It seems OpenAI is actively addressing concerns regarding its GPT language models following critiques that GPT-4 has been becoming “lazy,” “dumb,” and experiencing atypical errors, particularly noted on social media in late November.
Speculation has emerged that GPT-4.5 may have been quietly introduced to certain users, hinted at by some ChatGPT responses. Regardless of this theory's validity, there are indications of meaningful internal enhancements to GPT-4 in recent weeks.
As early as last Thursday, users began reporting enhancements in GPT-4’s performance. Wharton Professor Ethan Mollick, who had previously pointed out a decline in GPT-4’s capabilities in November, has remarked on a significant improvement in the model’s efficiency and response quality, albeit without any evidence of a transition to GPT-4.5. After relying on a code interpreter for corrections, he described the difference as “night and day,” contrasting sharply with his prior experiences of GPT-4 being “unreliable and a little dull for weeks.”
During this time, OpenAI also reopened ChatGPT Plus subscriptions last Wednesday, after halting registrations since November 14. In a post on X, Altman acknowledged user patience while they procured additional GPUs.
While it's uncertain if there’s a direct link between the reopening of ChatGPT Plus signups and the enhancements to GPT-4, the timing is notable. Signups had ceased shortly after OpenAI’s first DevDay, which introduced several new features for the premium version of the AI chatbot. The company had instituted a waitlist for ChatGPT Plus subscriptions, as post-DevDay signups overwhelmed its processing capacity.
User complaints had previously highlighted issues with GPT-4, where it would provide explanations on executing commands rather than executing tasks.
Following these complaints, reports of unusual GPT-4 behavior surfaced, diverging from the typical AI quirks users were accustomed to. One frequent issue involved GPT-4 "back-talking" users or requiring multiple clarifications before executing requests. Another prevalent complaint was its tendency to explain rather than complete commands.
The decline in GPT-4’s performance can be traced back to at least July, with a study indicating a significant drop in accuracy between March and June. Many, including OpenAI Product VP Peter Welinder, suggested that perceived deficiencies in answer quality might stem from psychological effects as the model underwent further updates. Some users have proposed that altering their queries could help achieve better results.
Although OpenAI has remained relatively quiet about its internal developments, Altman’s post regarding GPUs likely serves as a significant indicator of ongoing changes behind the scenes. Reports from April suggested that OpenAI would need over 30,000 GPU units to maintain its commercial performance throughout the rest of 2023, even before the surge in interest observed in November.
Speculation about GPT-4.5 has escalated, fueled by leaks detailing a potential new version. Rowan Cheung, founder of the @therundownai newsletter, shared pricing information and advanced multimodal capabilities associated with this development on X. When Cheung inquired about the authenticity of these details, Altman simply replied, “Nah.”
The GPT-4.5 speculation began with a purported image showing the new model's features and pricing. However, Altman dismissed the rumors as unfounded.
Nevertheless, users who believe they have encountered the proposed update assert that they are experiencing GPT-4.5, citing superior performance. Some even queried the chatbot about its model, receiving responses like “GPT-4.5 Turbo.” This has led many to theorize that OpenAI may be beta testing GPT-4.5 primarily through its mobile applications. However, not all users can replicate these findings, often being told they're using the standard GPT-4. OpenAI employee Will Depue referred to the variations as a “very weird and oddly consistent hallucination.”
Cheung pointed out a post from the official ChatGPT X page featuring emojis depicting a brain and clouds, which he interprets as OpenAI's subtle acknowledgment of the hallucination phenomenon. It’s notable that OpenAI provides early access to paid users, including ChatGPT Plus, developer API users, and enterprise clients. Given that these users are often the first to test new GPT versions and features, those experimenting with the model’s performance likely possess some form of developer API access.
Even if OpenAI is currently testing GPT-4.5, the timeline for any official updates remains unclear, particularly as the company stabilizes from recent organizational challenges. As the product continues to evolve, it’s apparent that any “version” of ChatGPT in use may be less important than the tangible improvements being realized.
Ranking
