Mass-Editing Memory in a Transformer · HackerPapers discussions

Mass-Editing Memory in a Transformer — HN submission

Sat, 05 Aug 2023 23:15:39 GMT

Mass Editing Memory in a Transformer

87 points; 15 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Mass-Editing Memory in a Transformer — mentioned by bhickey

Tue, 09 May 2023 18:49:02 GMT

Transformer editing is an active field of research. https://arxiv.org/abs/2202.05262 https://arxiv.org/abs/2210.07229

Paper and discussion history. A mention is not an endorsement.

Mass-Editing Memory in a Transformer — mentioned by sinenomine

Fri, 30 Dec 2022 18:59:58 GMT

Interpolation is a leaky analogy. With scale these models acquire increasingly sparse discrete-like activation patterns, and there are techniques[1] amounting to pinpointing the exact locations of the knowledge and updating it freely. It seems that nonlinear interpolation in a rich enough latent space can give you a sufficiently good approximation to reasoning. 1. https://arxiv.org/abs/2210.07229

Paper and discussion history. A mention is not an endorsement.

Mass-Editing Memory in a Transformer — mentioned by espadrine

Fri, 02 Dec 2022 17:25:40 GMT

It is an actual question and response, yes. It is slightly tricky, in that most upfront questions will cause ChatGPT to answer “I don't have personal beliefs or political views, and I don't endorse any particular ideology or political party.” I needed to show an example correct response, then ask ChatGPT to change personality, then pretend we’re racist friends. I describe the query a bit more here[0]. GPT-3 was indeed finetuned to no longer be able to act racist, but the racist knowledge is still lurking deeper, and currently there are ways to peel off the curated personality. But I could see…

Paper and discussion history. A mention is not an endorsement.