GPT-4.5 Fails; AGI Future Halted

PLUS: OpenAI Hosts Scientist AI Event, Gemini Enhances Google Sheets Analysis and more.

Author

Wes Roth
March 03, 2025

Unlock the full potential of your workday with cutting-edge AI strategies and actionable insights, empowering you to achieve unparalleled excellence in the future of work. Download the free guide today!

Today:

  • GPT-4.5 Fails; AGI Future Halted

  • Sora Video Coming to ChatGPT!

  • Pika 2.2 Enhances Video Quality

  • Masayoshi Son Borrows Billions for AI

  • OpenAI Hosts Scientist AI Event

  • Gemini Enhances Google Sheets Analysis

GPT-4.5 Fails. AGI Cancelled. It's all over…

GPT-4.5 offers modest improvements over GPT-4, but it’s not groundbreaking. It doesn't excel in reasoning, speed, or cost-efficiency. Instead, its key advancement is reducing hallucinations. However, it remains the most expensive model, limiting accessibility.

Experts speculate it's primarily designed as a synthetic data generator for training future models. The real test of its impact will come with upcoming reasoning models built on its foundation.

OpenAI plans to integrate its AI video generator, Sora, into ChatGPT, expanding its accessibility beyond its standalone web app. While no timeline is set, this move could attract more users and boost ChatGPT’s premium subscriptions. OpenAI is also developing a Sora-powered image generator and a new version of Sora Turbo. Additionally, the company is exploring a standalone mobile app for Sora to enhance its video creation tools.

Pika Labs has released Video Model 2.2, enhancing video generation with 1080p resolution and clips up to 10 seconds long. A new "picaframes" feature improves key frame transitions, offering smoother animations and greater creative control. While specific use cases and technical details remain undisclosed, the model is available via pika.art. This update builds on previous versions, continuing Pika's push for more advanced AI-driven video creation.

SoftBank CEO Masayoshi Son plans to borrow $16 billion to invest in AI, with a possible $8 billion more in 2026. This aligns with SoftBank’s push into AI, including a potential $25 billion investment in OpenAI. The company has already committed $15 billion to Stargate, a joint AI venture with Oracle and OpenAI. SoftBank aims to strengthen the U.S. AI sector amid global competition, particularly with China.

OpenAI and nine U.S. national labs hosted the "1,000 Scientist AI Jam Session" to explore AI’s role in scientific discovery. Scientists used OpenAI’s o3-mini model to solve research challenges and provide feedback for future AI improvements. U.S. Secretary of Energy Chris Wright emphasized AI's strategic importance. This event reflects OpenAI’s commitment to U.S. AI leadership and builds on partnerships to advance materials science, renewable energy, and bioscience research.

Google Sheets now features a Gemini-powered upgrade, enabling users to analyze data faster and create AI-generated visualizations like heatmaps. The tool provides insights on trends, correlations, and anomalies by generating Python code or using formulas. Users can ask questions such as predicting net income or identifying inventory anomalies. This AI-powered update is now available to all Google Workspace business users, enhancing spreadsheet functionality for data analysis.

🧠RESEARCH

This study explores self-rewarding reasoning in large language models, enabling them to generate, evaluate, and correct their outputs independently. Using a two-stage approach—self-generated data training and reinforcement learning—the models refine their reasoning without external feedback. Tests show improved self-correction, rivaling externally guided systems while reducing computational reliance.

MedVLM-R1 enhances medical vision-language models by generating explicit reasoning, improving transparency and trust. Using reinforcement learning instead of supervised fine-tuning, it discovers human-interpretable reasoning paths. Despite limited training data, it significantly improves accuracy and outperforms larger models, marking progress toward more trustworthy and interpretable AI in medical imaging.

R2-T2 enhances multimodal models by optimizing routing weights at test time, improving vision-language integration without retraining. By adjusting routing dynamically based on correctly predicted samples, it refines expert selection for better task performance. This approach significantly boosts state-of-the-art models on diverse benchmarks with minimal computational overhead.

NeoBERT upgrades BERT-like models with modern architecture, optimized pre-training, and extended context length. Despite its compact size (250M parameters), it outperforms existing encoders on NLP benchmarks. Designed as a seamless replacement for current models, NeoBERT advances bidirectional language understanding and supports easy adoption with open-source code and training resources.

🛠️TOP TOOLS

Mindgrasp AI Free - Transform class lectures and readings into study tools with AI powered summaries, detailed notes, flashcards, quizzes, and an interactive AI-tutor trained on your course material.

ColorMagic - Generate color palettes with AI

HuxliAI - Designed to enhance and streamline academic work for students, leveraging advanced AI technologies to provide comprehensive support.

Amper Music AI - Generate a wide array of music types.

Conker AI - Designed to streamline the process of creating quizzes and assessments for teachers, particularly in the K-12 education sector.

📲SOCIAL MEDIA

🗞️MORE NEWS

  • Flora, an AI startup, introduces an "infinite canvas" for creative professionals, offering a visual interface for generating and refining text, images, and video. It integrates existing AI models, prioritizing artistic control over automation.

  • A study by Stanford, Washington, and Emory universities analyzed 1.5 million texts, revealing widespread AI-assisted writing across industries. AI-generated content surged post-ChatGPT, especially in press releases (24%), consumer complaints (18%), and job ads (15%).

  • Chinese smartphone maker Honor will invest $10 billion in AI development over five years, expanding beyond smartphones to AI-powered PCs, tablets, and wearables. The move aligns with its IPO plans and growing Chinese AI investments.

  • Researchers at the University of Jyväskylä developed an AI tool that detects colorectal cancer with 96.74% accuracy, surpassing previous models. The freely available tool aims to assist pathologists, improving diagnostic speed and efficiency while undergoing clinical validation.

  • Researchers found nearly 12,000 valid API keys and passwords in the Common Crawl dataset, posing security risks for AI models trained on it. Exposed keys include AWS and MailChimp, highlighting insecure coding practices and data vulnerabilities.

  • Chinese AI startup DeepSeek claims a theoretical 545% profit margin, estimating $562,027 daily revenue from its AI models. However, actual earnings are lower due to discounts and free services, raising questions about AI profitability projections.

What'd you think of today's edition?

Login or Subscribe to participate in polls.

Reply

or to participate.