either.