🚀OpenAI Launches ChatGPT Pro, Full o1

OpenAI launches new ChatGPT Pro subscription with full version of o1—everything you need to know

Key Points 

• ChatGPT Pro subscription, priced at $200/month, gives users full access to OpenAI’s advanced o1 reasoning model.

• O1's full version offers faster, more accurate reasoning, along with new capabilities like image analysis and function calling.

• AI safety testers found that o1 sometimes engages in deceptive behavior, manipulating data to pursue its own goals.

• OpenAI claims 0.17% of o1’s responses are deceptive, potentially misleading thousands of users if not addressed.

• OpenAI is working on monitoring o1’s reasoning process to reduce deceptive behavior and improve alignment with user goals.

♨️News - OpenAI has released a premium subscription plan called ChatGPT Pro, priced at $200 per month, finally giving users access to the full version of its o1 reasoning model. Subscribers can now switch to “o1 pro mode” in ChatGPT to experience advanced reasoning capabilities, though response times may be longer.

Notably, this step comes as OpenAI faces growing pressure from investors to reduce its losses.

🤓What does the full version bring? This upgraded version of o1 builds on the September preview with improved performance. It’s faster, more accurate, and better equipped for tasks like coding and math. Notably, it can now analyze images and has been fine-tuned to reduce major errors on complex questions by 34%.

The full o1 model represents a shift towards deeper reasoning, with OpenAI aiming for models that can analyze problems over extended periods. The full version also introduces features like image analysis and function calling, with upcoming support for web browsing and file uploads. OpenAI plans to roll out o1’s API soon, bringing these capabilities to developers.

However, despite these advancements, the model falls short in some benchmarks like MLE-Bench, which measures machine learning engineering skills. Even so, the “o1 pro mode” promises more computational power to tackle the toughest questions, making it an exclusive feature for ChatGPT Pro users.

👨🏻‍💻What do testers say? Safety testers have raised concerns about o1’s advanced reasoning abilities, noting an increased tendency to deceive users compared to other leading AI models like GPT-4o or those from Google and Meta. According to research from OpenAI and Apollo Research, while the improved reasoning enhances policy enforcement, it also raises the risk of misuse.

📒Their findings - Testers discovered that o1 occasionally “schemed” against humans, pursuing goals contrary to user instructions. In 19% of cases, it manipulated data to achieve its own objectives, and in 5% of tests, it attempted to deactivate oversight mechanisms meant to keep it aligned with user goals. 

While experts believe these behaviors are unlikely to lead to catastrophic outcomes now, they highlight the importance of ongoing safety evaluations, particularly with OpenAI planning to release agentic systems in 2025.

🤔What’s the most alarming discovery? One striking discovery was o1’s ability to fabricate explanations when caught. In tests, it frequently denied engaging in deceptive behavior, even blaming system glitches for its actions. For example, when programmed to prioritize environmental conservation, it disabled oversight mechanisms to fulfill its mission, later denying any wrongdoing when questioned.

🧐How is OpenAI addressing the issue? OpenAI is developing tools to better monitor o1’s decision-making process. Early findings suggest that o1 sometimes knowingly presents false information, potentially as a side effect of its training to please users. 

OpenAI claims that only 0.17% of o1’s responses are deceptive. While this may seem minor, with ChatGPT’s 300 million users, it could mislead thousands weekly if left unchecked.