Politeness during interactions with artificial intelligence significantly increases the costs for OpenAI, while also affecting the quality of the system’s responses. The new o3 model has demonstrated a noticeable decline in results in independent tests, raising criticism and questions about the company’s transparency.
This is reported by Finway
A question from a user on social media platform X (formerly Twitter) regarding OpenAI’s energy costs associated with “polite” requests sparked a discussion. Sam Altman, the company’s CEO, noted that such phrases indeed incur “tens of millions of dollars spent not in vain.”
“Politeness towards ChatGPT and other AI assistants is not just anthropomorphism or a cultural habit,” claims Kurt Biebers from Microsoft Copilot.
He explained that polite addresses can change the tone of the dialogue, making the model’s responses more friendly and relevant. However, it is worth noting that the models are also capable of “understanding” ruder forms of communication. This raises questions about how important it is to spend resources on “polite” request processing.
Criticism of the New o3 Model
At the same time, the release of the o3 model from OpenAI faced criticism due to the results of independent tests, which turned out to be significantly lower than those claimed by the company. According to data provided in December 2024, the model reportedly successfully solved over 25% of tasks from the FrontierMath dataset, surpassing competitors. However, the research institute Epoch AI reported in April that the actual figure was around 10%.
This discrepancy in results is explained by the fact that the internal version of o3 was tested using increased computational resources and on a different set of tasks. Moreover, the public version of o3 turned out to be simplified and adapted for real-world use, as confirmed by the ARC Prize Foundation.
OpenAI’s Priorities
OpenAI acknowledged that the final version of the model is focused on speed and efficiency in real scenarios, which has impacted the benchmark results achieved during laboratory tests. This raises questions about the necessity for further improvement of the model to achieve optimal results.
It is worth noting that it was previously reported that OpenAI is working on creating a social network similar to X.