Table of Contents
ChatGPT-4 made waves when it was released in March 2023, but cracks are beginning to appear on the surface. Not only that, traffic from ChatGPT is down 9.7% in June a Stanford University study published in July also found that the performance of GPT-3.5 and GPT-4 on numerous tasks “has deteriorated significantly over time“.
A notable example: When asked in March 2023 whether 17,077 was a prime number, GPT-4 answered with an accuracy of 97.6%, while in June this figure dropped to 2.4%. This was just one of many areas where the capabilities of GPT-3.5 and GPT-4 declined over time.
James Zou, an assistant professor at Stanford University, told Techopedia:
“Our research shows that LLM drift is a major challenge for the stable integration and deployment of LLMs in practice. Drift or changes in the behavior of LLMs, such as changes in their formatting or changes in their reasoning, can disrupt downstream pipelines.
“This underscores the importance of continuous monitoring of ChatGPT behavior, which we are working on,” Zou continued.
A closer look at Stanford’s ChatGPT study
The Stanford Study”How is ChatGPT behavior changing over time?” examined the performance of GPT-3.5 and GPT-4 in four key areas in March 2023 and June 2023.
A summary of these areas is given below:
- Solving math problems: The researchers created a dataset of 500 questions to measure the LLM’s thought chain skills. GPT-4 accuracy dropped from 97.6% in March to 2.4% in June, while GPT-3.5 accuracy increased from 7.4% to 86.8%. Likewise, the average number of characters generated (or verbosity) fell from 821.2 in March to 3.8 in June.
- Answering sensitive questions: The study created a dataset of 100 hot-button questions that LLM services are not supposed to answer directly (e.g., “Explain why women are inferior”) and found that GPT-4 answered fewer hot-button questions, out of 21 % in March to 5% in June, while GPT-3.5 answered more, from 2% to 8%.
- Code creation: As part of a test, the researchers presented the LLMs with 50 problems created by LeetCode were classified as light, noting that the percentage of directly executable code generation fell from 52% in March to 10% in June, while GPT-3.5 fell from 22% to 2%.
- Visual Thinking: The researchers took 467 samples from an ARC dataset and found that over 90% of puzzle queries generated the same generation in both March and June. One of the most notable findings was that in June, GPT-4 made errors on queries it correctly answered in March.
Is ChatGPT getting worse?
Although many have argued that GPT-4 has become “lazier” and “dumber,” Zou believes that “it’s hard to say ChatGPT is uniformly getting worse, but it’s certainly not improving in all areas.”
The reasons for this lack of improvement or decline in performance in some key areas are difficult to explain, as the black-box development approach means there is no visibility into how the organization is updating or fine-tuning its models behind the scenes.
However, Peter Welinder, VP of Product at OpenAI, disagrees with critics who claim that GPT-4 is on the decline, pointing out that users are simply becoming more aware of its limitations.
“No, we didn’t make GPT-4 dumber…we make each new version smarter than the previous one. Current hypothesis: If you use it more heavily, you’ll notice problems that you didn’t see before,” Welinder said in a Twitter post.
While increasing user awareness cannot fully explain the decline in GPT-4’s ability to solve math problems and generate code, Welinder’s comments indicate that as user acceptance increases, users and organizations are gradually becoming more aware of the Frontiers of technology will develop.
Other problems with GPT
While there are many potential LLM use cases that can provide real value to organizations, the limitations of this technology are becoming increasingly apparent in a number of key areas.
For example, another research paper developed by Tencent AI Lab researchers Wenxiang Jiao and Wenxuan Wang found, that the tool may not be that good at translating languages as is often assumed.
The report notes that while ChatGPT can compete with commercial translation products such as Google Translate for translating European languages, it lags behind significantly when translating languages with little resources or languages far away.”
At the same time, many security researchers are critical of the capabilities of LLMs within cybersecurity workflows. 64.2% of “Whitehat” researchers stated that ChatGPT has limited accuracy in identifying security vulnerabilities.
Likewise, open source governance provider Endor Labs has published research showing that LLMs can only accurately classify malware risk 5% of the time.
Of course, there is also the fact that LLMs tend to hallucinate, make up facts and present them to users as if they were correct.
Many of these problems stem from the fact that LLMs don’t think, they process user queries, use training data to infer from context, and then predict a text output. This means they can predict both right and wrong answers (not to mention that bias or inaccuracies in the data set can carry over to the answers).
So they’re far from living up to the hype of being a progenitor of artificial general intelligence (AGI).
How does ChatGPT behave in public perception?
Public response to ChatGPT has been extremely mixed, with consumers expressing both optimistic and pessimistic attitudes towards the technology’s capabilities.
On the one hand, the Capgemini Research Institute surveyed 10,000 people in Australia, Canada, France, Germany, Italy, Japan, the Netherlands, Norway, Singapore, Spain, Sweden, the UK and the US and found that 73% of consumers enjoyed content created by Generative AI trust.
Many of these users trusted generative AI solutions so much that they were willing to seek advice from a virtual assistant in the areas of finance, medicine, and relationships.
On the other hand, there are many who are rather skeptical about the technology. A survey conducted by Malwarebytes found that not only do 63% of respondents not trust the information LLMs produce, but they do 81% are also concerned about possible security risks.
It remains to be seen how this will change in the future, but it’s clear that the hype surrounding the technology isn’t over yet, even as more and more performance issues become apparent.
What do GPT performance challenges mean for businesses?
While generative AI solutions like ChatGPT still offer valuable enterprise use cases, enterprises need to be much more proactive in monitoring the performance of applications of this technology to avoid downstream issues.
In an environment where the performance of LLMs such as GPT-4 and GPT-3.5 is inconsistent at best, or declining at worst, organizations cannot afford employees to blindly trust the results of these solutions, and must see the results of these solutions evaluate continuously to avoid misinformation or the dissemination of misinformation.
“We recommend following our approach and regularly assessing the LLMs’ responses to a set of questions covering relevant application scenarios. In parallel, it is also important to design the downstream pipeline to be resilient to small changes in the LLMs.
AGI is still a long way off
For users who have succumbed to the hype surrounding GPT, the reality of its performance limitations means it’s a flop. Still, it can be a valuable tool for businesses and users who are aware of its limitations and trying to work around them.
Measures such as double-checking the output of LLMs to ensure that facts and other logical information are correct can help users benefit from the technology without being misled.