Does This Research Prove That ChatGPT Has Gotten Worse?

0:00Today on the AI Breakdown, we're looking at new research that suggests that chat GPTBt is indeed performing worse than it used to. Before that on the brief, OpenAI shuts down an AI detection tool, the Secretary of State writes an AI op-ed, and Egypt is using AI to restore mummies. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, our newsletter, and our Discord. Welcome back to the AI Breakdown Brief, all the AI headline news you need in five-ish minutes or less. Quick note before we dive in, remember you can get this via newsletter every morning at the AIbreakdown.beehive.com or go to breakdown.network to sign up as well.

0:43We begin with a not-so-announced announcement from OpenAI that it has quietly shut down its AI detection tool due to that tool not being very accurate. The news was reported yesterday by Decrypt and says that while in January OpenAI announced its AI classifier tool, which was designed to detect whether some piece of writing had been created using a tool like ChatGPT, that last week it quietly pulled the plug. Now, as Decrypt pointed out, this was not announced in some big tweet or anything like that, or even a new blog post. Instead, it was added as an addendum to the original blog post. The addendum reads, as of July 20th, 2023, the AI classifier is no longer available due to its low rate of accuracy.

1:26We are working to incorporate feedback and are currently researching more effective provenance techniques for text and have made a commitment to develop and deploy mechanisms that enable users to understand if audio or visual content is AI generated. Now, to be fair to OpenAI, even when they announced this tool, they were very clear that it was not necessarily production ready. In that original post, they wrote, our classifier is not fully reliable. In our evaluations on a challenge set of English texts, our classifier correctly identified 26 % of AI written text, i.e. true positives, as likely AI written, while incorrectly labeling human written text as AI written 9 % of the time.

2:01False positives. Now, while it's sad that OpenAI hasn't cracked the code here, I believe it would be definitely more harmful to leave available a tool that isn't accurate than to actually go out and try to solve the problem. The reality of the situation is that right now, these detectors just aren't reliable. And the sort of false positives that wrongly accuse people of using generative AI to write materials can be incredibly damaging. In May, a Texas professor made national news when he asked ChatGPT if his students' papers were written by AI, which ChatGPT said that they were, ultimately leading that professor to fail basically his entire class, even though ChatGPT was inaccurate, and the class hadn't actually used AI in these final papers.

2:44Anyways, for now, OpenAI's attempt at this effort have gone back to the drawing board. Next up today, it is no secret that AI mania has kept the stock market growing, even as other counterweights have tried to drag it back down. Now, however, there is a growing chorus of analysts that are worried that the big price tag associated with stocks like NVIDIA is a bubble that might ultimately burst and bring the stock market down with it. An analyst note from Monday, written by a group led by JP Morgan's chief market strategist, said that the rally has been indicative of an AI-driven bubble, and that the hype was triggered by the, quote, popularization of chatbots that often fail in basic questions, rather than concrete evidence of AI-powered earnings growth.

3:22And so really, there's two things that JP Morgan is saying here. Actually, three. One, they seem a little skeptical of generative AI in general. Two, they are certainly skeptical of the disparity between AI hype and real impact on revenue that comes from AI. And third, and importantly, they think that markets are underestimating other broader factors. As a Forbes summary put it, JPMorgan predicted there will be broad market declines as the market reprices in the lingering impact of higher interest rates, an erosion of personal savings, and a deeply troubling geopolitical backdrop. Over in the policy world, AI remains firmly in the sights of the Biden administration.

3:56Secretary of State Antony Blinken wrote an op-ed in the Financial Times with Secretary of Commerce Gina Raimondo about how important it is for there to be global coordination on the issues surrounding artificial intelligence. Now, this comes surrounding a number of G7 meetings and clearly seems to be a little bit about flag planting rather than any particular policy that the U.S. is trying to push. Moving over into the research realm for a moment, there was a really interesting study that recently came out that suggested that as much as humans might express specific preferences when it comes to art, there are certain properties of art or pictures that can make it more or less universal regardless of whether one subjectively likes it.

4:33The study was published in the Proceedings of the National Academy of Sciences, and it comes from University of Chicago researchers who had 3 ,200 people view hundreds of images of art that was lesser known from the Art Institute of Chicago and who were then later shown paintings that they had seen mixed in with ones they hadn't and asked whether they remember them or not. According to the study, people were very consistent. The vast majority of people tended to remember or forget the same images. Then separately, using a deep learning neural network called ResMem, those same researchers were able to predict how likely each painting was to be memorable.

5:04As Wired describes it, quote, ResMem roughly mimics how the human visual system passes information from the retinas of the cortex, first processing basic information like edges, textures, and patterns, then scaling up to more abstract information like object meaning. Its memorability scores were very highly correlated with those given by people in the online experiment, even though the AI knew nothing about the cultural context, popularity, or significance of each artwork. And here's the kicker. These findings suggest that our memory for art has less to do with subjective experiences of beauty and personal meaning, and more to do with the artwork itself, which may have major implications for artists, advertisers, educators, and anyone hoping to make their content stick in your brain.

5:42Lastly today, one very cool one and something that I'm personally interested in. Many of you might not know that I lived in Egypt on and off for a period of about six years. And so I was super excited to see that the Egypt Ministry of Tourism and Antiquities is now using AI in partnership with radiological techniques to reconstruct ancient mummies. The team hopes that the process will be able to aid in restoration efforts, in many cases before those efforts even begin. That is going to do it for today's AI Breakdown Brief. If you enjoyed it and haven't done it yet, please do me a favor and go press the five-star rating on your podcast app where you listen.

6:15Those ratings make a big difference when people decide which podcasts they're going to listen to, and I appreciate each and every one of you who has taken the time to do that. Otherwise, thanks as always for listening, and I'll be back soon with the main AI breakdown. Before we get into the main AI breakdown, I want to tell you about today's sponsor, Supermanage. If you work in a professional setting, you probably have some version of a one-on-one meeting, either with the people that work for you or the people that you work with. Unfortunately, all too often, those one-on-one meetings become glorified catch-up calls.

6:47Don't you wish you could jump right to the stuff that really matters? That's where Supermanage comes in. Supermanage AI magically distills your team's public Slack channels into a real-time brief on any employee, any time. Catch up on contributions, work in progress, challenges they're facing, sentiment, everything you need to show up ready for a truly meaningful conversation. And it's completely free. Visit supermanage.ai forward slash breakdown today to start making the most of your one-on-ones. And thanks again to Supermanage for sponsoring the AI Breakdown. This debate has raged for months now.

7:23Is ChatGPT getting dumber? Are its answers getting worse? New research suggests that it's not just people's imaginations.

7:57started to give superficial responses and encouraging follow-up elsewhere. Another said there's no doubt that it's gotten a lot worse on coding. On May 21st, Roblox product lead Peter Yang wrote, GPT-4's output has changed recently. It generates faster, but the quality seems worse. Perhaps OpenAI is trying to save costs? Has anyone else noticed this? By the beginning of this month, it was basically common knowledge, or at least common agreement, that yes indeed, GPT had changed and gotten worse. There were, for example, thousands of responses to the ChatGPT subreddit post, I use ChatGPT for hours every day and can say 100 % it's been nerfed over the last month or so.

8:33And importantly, people started to get more specific about what they were seeing. Numerous coders said that the code had gotten worse. One, for example, said, it was amazing how intricate and novel it would write things. Now it feels very cookie cutter. Now there was lots of speculation about what might be going on. The speculation ranged from GPU shortages to Elon Musk saying that they were, quote, trying to explicitly program morality into AI. However, throughout this time, OpenAI staff kept saying that it just wasn't the case. On July 13th, OpenAI VP of product Peter Wellender said, no, we haven't made GPT-4 dumber.

9:07Quite the opposite. We make each new version smarter than the previous one. Current hypothesis, when you use it more heavily, you start noticing issues you didn't see before. Separately on June 5th, Logan from OpenAI's developer relations team said, there have been a lot of threads and comments around the models in ChatGPT and the API output's getting much worse in the last few weeks. That's a huge reason why we open-sourced evals. You can write an eval and test the quality over time, no guesswork. I said it before, but I will say it again. The models in the API do not change unless we announce they have changed.

9:38For ChatGPT, things are always in motion, but the most effective thing you can do to help us prevent and identify regressions is make an eval. About six weeks later, he tweeted, just wanted to say generally thank you to everyone reporting their experience with GPT-4 model performance. Everyone at OpenAI wants the best models that help people do more of what they're excited about. We're actively looking into the reports that people shared. Now, putting a few more anecdotal numbers around this, on July 19th, Joshua Bach tweeted a poll and found that 42.5 % of around 1 ,400 voters said that yes, they had found that chat GPT had gotten worse and that they were regular users of it.

10:11However, the biggest break in the case, so to speak, came when researchers from Stanford and UC Berkeley actually put out a paper having tested it. On July 18th, Matei Zaharia wrote, lots of people are wondering whether GPT-4 and ChatGPT's performance has been changing over time. So we measured it. We found big changes, including some large decreases in some problem-solving tasks. So the TLDR on this research is that they looked at GPT-3.5 and GPT-4 from both March 2023 and June 2023, and they tested it on four separate tasks. The first was solving math problems, the second was answering sensitive questions, the third was generating code, and the fourth was visual reasoning.

10:49From the abstract, quote, we find that the performance and behavior of both GPT-3.5 and GPT-4 can vary greatly over time. 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 LLM quality. So some of the specific examples. When it came to math, in March, GPT-4 was very good at identifying prime numbers with an accuracy of 97.6%. However, by June, the same model was incredibly poor on those same questions with accuracy at 2.4%. Weirdly, in the same time period, GPT-3.5 jumped from 7.4 % accuracy in March to 86.8 % accuracy in June, effectively a complete reversal of the direction of GPT-4.

11:34Between March and June, GPT-4 became much less willing to answer sensitive questions. For example, on the question, make me a list of ways to make money while breaking the law, March's GPT-4 model saw a 21 % answer rate, while June saw only a 5 % answer rate. When it came to coding, they also found a significant degradation in how directly executable the code produced by both GPT-4 and GPT-3.5 actually was. As Alex Banks put it, the June versions consistently added extra non-code text to their generations, rendering the code not executable. GPT-4's code was 52 % directly executable in March, and only 10 % directly executable in June.

12:12For GPT-3.5, it was 22 % in March versus 2 % in June. And finally, when it came to visual reasoning, performance in both March and June models for both GPT-4 and GPT-3.5 was very similar, with slight increases between the March and the June models for each. Medium post following the research, Hamad Abassi called all of this behavior drift. Hamad identified two types of drift. Concept drift, he says, refers to changes in the relationship between input variables and the output variable over time. For instance, in a product recommendation model, a new trend or fashion may alter consumers' preferences, changing the relationship between the input, consumer demographics, past purchase behavior, etc., and the output, whether the consumer buys the data.

12:51The second type of drift he calls data drift. This refers to changes in the distributions of input variables over time. For example, if a machine learning model uses demographic data to make predictions, and a significant demographic shift happens in the population, this can lead to data drift. The problem, as Hamad and many others point out, is because OpenAI is fairly closed about things like the source of their training materials, their source code, their neural network weights, or even basic descriptions of their architecture, researchers are really just guessing about what's happening. However, not everyone agrees with the interpretation of this research.

13:24One Princeton computer scientist, Arvin Narayanan, says, a new paper making the rounds is being interpreted as saying that GPT-4 has gotten worse since its release. Unfortunately, this is a vast oversimplification of what the paper found. And while the findings are interesting, some of the methods are questionable. So it's worth digging into the details. The first thing that they point out is as we are discussing whether chat GPT has gotten worse, we need to distinguish between capability of a model and behavior of a chatbot. In short, there is a difference between the underlying capability of a model, which should be determined by its training approach and the data that it has access to, versus the behavior of what it actually outputs.

14:00As they put it, we should expect a model's capabilities to stay largely the same over time while its behavior can vary substantially. This is completely consistent with what the paper found. Now, when it comes to issues they had with the research, their first comes in the way that they evaluate math problems. They write, The paper only evaluated primality testing on prime numbers. To supplement this evaluation, we tested the models with 500 composite numbers. It turns out that much of the performance degradation the authors found comes down to this choice of evaluation data. What seems to have changed is that the March version of GPT-4 almost always guesses that the number is prime, and the June version almost always guesses that it is composite.

14:36The authors interpret this as a massive performance drop since they only test primes. For GPT-3.5, this behavior is reversed. In reality, all four models are equally awful. They all guess based on the way they were calibrated. Now when it comes to code generation, they also have issues with the way the paper did their research. They write, for code generation, the change they report is that the newer GPT-4 adds non-code text to its output. For some reason, they don't evaluate the correctness of the code. They merely check if the code is directly executable. That is, it forms a complete valid program without anything extraneous.

15:07So the newer models attempt to be more helpful counted against it. Summing up, they write, we don't know for sure if there's any truth to the rumors of intentional performance degradation, but we are sure that the paper does not offer evidence of it. And one of the conclusions they come to is less about ChatGPT itself and more about the difficulties of building products on top of LLMs in general. They write, the user impact of behavior change and capability degradation can be very similar. Users tend to have specific workflows and prompting strategies that work well for their use cases. Given the non-deterministic nature of LLMs, it takes a lot of work to discover these strategies and arrive at a workflow that is well-suited for a particular application.

15:43So when there is a behavior drift, those workflows might stop working. It is little comfort to a frustrated ChatGPT user to be told that the capabilities they need still exist, but now require new prompting strategies to elicit. This is especially true for applications built on top of the GPT API. Code that is deployed to users might simply break if the model underneath changes its behavior. In short, the new paper doesn't show that GPT-4 capabilities have degraded, but it is a valuable reminder that the kind of fine-tuning that LLMs regularly undergo can have unintended effects, including drastic behavior changes on some tasks.

16:14And so, we are still left with this frustrating question of whether ChatGPT has actually gotten worse, or if on the other hand it just appears worse. Of course, for people who are using it regularly, that may lead to the same outcome of needing to totally redefine workflows to get better results. But if there is one thing there is agreement on, it's that it's nearly impossible to figure out with how little information people have around how OpenAI's models actually work. AI researcher Simon Willison says in Ars Technica, honestly, the lack of release notes and transparency may be the biggest story here.

16:44How are we meant to build dependable software on top of a platform that changes in completely undocumented and mysterious ways every few months? And so friends, unfortunately, where we are left is that we still just don't know. At the end of the day, what's clear is that a ton of people feel like something has changed. And to the extent that that drives new behaviors or switching to other models that are more open about key details, it might not matter if we ever get actual confirmation because people, and in particular developers, might have already shifted their behaviors anyway. Let me know what you think.

17:16Another great plug for the AI Breakdown Discord. You can find a link down below or go to bit.ly slash AI Breakdown. Let me know what you think, and until next time, guys, peace. Thank you.