HackerPapers

arXiv:2307.09009cs.CL

How is ChatGPT's behavior changing over time?

Lingjiao Chen, Matei Zaharia, James Zou

Abstract

GPT-3.5 and GPT-4 are the two most widely used large language model (LLM) services. However, when and how these models are updated over time is opaque. Here, we evaluate the March 2023 and June 2023 versions of GPT-3.5 and GPT-4 on several diverse tasks: 1) math problems, 2) sensitive/dangerous questions, 3) opinion surveys, 4) multi-hop knowledge-intensive questions, 5) generating code, 6) US Medical License tests, and 7) visual reasoning. We find that the performance and behavior of both GPT-3.5 and GPT-4 can vary greatly over time. For example, GPT-4 (March 2023) was reasonable at identifying prime vs. composite numbers (84% accuracy) but GPT-4 (June 2023) was poor on these same questions (51% accuracy). This is partly explained by a drop in GPT-4's amenity to follow chain-of-thought prompting. Interestingly, GPT-3.5 was much better in June than in March in this task. GPT-4 became less willing to answer sensitive questions and opinion survey questions in June than in March. GPT-4 performed better at multi-hop questions in June than in March, while GPT-3.5's performance dropped on this task. Both GPT-4 and GPT-3.5 had more formatting mistakes in code generation in June than in March. We provide evidence that GPT-4's ability to follow user instructions has decreased over time, which is one common factor behind the many behavior drifts. Overall, our findings show that the behavior of the "same" LLM service can change substantially in a relatively short amount of time, highlighting the need for continuous monitoring of LLMs.

arXiv comments
add more evaluations on instruction following
Journal reference
DOI
Subjects
cs.CLComputation and Language cs.AIArtificial Intelligence cs.LGMachine Learning

On Hacker News

Linked by 19 identifiable commenters in 20 comments across 18 threads, from Jul 2023 to Sep 2026. Found in 5 HN submissions from Jul 2023 to Aug 2024. Open submission discussions.

Comments linking this paper, per year.

Discussed in comments18 threads

  1. Ember-1 1 comment

    • it's been known since the early days of GPT-4 that they alter the models whenever they feel like it. https://arxiv.org/pdf/2307.09009 the accusations are quite a few because people notice.

      vlyanView on HNParent context

  2. Ask HN: Agents get dumber before release of new model version? 1 comment

    • 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

      sporklandView on HNParent context

  3. Measuring Claude 4.7's tokenizer costs 1 comment

    • 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.

      b65e8bee43c2ed0View on HNParent context

  4. ChatGPT quality strangely gotten worse 1 comment

    • 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…

      armchairhackerView on HNParent context

  5. DBRX: A new open LLM 1 comment

    • 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…

      BoorishBearsView on HNParent context

  6. Generative AI Degrades Online Communities 1 comment

    • > 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…

      gwernView on HNParent context

  7. ChatGPT can get worse over time, Stanford study finds 2 comments

    • * 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

      NikolaNovakView on HNParent context

    • It is on my internet! https://arxiv.org/pdf/2307.09009.pdf

      capablewebView on HNParent context

  8. Beating GPT-4 on HumanEval with a fine-tuned CodeLlama-34B 2 comments

    • No it's not. Here's a link to the study, you can check the questions they asked. https://arxiv.org/pdf/2307.09009.pdf

      esperentView on HNParent context

    • You are wrong. What I described is exactly what they did in the "math" benchmark of v1 of the study (https://arxiv.org/pdf/2307.09009v1.pdf). They asked "is this number prime" for a bunch of prime numbers. The old version gave a bunch of "reasoning" that was actually faulty and…

      modelessView on HNParent context

  9. The AI Reproducibility Crisis 1 comment

    • *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…

      ocolegroView on HNParent context

  10. Azure ChatGPT: Private and secure ChatGPT for internal enterprise use 1 comment

    • 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.

      shubbView on HNParent context

  11. GPT Unicorn has drawn a unicorn 1 comment

    • 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…

      n2d4View on HNParent context

  12. Why ChatGPT Is Getting Dumber at Basic Math 1 comment

    • 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.…

      ndr_View on HNParent context

  13. Why is OpenAI making ChatGPT increasingly dumber? What's the motivation? 1 comment
  14. Electronic Structure of LK-99 1 comment
  15. ChatGPT went from correctly answering simple math 98% of the time to just 2% 1 comment
  16. Ask HN: Is SICP/HtDP still worth reading in 2023? Any alternatives? 1 comment

    • 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…

      embeng4096View on HNParent context

  17. ChatGPT is getting worse over time, Stanford study finds 1 comment
  18. GPT-4 is getting worse over time, not better 1 comment

    • I have not read the paper yet (in my backlog, here's the paper: https://arxiv.org/pdf/2307.09009.pdf), but it's important note that the paper is entitled "How Is ChatGPT’s Behavior Changing over Time?" not that it's necessarily "getting worse." Here's a more nuanced (not an AI…

      lhlView on HNParent context

Submitted to Hacker News

  • How is ChatGPT's behavior changing over time?289 points178 comments
  • Ask HN: Has degradation in the quality of ChatGPT and Claude been proven?42 points40 comments

    Has there been any consensus on this phenomenon I've seen where there are reports of decrease in model performance? This decrease in quality seems to take many forms: laziness, lower expressiveness, mistakes, laziness etc. One the other hand there are people who claim there hasn't been any degradation at all[0]? If there is indeed no degradation how could the perceived degradation be explained? [0] https://community.openai.com/t/declining-quality-of-openai-m...…

  • The AI Reproducibility Crisis4 points3 comments

    I've really been struggling of late to replicate recent findings in research that has built on top of GPT-3.5/GPT-4. This leads me to believe there is growing yet largely unnoticed issue is taking root in recent AI research. I've termed this the "AI Reproducibility Crisis". The principle is simple, if accessible private models are silently changing in time, then previous results cannot be replicated. *Key Issues*: 1. Users have reported significant performance shifts post the May release. 2.…

  • How is ChatGPT's behavior changing over time?3 points0 comments
  • How is ChatGPT's behavior changing over time?2 points0 comments

Connected through shared HN threads, not inferred similarity or a recommendation.