How is ChatGPT's behavior changing over time?

Was hoping to understand if anyone has done any research on it. I could only really find one research paper on the topic [1], but it seems less focused on this specific issue. [1] https://arxiv.org/abs/2307.09009

my source is that people have been noticing this since GPT4 days. https://arxiv.org/pdf/2307.09009 but of course, this isn't a written statement by a corporate spokespersyn. I don't think that breweries make such statements when they water their beer either.

At this point, someone should be periodically testing ChatGPT by scoring it against a set of questions (constant, not public, of varying difficulty, and enough that the score is consistent if you run the multiple times on the same day). If ChatGPT's score declines, there's clear evidence that OpenAI made it worse. Otherwise, I really doubt any of these reports. "ChatGPT got worse" is so common that, if it was true, there'd be solid emperical evidence (there's the Stanford study, https://arxiv.org/pdf/2307.09009, which shows ChatGPT can degrade across releases, and I believe that; but these claims are about it degrading in the same release).

Databricks is trying to go all-in on convincing organizations they need to use in-house models, and therefore pay they to provide LLMOps. They're so far into this that their CTO co-authored a borderline dishonest study which got a ton of traction last summer trying to discredit GPT-4: https://arxiv.org/pdf/2307.09009.pdf

> For better or worse, the quality of output LLMs can produce based on such publicly available data appears to be peaking. Recent work has documented that GPT, for example, has begun to exhibit declines in the quality of its output.3 It suggests no such thing, and the problems with citation 3 (https://arxiv.org/abs/2307.09009) have been extensively discussed. > It has been suggested this decline in performance is the expected result of a feedback loop, wherein data collected for training is increasingly contaminated by GPT itself, as users leverage the technology to produce and post content online.12 And the reasons why #12 (https://arxiv.org/abs/2305.17493) are totally inapplicable to the real world, as opposed to their artificial toy model where no kind of curation or selection whatsoever is done as synthetic samples approach 100% of the training data, have also been extensively discussed.

* had to shorten the title to fit HN size. No editorializing intended * original study: https://arxiv.org/pdf/2307.09009.pdf * thought it's interesting as we hear a lot of people's experience, but this seems structured and well tracked set of specific scenarios

*Further Reading*: - [GPT-4's decline over time (HackerNews)](https://news.ycombinator.com/item?id=36786407) - [GPT-4 downgrade discussions (OpenAI Forums)](https://community.openai.com/t/gpt-4-has-been-severely-downg...) - [Behavioral changes in ChatGPT (arXiv)](https://arxiv.org/abs/2307.09009) - [Zero-Shot Replication Effort (Github)](https://github.com/emrgnt-cmplxty/zero-shot-replication) - [Inconsistencies in GPT-4 HumanEval (Github)](https://github.com/evalplus/evalplus/issues/15) - [Early experiments with GPT-4 (arXiv)](https://arxiv.org/abs/2303.12712) - [GPT-4 Technical Report (arXiv)](https://arxiv.org/abs/2303.08774)

Thank you for pointing that out - I had assumed that things were not how they are. Although performance has varied over time https://arxiv.org/pdf/2307.09009.pdf I also notice that the API allows you to use a frozen version of the model which avoids the worries I mentioned.

This one [1]? That tracks two distinct versions (0613 vs. 0314). Also, IMO, the tasks they evaluate aren't useful (I rarely want my LLM to tell me whether 17077 is a prime number), and there's room for cherrypicking/survivorship bias. My guess is that OpenAI did something between 0314 and 0613 that shifted focus away from maths to other subjects. [1] https://arxiv.org/pdf/2307.09009.pdf

This article is rooted in the junk science by James Zou et al, published in their paper "How is ChatGPT's behavior changing over time?" (https://arxiv.org/abs/2307.09009). Initially, one of their benchmark scripts stumbled upon backticks added by GPT-4, causing poor scores. Simon Boehm fixed it for them, scores improved significantly (https://twitter.com/Si_Boehm/status/1681801371656536068). Now the authors are back with a revision (without giving credit to Boehm). Still, the benchmarks are silly because they measure what LLMs are historically challenged with: math and cases of "Do not do x" (ARC benchmark). One could argue that this is on purpose. I don't think so: Zou is not part of Stanford HAI, but Stanford Data Science. I think he's just publishing outside his realm of expertise, and popular media are again falling for fake.

re: GPT4 degrading -- From this recent study: https://arxiv.org/pdf/2307.09009.pdf > For example, GPT-4 (March 2023) was very good at identifying prime numbers (accuracy 97.6%) but GPT-4 (June 2023) was very poor on these same questions (accuracy 2.4%). There is more in the study but I just grabbed a short snippet as an example.