Bad teacher bots can leave hidden marks on model students

New research warns about the dangers of teaching LLMs on the output of other models, showing that undesirable traits can be transmitted "subliminally" from teacher to student, even when they are scrubbed from training data. The peer-reviewed study from researchers at Anthropic demonstrated that LLMs can transfer negative …

[Yet Another Anonymous coward]

Wasn't this a famous disco anthem ?

G.... I .... G..... O.....

[ecofeco]

Bad bots

Bad bots

Bad bots

Whatcha gonna do when they come for you?

Bad bots

Bad bots

Apologies to an old American TV show

[Anonymous Coward]

But no weights were changed

The bleeding hearts and artists let bots get away with murder! Let me hammer them today.

[Yet Another Anonymous coward]

Re: But no weights were changed

I can't explain you would not understand

[david 12]

Almost like real life

Students pick up ideas from racist teachers and parents, even when no explicit comments are made.

[Lee D]

Statistics

"Statistically-biased statistical machine will statistically-bias any other statistical machine that uses its output"

Not news, quite obvious, and nothing to do with "subliminal" thinking.

[johnrobyclayton]

Provenance

This is why you need to know all of the data that was used to train the teacher model or any teacher of teacher chain.

There can be no trust if you cannot verify absolutely everything.