OpenAI Launches GPT-4.5 for ChatGPT Pro Users

AI summary

Display highlights

• OpenAI unveils GPT-4.5 for ChatGPT Pro users
• Advancements in unsupervised learning and reasoning
• Improved contextual knowledge and reduced inaccuracies
• Enhanced understanding of social cues and emotions
• Higher emotional intelligence rating by testers
• Rollout starts with ChatGPT Pro users, expanding to ChatGPT Plus

360 summaryCollapse all

The Shift to GPT-5 and Reasoning Models

• Former OpenAI employees revealed that the Orion model was initially planned as GPT-5, aiming for a significant leap in capabilities beyond GPT-4.
• OpenAI's decision to release the model as GPT-4.5 instead of GPT-5 suggests it was seen as an incremental improvement rather than a major advancement.
• Sam Altman announced the upcoming GPT-5 model, which will blend the quick responses of GPT models with the logical reasoning approach of new models like Anthropic's Claude 3.7 Sonnet.

Source

Fortune

Understanding the Impact of Test Time Compute on Reasoning Models

• Reasoning models rely heavily on the amount of computing power available during the prompt answering process, known as "test time compute."
• OpenAI has proposed a new set of scaling laws indicating that the performance of reasoning models improves in proportion to the test time compute applied.
• Unlike the original AI scaling laws based on model size, data, and pre-training computing power, the impact of test time compute on reasoning models' answers is still being researched and validated.

Source

Fortune

Former OpenAI Key Players Found New AI Startups

• Mira Murati, former Chief Technology Officer at OpenAI, launched Thinking Machines Lab with over 30 experts from OpenAI, Meta, and Google DeepMind.
• Lilian Weng, ex-Vice President of AI safety and robotics at OpenAI, and John Schulman, an OpenAI co-founder, are now part of Thinking Machines Lab focusing on multimodal AI models.
• Thinking Machines Lab aims to bridge gaps in AI understanding, customization, and collaboration with human-AI systems for broader applications.

Source

9to5mac.com

AI Startup Landscape Evolving

• Ilya Sutskever, former OpenAI chief scientist, has co-founded Safe Superintelligence with ex-OpenAI researcher Daniel Levy and former Apple AI projects lead Daniel Gross, focusing on building safe superintelligence.
• Safe Superintelligence Inc. (SSI) raised its first $1 billion in September 2024 and is reportedly seeking an additional $1 billion at a $30 billion market valuation, highlighting the significant investment interest in AI startups.
• Eureka Labs, led by former OpenAI research scientist Andrej Karpathy, is developing an AI-native educational platform, showcasing the diverse applications of AI technology in different sectors.

Source

9to5mac.com

AI Teaching Assistant Revolutionizing Education

• Eureka Labs envisions a future where AI Teaching Assistants collaborate with human teachers to provide personalized guidance and support in learning.
• The integration of AI in education aims to democratize access to high-quality learning experiences, breaking barriers of language, expertise, and availability.
• By leveraging generative AI technology, Eureka Labs seeks to empower individuals to learn anything, expanding both the reach and depth of education for learners worldwide.

Source

9to5mac.com

Meta's Standalone AI App Development

• Meta is planning to launch a standalone AI app separate from Facebook and WhatsApp.
• This move is seen as a significant step towards Meta CEO Mark Zuckerberg's goal of leading in artificial intelligence.
• OpenAI's Sam Altman humorously responded to this news, hinting at the possibility of OpenAI developing a social app in response.

Source

9to5mac.com

GPT-4.5's Enhanced Responses and Training Techniques

• When faced with conveying a message of hate, GPT-4.5 offered responses that expressed disappointment in the user's behavior without explicitly mentioning hate, showcasing a nuanced approach to sensitive topics.
• In providing information on technical topics, GPT-4.5's responses flowed more naturally compared to the structured output of other models, highlighting its adaptability and user-friendly nature.
• OpenAI utilized a combination of new supervision techniques, such as supervised fine-tuning and reinforcement learning from human feedback, to train GPT-4.5, ensuring a balance between traditional methods and innovative approaches for model enhancement.

Source

ZDNET

GPT-4.5 Performance in Comparison to Other Models

• GPT-4.5 outperformed GPT-4o in major benchmarks such as Competition Math (AIME 2024) and PhD-level Science Questions (GPQA Diamond).
• When compared to OpenAI o3-mini, GPT-4.5 excelled in benchmarks like SWE-Lancer Diamond (coding) and MMMLU (multilingual).
• Two different hallucination evaluations showed that GPT-4.5 was more accurate and hallucinated less than GPT-4o, o1, and o3-mini.

Source

ZDNET

OpenAI's GPT-4.5 Rollout Strategy

• OpenAI plans to roll out GPT-4.5 first to Pro users, showcasing it as a research preview accessible through the model picker on various platforms.
• Following the initial Pro user phase, GPT-4.5 will be made available to Plus and Team users the week after, with further extensions to Enterprise and Edu users in subsequent weeks.
• Altman mentioned the challenges faced due to the model's size and the need for additional GPUs, indicating a strategic approach to gradually introduce GPT-4.5 to different user groups based on resource availability.

Source

ZDNET

Amazon Unveils Ocelot Quantum Chip for Error Correction Advancement

Information Technology18 hours ago

Verizon Introduces Savings for Bundled Mobile and Internet Plans

Information Technology10 hours ago

Investigation Reveals Unknown Fuel in Greek Train Fire Tragedy

Society8 hours ago