文 | 硅基研究室，作者｜kiki
Whether it is public or media revelations, although OpenAI founder Sam Altman has repeatedly spoiled GPT-5-related news, compared with the founder's hardware, computing power and other ambitions, the most difficult problem at the moment is that a group of loyal users of ChatGPT are finding that GPT-4 is becoming more and more "lazy".
The so-called "lazy" refers to ChatGPT's personal user experience. Recently, on the OpenAI online forum, many users are complaining about the degradation of GPT-4's performance, including poor inference and slower response. One user even bluntly said: "ChatGPT is completely unusable, and today I actually made coffee while waiting for an answer." ”
Compared with the lazy GPT-4, OpenAI is telling more commercialization stories to accelerate monetization.
The first is to look for more individual users on the conversational product ChatGPT, such as the opening of no-login use a few days ago, to compete for more traffic and data. On the other hand, there is the prospect of commercialization for enterprises. ChatGPT COO Brad Lightcap recently revealed that 600,000 users are already working on ChatGPT's enterprise-grade products (including ChatGPT Enterprise and ChatGPT Team). According to Brad Lightcap, 92% of Fortune 500 companies are using ChatGPT in some form, with 100 million people actively using ChatGPT every week. The second is the layout around hardware, computing power and globalization.
For OpenAI, "both want and want" is a certain inevitability, and "model laziness" is also the direct result of its accelerated monetization and horse racing. It's just that for Sam Altman and his OpenAI, there are still many problems and troubles that need to be solved before at least GPT-5 can be released.
GPT-4 is getting lazy again
"I used to be an OpenAI evangelist and always told people how great GPT is and how to use it. But now I don't recommend it anymore because it has become difficult to use effectively. A user recently wrote on the OpenAI online forum.
Some loyal users have chosen to abandon the use of GPT-4 Source: OpenAI forum
In mid-March of this year, a post titled "How to deal with "lazy" GPT-4" was crowded with OpenAI's lazy "victims". They found that while using ChatGPT, the response time was getting slower and the answers to questions were less accurate than expected.
Some individual users are dissatisfied with OpenAI Source: OpenAI Forum
Some people complain that GPT-4 does not follow the instructions they give, and when the user needs it to give the full code, GPT-4 will give a truncated code with placeholders. It's also more prone to errors when asked to give examples for updating the code. Others are also finding that ChatGPT is also getting more and more nonsense, and GPT-4 is now more elusive when asked about its "source", its background, and how it generates answers. Some users even complained about its response speed: "Today I actually brewed coffee while waiting for an answer." ”
Developers complain that GPT-4 is lazy Source: OpenAI Forum
In fact, this isn't the first time ChatGPT has become lazy.
As early as July last year, many users expressed their dissatisfaction on social media and OpenAI developer forums, and the lazy and stupid GPT-4 had problems such as weakened logic, wrong responses, difficulty following instructions, and only remembering the latest prompts. Earlier this year, amid more complaints, Sam Altman directly admitted the fact that GPT-4 "has been lazy" and said that he "has issued a fix to resolve the complaint".
Sam Altman GPT-4「 」,图源:X
At the moment, this round of "lazy complaints" does not seem to be waiting for OpenAI's reply. ChatGPT users in the community are discussing why GPT-4 is lazy and finding solutions to these problems on their own. "It used to be smart, now it's a complete idiot" "I feel like I've been lied to", anger is growing within the community.
As for the reason for "laziness", users even speculated whether OpenAI secretly replaced the model behind it with GPT-3.5, and some believe that OpenAI is focusing more on its "enterprise-level customers" rather than "civilians".
The complaints and complaints of users about the decline in GPT-4's performance actually confirm two points, one is that users have accumulated user minds about GPT-4, and they are eager for OpenAI to launch new products. Second, this is likely to be an important time for OpenAI to release GPT-5, AI angel investor Allie K. Miller said: "They (OpenAI) have a user base, they have subscriptions, and if they find that the number of users is declining, maybe they will release an updated version of the model, timing is key." ”
Users are looking for alternatives, does OpenAI still have any food left?
A not-so-good sign for OpenAI is that as GPT-4 has become lazy, even its loyal users have begun to look for alternatives.
In the above post complaining about OpenAI's laziness, many users mentioned that at least in terms of coding capabilities, Anthropic's Claude Opus 3 appears to be reliable, and it seems to be on par with GPT-4 in terms of performance in actual use. In the Chatbot Arena leaderboard as of March 29, Claude Opus 3 beat GPT-4 to occupy the first position.
As of March 29, Claude Opus 3 beat GPT-4 and ranked first Source: Hugging Face
In the past AI model competition, it has become a consensus that various technology companies "catch up with GPT-4", and "comparison with GPT-4" in various evaluations and performance tables has also become an iron rule. The source of the Chatbot Arena ranking comes from the actual votes of users, and Claude Opus 3 temporarily outperformed GPT-4 on the user experience side, which also shows that at least in the eyes of individual users, OpenAI's so-called user moat is not strong.
On the list, in addition to the old rivals Anthropic's Claude and Google's Gemini, it is known as "French OpenAI", and the name of Mistral AI, which focuses on MoE architecture, is also impressively listed. Previously, Mistral-Medium, the "medium cup version" model, was very popular with developers because of its open source and powerful performance, and the "big cup version" Mistral Large released at the end of February went straight to OpenAI's GPT-4, and by the way, it also officially announced its closed-source cooperation with Microsoft.
There are wolves and tigers, and the jury is still out on when GPT-5 will be released, and how much surplus food does OpenAI have?
OpenAI and Sam Altman are naturally not idle, and in the just-concluded Q1 of 2024, the schedule of the world's hottest artificial intelligence start-up has written three things: one is to find the next entrance to AI, the second is to have greater computing ambitions, and the third is the road to globalization.
First, look for the next entry point to AI, which is OpenAI's ambitious hardware plan. In addition to the humanoid robot Figure 01 that was out of the circle before, according to the latest foreign media reports, Sam Altman is planning to launch an artificial intelligence-powered personal device with a mysterious company founded by former Apple designer Jony Ive, and earlier, Altman led the investment in AI hardware start-up Humane, and OpenAI is also discussing calling it GPT-4 with Vision's object recognition software is embedded in smart glasses from Snapchat's parent company, Spectacles.
Secondly, there is a larger computing power plan. Altman has repeatedly stated in public that computing power is the reason for restricting the evolution of models, and the number of AI server chips is seriously insufficient. In order to seek a more stable supply of chips, in addition to investing in GPU chip companies and establishing chip companies, OpenAI and Microsoft, the "most ironclad ally", were revealed to be planning to spend $100 billion to build an AI supercomputer called "Stargate".
However, in addition to the hardware and chip plan, OpenAI is also in the next global chess game in the past year.
In January this year, Altman started his journey to South Korea, exploring the possibility of cooperation with South Korean chip giants Samsung and SK. Recently, Altman also traveled to Japan to meet with Japanese Prime Minister Fumio Kishida and said that OpenAI is considering Japan as its first office in Asia. Speaking to local reporters in Tokyo, Altman said, "It's amazing to see this technology being adopted in Japan. According to incomplete statistics from the "Silicon-based Research Laboratory", in addition to Japan, OpenAI currently has international offices in London and Dublin.
In the face of greater ambitions, there are new and more intractable troubles
However, in addition to GPT-5 and various ambitious plans, there are many new troubles that need to be solved in front of OpenAI.
As mentioned above, "the model is getting lazy", whether OpenAI will continue to patch up the old model or release GPT-5's big move early has not yet been determined.
Another new nuisance is the data problem. Video giant YouTube recently went to war with OpenAI, saying that it would violate YouTube's rules if OpenAI used its videos to develop the Wensheng video model Sora without permission. OpenAI's chief technology officer, Mira Murati, said in an interview that she didn't know if Sora had been trained on YouTube videos and that the company had not revealed where the data came from.
In a recent report, the New York Times uncovered the source of OpenAI's data. Back in late 2021, OpenAI needed more data, so researchers created a speech recognition tool called Whisper, which can transcribe audio from YouTube videos to generate new conversational text, and ended up transcribing more than 1 million hours of YouTube videos, despite some OpenAI employees discussing possible violations of YouTube's rules. OpenAI did not immediately respond to the comments of the aforementioned media outlets.
The Scaling Law of Big Computing Power and Big Data has laid the foundation for all of OpenAI's work. But nowadays, the problem of data acquisition and use is a problem that artificial intelligence companies, including OpenAI, must face, which involves two levels, one is the competition between large companies at the commercial level, and better user data determines the performance of the model. The second is the issue of user privacy at the social level, and these technology companies must be responsible for the security of user privacy.
In addition to the data issue, it is also possible to ignore the changing sentiment towards technology companies. According to The Information, the current valuation of AI start-ups may be falling from their peaks and returning to reality. The decline in startup price-to-earnings ratios reflects the fact that the outlook for AI startups is not as optimistic as one might think. What more investors want to know is that as the boundaries of the business expand, when these start-ups get more money, in addition to how they spend their money, how they make money, how they deal with more competition, these questions become more important.
For OpenAI, they may need to respond more to the problem of GPT-4 becoming lazy, and a better product experience is still the core moat. As one user on the forum put it, "If there was a better product, I would jump ship like a burning ship." "OpenAI is building a bigger ship, but its users may be choosing to leave at the moment.
Resources:
- 1、Business Insider:Uh-oh — it looks like ChatGPT's AI model got lazy again
- 2、The Information：AI Valuations May Be Coming Down to Earth; A Glimpse of OpenAI’s Search Engine
- 3、The New York Times：How Tech Giants Cut Corners to Harvest Data for A.I.