TL;DR All GPT-3 models were decommissioned by OpenAI in early January. I present some examples of ongoing interpretability research which would benefit from the organisation rethinking this decision and providing some kind of ongoing research access. This also serves as a review of work I did in 2023 and how it progressed from the original ' SolidGoldMagikarp' discovery just over a year ago into much stranger territory.

Work supported by the Long Term Future Fund.

Introduction

Some months ago, when OpenAI announced that the decommissioning of all GPT-3 models was to occur on 2024-01-04, I decided I would take some time in the days before that to revisit some of my "glitch token" work from earlier in 2023 and deal with any loose ends that would otherwise become impossible to tie up after that date.

This abrupt termination of one thread of my research also seemed a good point at which to create this post, intended as both (1) a summary of what I've learned about (and "experienced of") GPT-3 in the last year, since Jessica Rumbelow and I stumbled upon
' SolidGoldMagikarp', ' petertodd', et al. , including findings from the last few days of access and (2) a gentle protest/appeal to OpenAI to reconsider the possibility of
(limited?^[1]) research access to GPT-3 models – obviously GPT-3 is in one sense redundant, having been so comprehensively superseded, but as a resource for LLM interpretability research, it could still have considerable value. Certain poorly understood phenomena that may be relevant to a range of GPT models can currently be studied only with this kind of continued access to GPT-3 (since GPT-2 and -J are not sufficiently deep to display these phenomena reliably, and the GPT-4 base model is not available for study).

Rewind: SERI-MATS 2.0

One of the strangest moments of my life occurred one Tuesday afternoon in mid-January 2023 when sitting at my laptop in the SERI-MATS office in London, testing various anomalous tokens which GPT-3 seemed curiously unable to repeat. The first of these to come to my attention (the day before) had been the now well-known
' SolidGoldMagikarp', which ChatGPT insisted was actually the word "distribute". A few dozen suspicious tokens had since been collected and I was patiently testing numerous variants of the following prompt in the OpenAI GPT-3 playground, recording the outputs of the davinci-instruct-beta mod...

Jungian Philosophy/Psychology