The first GPT-5 chat record across the entire network has been exposed!
Zhidongxi reported on August 4th that today, OpenAI CEO Sam Altman posted the dialogue record of GPT-5 on X platform, previewing the user experience of GPT-5 in advance.
In the conversation, Altman asked GPT-5 to recommend several thought-provoking TV series centered on AI. GPT-5 sent Altman a list, with the top recommendation being the American TV series "Pantheon" that Altman had previously recommended.
Altman also tweeted that the Software as a Service (SaaS) industry will soon enter the "fast fashion" era, which may mean that the upcoming GPT-5 will profoundly impact software development processes, significantly improving software iteration speed and reducing development costs.
However, many netizens are not impressed by GPT-5's performance. It can be seen that GPT-5 uses a large number of hyphens in its responses, which is one of the sources of the "AI flavor" widely criticized by users.
Moreover, GPT-5 did not perform reasoning, which led to limited reflective ability. The second recommended series deviated from the initial requirement, having nothing to do with AI but instead relating to quantum computing.
In the comments section, a post with over 20,000 views may reflect the collective sentiment of netizens: "GPT-5 seems no different from GPT-4o."
The poster further complained that GPT-5's word choice is unnecessarily exaggerated and fancy, with language quirks identical to GPT-4o.
Altman has recently been prominently promoting GPT-5's capabilities, claiming that "GPT-5 is smarter than us in almost every aspect." The unanimous skepticism reflects the obvious gap between the exposed GPT-5 performance and user expectations.
Recently, foreign media The Information exposed many details behind GPT-5's "difficult birth", revealing the numerous challenges OpenAI faces in technological breakthroughs, team management, and negotiations with partners.
In fact, the release of GPT-5 has been severely delayed. Originally scheduled to be released months ago, it was forced to be downgraded to GPT-4.5 due to limited capability improvements. Researchers discovered that fine-tuning techniques suitable for smaller models do not apply to ultra-large-scale models; additionally, when converting reasoning models to "student models" suitable for chat and API use, their performance significantly declined.
According to informed sources, OpenAI's next-generation flagship model GPT-5 has improved in programming and mathematical tasks compared to existing models, with generated code focusing more on user experience and aesthetics. It is also more efficient in supporting AI agents to execute complex tasks, requiring less human intervention.
However, some informed sources believe that its progress is difficult to compare with the generational leap from GPT-3 to GPT-4.
The progress of universal validators is helping OpenAI develop GPT-5, showing improvements not only in highly verifiable tasks like programming but also in subjective areas such as creative writing.
The entire industry, including xAI and Google, is increasing investment in reinforcement learning. Tworek, responsible for OpenAI's reinforcement learning system, publicly stated that the reinforcement learning system behind OpenAI models is actually the core of AGI.
These new developments also explain why OpenAI executives recently claimed to investors that they are confident in achieving "GPT-8".
Although GPT-5 is still far from AGI, it has some more attractive new features beyond programming and reasoning. According to internal Microsoft testing feedback, GPT-5 has improved the quality of generated code and text without significantly increasing computational resource consumption.
A Microsoft employee said this is because GPT-5 is better at judging the computational intensity required for different tasks, thereby achieving more efficient resource allocation.
Automated programming has become a key focus for OpenAI. One reason is that competitor Anthropic gained an early advantage last year in providing code generation models to developers and tools like Cursor.
OpenAI internally believes that automated programming is not only crucial for the company's future business but also key to automating AI research work.
03. Meta Poaches Employees, Causing Team Turmoil, with Some Staff Refusing to Share New Technology with Microsoft
Altman previously stated that with the current technological path, OpenAI can achieve AI with human-level intelligence, or Artificial General Intelligence (AGI).
However, on the path to achieving AGI, technology is not the only challenge. As the most prominent AI startup, OpenAI constantly faces talent poaching from competitors.
Recently, Meta has hired over ten OpenAI researchers, including those involved in OpenAI's recent core technological advancements. Meta offered "top star" level compensation packages, with some employees receiving compensation packages worth billions of dollars.
This wave of departures and subsequent personnel reorganization has put pressure on OpenAI's senior employees. Last week, OpenAI's Research Vice President Jerry Tworek expressed dissatisfaction with team adjustments in the company's internal Slack to Research Head Mark Chen, saying he needed a week off to reassess, but ultimately did not take leave.
Additionally, some senior researchers are resisting sharing their technological inventions with Microsoft, despite the agreement that Microsoft can use OpenAI's technology until 2030.
While OpenAI and its largest external shareholder Microsoft have close financial relations, there are always frictions around the cooperation agreement terms, with both sides trying to gain more concessions during OpenAI's restructuring of its for-profit division and paving the way for future listing.
According to two people involved in the negotiations, talks are moving in a positive direction. Some points are still under discussion, but some content is becoming clearer, such as Microsoft's expected acquisition of about 33% of shares in OpenAI's for-profit entity.
04. Conclusion: How Long Can OpenAI Maintain Its Advantage?
When GPT-5 is officially released, it will carry extremely high expectations. Last week, Altman mentioned in a podcast that he once posed a question even he couldn't understand, yet GPT-5 could easily answer it.
However, OpenAI's slowdown in model performance over the past year, coupled with repeated discrepancies between promotion and actual capabilities, has raised doubts: Can OpenAI continue to lead in AI capabilities against closed-source competitors like Google and Anthropic, as well as top open-source models like DeepSeek, Qwen, and Kimi?
This article is from the WeChat public account "Zhidongxi", author: Chen Junda, editor: Li Shuiqing, published with authorization from 36kr.
