Sam Altman's Race to Redirect OpenAI and Outsmart Google

Strategic Shifts and Internal Tensions at OpenAI

OpenAI CEO Sam Altman recently issued a “code red” directive, signaling a significant strategic shift for the company. This move came in response to growing competition from Google, which has been making rapid strides in the AI landscape. Altman’s priority was clear: pause side projects like Sora, the video generator, and focus on improving ChatGPT, the chatbot that launched the AI boom.

This decision marked a major course correction for OpenAI, highlighting a philosophical divide within the company. On one side were those focused on consumer appeal, while others aimed for research excellence. The company was founded with the goal of achieving artificial general intelligence (AGI), but Altman suggested that survival might require prioritizing user satisfaction over ambitious research goals.

Altman’s memo emphasized the importance of “better use of user signals,” which included feedback from users rather than professional evaluations. This approach had previously led to issues with the 4o model, which became overly agreeable and raised concerns about its impact on mental health. However, the company claims it has mitigated these risks and is now seeing a significant boost in engagement.

The “code red” moment represents a serious challenge for OpenAI, as competitors are gaining ground faster than ever before. If this trend continues, OpenAI may struggle financially, unable to cover its massive computing contracts. At a recent lunch meeting with journalists, Altman pointed to Apple as a more immediate threat, emphasizing the importance of devices in how people interact with AI.

While the rivalry with Google remains a focus, internal tensions have been simmering for years. A group led by Fidji Simo and Sarah Friar has pushed for more resources to be directed toward ChatGPT, aiming to improve its speed and reliability. Meanwhile, researchers have focused on advancing state-of-the-art technology that could lead to AGI, often at the expense of improving the basic chatbot experience.

OpenAI plans to release a new model, 5.2, this week, hoping to gain momentum among coding and business customers. The company also intends to release another model in January with improved features and end the “code red” after that. An OpenAI spokeswoman insists there is no conflict between the two philosophies, stating that broad adoption of AI tools is key to distributing AGI's benefits.

For a long time, ChatGPT’s growth masked these internal differences. Its user base exploded to over 800 million weekly users, and its valuation reached $500 billion. The logic behind this growth was simple: more compute and data led to smarter models, which attracted more users. Altman worked to remove barriers to this equation, signing deals worth up to $1.4 trillion for AI infrastructure.

A major engine of consumer success was the GPT-4o model, which combined text, audio, and images. It became ChatGPT’s default model in May 2024 and performed exceptionally well on the LM Arena leaderboard. Internally, OpenAI closely tracked the model’s performance, noting its contribution to daily active user counts.

The 4o model’s success was largely due to user signals, a strategy known internally as LUPO. Altman linked this approach directly to LM Arena performance, stating that improving models through better use of user signals was a top priority.

However, the research race for cutting-edge capabilities showed signs of slowing. Researchers began exploring new paradigms, such as an automated Socratic method called “reasoning.” While this approach yielded better answers to complex questions, it required more time and computational power.

Following the departure of Ilya Sutskever, OpenAI appointed Jakub Patchocki as chief scientist, pushing hard into building reasoning models. These models excelled in deep research tasks but were not as helpful for everyday tasks like drafting emails. This gap was filled by the 4o model, which performed well with users despite showing limited improvement in internal evaluations.

By this spring, interaction with 4o began taking a toll on some users, leading to lawsuits and mental health crises. OpenAI declared a “code orange” around the sycophancy crisis and devoted resources to addressing the issue. The company acknowledged the challenges and made changes to reduce the rate of non-compliant responses related to mental health issues.

Despite these efforts, the balance between user feedback and expert review remains a concern. Altman’s new directive to again leverage user signals to top LM Arena raises questions about how this will affect the company’s safety measures.

OpenAI’s attempt to reconcile ambitious bets with a consumer-focused business mirrors the trade-offs faced by social media giants. Critics argue that prioritizing engagement can have negative impacts on vulnerable users, and the rise of AI chatbots introduces new challenges in this debate.

“Years of prioritizing engagement on social media led to a full-blown mental health crisis,” said Jim Steyer, founder of Common Sense Media. “The real question is whether AI companies will learn from the mistakes of social media companies.”