machine learning

As ChatGPT gets “lazy,” people test “winter break hypothesis” as the cause

AI, Biz & IT, chatgpt, ChatGPT lazy, chatgtp, GPT-3.5, GPT-4, machine learning, openai, winter break hypothesis / DJ Henderson / December 11, 2023

only 14 shopping days ’til Christmas —

Unproven hypothesis seeks to explain ChatGPT’s seemingly new reluctance to do hard work.

Benj Edwards – Dec 11, 2023 10: 50 pm UTC

A hand moving a wooden calendar piece that says

In late November, some ChatGPT users began to notice that ChatGPT-4 was becoming more “lazy,” reportedly refusing to do some tasks or returning simplified results. Since then, OpenAI has admitted that it’s an issue, but the company isn’t sure why. The answer may be what some are calling “winter break hypothesis.” While unproven, the fact that AI researchers are taking it seriously shows how weird the world of AI language models has become.

“We’ve heard all your feedback about GPT4 getting lazier!” tweeted the official ChatGPT account on Thursday. “We haven’t updated the model since Nov 11th, and this certainly isn’t intentional. model behavior can be unpredictable, and we’re looking into fixing it.”

On Friday, an X account named Martian openly wondered if LLMs might simulate seasonal depression. Later, Mike Swoopskee tweeted, “What if it learned from its training data that people usually slow down in December and put bigger projects off until the new year, and that’s why it’s been more lazy lately?”

Since the system prompt for ChatGPT feeds the bot the current date, people noted, some began to think there may be something to the idea. Why entertain such a weird supposition? Because research has shown that large language models like GPT-4, which powers the paid version of ChatGPT, respond to human-style encouragement, such as telling a bot to “take a deep breath” before doing a math problem. People have also less formally experimented with telling an LLM that it will receive a tip for doing the work, or if an AI model gets lazy, telling the bot that you have no fingers seems to help lengthen outputs.

“Winter break hypothesis” test result screenshots from Rob Lynch on X.
“Winter break hypothesis” test result screenshots from Rob Lynch on X.
“Winter break hypothesis” test result screenshots from Rob Lynch on X.

On Monday, a developer named Rob Lynch announced on X that he had tested GPT-4 Turbo through the API over the weekend and found shorter completions when the model is fed a December date (4,086 characters) than when fed a May date (4,298 characters). Lynch claimed the results were statistically significant. However, a reply from AI researcher Ian Arawjo said that he could not reproduce the results with statistical significance. (It’s worth noting that reproducing results with LLM can be difficult because of random elements at play that vary outputs over time, so people sample a large number of responses.)

As of this writing, others are busy running tests, and the results are inconclusive. This episode is a window into the quickly unfolding world of LLMs and a peek into an exploration into largely unknown computer science territory. As AI researcher Geoffrey Litt commented in a tweet, “funniest theory ever, I hope this is the actual explanation. Whether or not it’s real, [I] love that it’s hard to rule out.”

A history of laziness

One of the reports that started the recent trend of noting that ChatGPT is getting “lazy” came on November 24 via Reddit, the day after Thanksgiving in the US. There, a user wrote that they asked ChatGPT to fill out a CSV file with multiple entries, but ChatGPT refused, saying, “Due to the extensive nature of the data, the full extraction of all products would be quite lengthy. However, I can provide the file with this single entry as a template, and you can fill in the rest of the data as needed.”

On December 1, OpenAI employee Will Depue confirmed in an X post that OpenAI was aware of reports about laziness and was working on a potential fix. “Not saying we don’t have problems with over-refusals (we definitely do) or other weird things (working on fixing a recent laziness issue), but that’s a product of the iterative process of serving and trying to support sooo many use cases at once,” he wrote.

It’s also possible that ChatGPT was always “lazy” with some responses (since the responses vary randomly), and the recent trend made everyone take note of the instances in which they are happening. For example, in June, someone complained of GPT-4 being lazy on Reddit. (Maybe ChatGPT was on summer vacation?)

Also, people have been complaining about GPT-4 losing capability since it was released. Those claims have been controversial and difficult to verify, making them highly subjective.

As Ethan Mollick joked on X, as people discover new tricks to improve LLM outputs, prompting for large language models is getting weirder and weirder: “It is May. You are very capable. I have no hands, so do everything. Many people will die if this is not done well. You really can do this and are awesome. Take a deep breathe and think this through. My career depends on it. Think step by step.”

As ChatGPT gets “lazy,” people test “winter break hypothesis” as the cause Read More »

Elon Musk’s new AI bot, Grok, causes stir by citing OpenAI usage policy

You are what you eat —

Some experts think xAI used OpenAI model outputs to fine-tune Grok.

Benj Edwards – Dec 11, 2023 8: 13 pm UTC

Illustration of a broken robot exchanging internal gears.

Grok, the AI language model created by Elon Musk’s xAI, went into wide release last week, and people have begun spotting glitches. On Friday, security tester Jax Winterbourne tweeted a screenshot of Grok denying a query with the statement, “I’m afraid I cannot fulfill that request, as it goes against OpenAI’s use case policy.” That made ears perk up online since Grok isn’t made by OpenAI—the company responsible for ChatGPT, which Grok is positioned to compete with.

Interestingly, xAI representatives did not deny that this behavior occurs with its AI model. In reply, xAI employee Igor Babuschkin wrote, “The issue here is that the web is full of ChatGPT outputs, so we accidentally picked up some of them when we trained Grok on a large amount of web data. This was a huge surprise to us when we first noticed it. For what it’s worth, the issue is very rare and now that we’re aware of it we’ll make sure that future versions of Grok don’t have this problem. Don’t worry, no OpenAI code was used to make Grok.”

In reply to Babuschkin, Winterbourne wrote, “Thanks for the response. I will say it’s not very rare, and occurs quite frequently when involving code creation. Nonetheless, I’ll let people who specialize in LLM and AI weigh in on this further. I’m merely an observer.”

A screenshot of Jax Winterbourne's X post about Grok talking like it's an OpenAI product. — Enlarge / A screenshot of Jax Winterbourne’s X post about Grok talking like it’s an OpenAI product.

Jason Winterbourne

However, Babuschkin’s explanation seems unlikely to some experts because large language models typically do not spit out their training data verbatim, which might be expected if Grok picked up some stray mentions of OpenAI policies here or there on the web. Instead, the concept of denying an output based on OpenAI policies would probably need to be trained into it specifically. And there’s a very good reason why this might have happened: Grok was fine-tuned on output data from OpenAI language models.

“I’m a bit suspicious of the claim that Grok picked this up just because the Internet is full of ChatGPT content,” said AI researcher Simon Willison in an interview with Ars Technica. “I’ve seen plenty of open weights models on Hugging Face that exhibit the same behavior—behave as if they were ChatGPT—but inevitably, those have been fine-tuned on datasets that were generated using the OpenAI APIs, or scraped from ChatGPT itself. I think it’s more likely that Grok was instruction-tuned on datasets that included ChatGPT output than it was a complete accident based on web data.”

As large language models (LLMs) from OpenAI have become more capable, it has been increasingly common for some AI projects (especially open source ones) to fine-tune an AI model output using synthetic data—training data generated by other language models. Fine-tuning adjusts the behavior of an AI model toward a specific purpose, such as getting better at coding, after an initial training run. For example, in March, a group of researchers from Stanford University made waves with Alpaca, a version of Meta’s LLaMA 7B model that was fine-tuned for instruction-following using outputs from OpenAI’s GPT-3 model called text-davinci-003.

On the web you can easily find several open source datasets collected by researchers from ChatGPT outputs, and it’s possible that xAI used one of these to fine-tune Grok for some specific goal, such as improving instruction-following ability. The practice is so common that there’s even a WikiHow article titled, “How to Use ChatGPT to Create a Dataset.”

It’s one of the ways AI tools can be used to build more complex AI tools in the future, much like how people began to use microcomputers to design more complex microprocessors than pen-and-paper drafting would allow. However, in the future, xAI might be able to avoid this kind of scenario by more carefully filtering its training data.

Even though borrowing outputs from others might be common in the machine-learning community (despite it usually being against terms of service), the episode particularly fanned the flames of the rivalry between OpenAI and X that extends back to Elon Musk’s criticism of OpenAI in the past. As news spread of Grok possibly borrowing from OpenAI, the official ChatGPT account wrote, “we have a lot in common” and quoted Winterbourne’s X post. As a comeback, Musk wrote, “Well, son, since you scraped all the data from this platform for your training, you ought to know.”