Sam Altman discusses the uneven GPT-5 launch, the return of 4o, and the 'chart crime.'

Sam Altman discusses the issues with GPT-5's launch and measures to regain stability and user trust.

: During a Reddit AMA, OpenAI's Sam Altman addressed concerns regarding the unexpected challenges faced during the GPT-5 launch, including user dissatisfaction compared to the earlier GPT-4o model. Altman explained that a malfunction with the real-time router led to discrepancies in performance, making GPT-5 seem less effective, and promised transparency and improvements. Additionally, upon popular demand, he expressed willingness to reconsider enabling 4o for Plus users, emphasizing his commitment to refining model decision boundaries and ensuring stability of the rollout. The session also touched upon a presentation mishap, humorously criticized as 'chart crime,' with Altman acknowledging the error and promising corrective measures moving forward.

Sam Altman, the CEO of OpenAI, took to Reddit for an ask-me-anything session to discuss the recent launch of GPT-5. The event allowed users to share their experiences and inquire about the issues that accompanied the release. Many users voiced their dissatisfaction, noting that GPT-5 was underperforming compared to its predecessor, GPT-4o. Altman addressed these concerns by revealing that the real-time routing feature that dictates which model responds to certain prompts had malfunctioned, resulting in unexpected performance issues that day.

The critical error behind the sluggish performance of GPT-5 was due to what Altman described as the router's 'sev,' a significant development error, rendering the auto-switcher ineffective. This meant that responses weren't being managed optimally across the different models, leading to user frustrations and the perception that GPT-5 was 'dumber.' To resolve this, Altman assured users that interventions were being made to improve the decision boundary configurations and that transparency would be enhanced regarding the model selection process during responses.

Altman's interaction with the community revealed the high demand for reinstating GPT-4o, at least for OpenAI's Plus subscribers. Many users expressed a strong preference for 4o's robustness and familiarity, which they found now lacking in the latest model. Responding to the overwhelming request, Altman didn't dismiss the possibility. He stated that the team would consider this option, as they collected more data to determine the best trade-offs that prioritize user satisfaction and product effectiveness.

On a lighter note, the session took a humorous turn with discussions revolving around what participants dubbed the 'chart crime.' During GPT-5's presentation, a graph was displayed incorrectly, featuring a disproportionately large bar that did not accurately reflect the data being presented. Altman admitted to the error through X, recognizing it as a 'mega chart screwup.' The mistake, although minor, opened the floodgates for jokes and underscored the importance of accuracy in presentations. Despite the gaffe, Altman appreciated the community's feedback, emphasizing ongoing efforts to address and rectify these aspects in future presentations.

Amid these discussions, Altman pledged to double the rate limits for Plus users, aiming to allow them ample opportunity to test and adjust to GPT-5 without the constraint of prompt limitations. By doing so, he hoped to foster a learning environment where users could adapt to the new model's capabilities more comfortably. Collectively, these steps highlight OpenAI's commitment to continuous improvement and open dialogue with its user base, as Altman concluded with a promise to work diligently towards stabilizing GPT-5 and maintaining open communication channels for future enhancements.

Sources: Reddit, OpenAI, TechCrunch

Generative AIinnovationMachine Learning