Sam Altman on GPT-5 delays, GPT-4o return, and AI controversies
▼ Summary
– OpenAI CEO Sam Altman and the GPT-5 team addressed user concerns during a Reddit AMA, including requests to bring back GPT-4o.
– A real-time router in GPT-5 determines which model to use for prompts, but initial rollout issues made GPT-5 seem “dumber” than GPT-4o.
– Altman promised improvements to the router system and transparency about which model responds to queries, aiming to enhance performance.
– Users strongly lobbied for GPT-4o’s return for Plus subscribers, prompting Altman to consider allowing continued access while evaluating tradeoffs.
– Altman acknowledged the “chart crime” error from GPT-5’s presentation but avoided discussing it in the AMA, focusing instead on fixing user-reported issues.
OpenAI CEO Sam Altman addressed user concerns about GPT-5’s performance during a recent Reddit Q&A, acknowledging technical hiccups while promising improvements. The discussion revealed frustrations from users who found the new model less effective than its predecessor, GPT-4o, prompting Altman to confirm the team was investigating ways to bring back the older version for paying subscribers.
One major issue stemmed from GPT-5’s real-time routing system, designed to allocate prompts between faster responses and more thoughtful analysis. Altman admitted the feature malfunctioned during rollout, making the AI appear “way dumber” than intended. He assured users that fixes were already in progress, including better transparency about which model handles each query. Additionally, Plus subscribers will see doubled rate limits to help them adapt to the new system without exhausting their monthly allowances.
The Reddit session also revisited OpenAI’s now-infamous presentation blunder, a misleading chart that exaggerated benchmark results. Though Altman avoided addressing it directly during the Q&A, he had previously called it a “mega chart screwup” on social media. The error sparked widespread mockery, with critics joking about AI’s inability to handle basic data visualization. Independent reviewer Simon Willison noted that while GPT-5 performed well overall, generating accurate tables remained a weak spot.
Despite these setbacks, Altman emphasized OpenAI’s commitment to refining the model based on user feedback. The company plans to stabilize performance and explore options for reinstating GPT-4o access for Plus members, though no final decision has been made. As the team works through these challenges, the focus remains on delivering a smarter, more reliable experience for all users.
(Source: TechCrunch)
