Finally, a GPT 4.5 Review Worth Reading

⚡️ Headlines

🤖 AI

Judge denies Elon Musk's request to block OpenAI for-profit conversion but welcomes trial - A federal judge has denied Elon Musk's request to prevent OpenAI's transition to a for-profit entity but has offered to expedite a trial to address Musk's breach of contract claims against OpenAI and its CEO, Sam Altman. [Associated Press].

Nvidia-backed CoreWeave to acquire AI developer platform Weights & Biases - CoreWeave plans to acquire Weights & Biases to integrate its cloud infrastructure with AI model training services, in a deal valued at approximately $1.7 billion. [Reuters].

Amazon's AWS forms new group focused on agentic AI - Amazon Web Services has established a new group dedicated to agentic artificial intelligence, aiming to automate tasks without user prompts and unlock new possibilities for customers. [Reuters].

Turnitin's new 'canvas' lets students use AI and show their work - Turnitin has announced Turnitin Clarity, an online platform that allows students to utilize approved AI tools in their assignments, providing educators with a structured way to monitor and assess AI-assisted work. [Axios].

Google's rumored 'Pixie' assistant coming to Pixel 10 as 'Pixel Sense' - Google is reportedly launching 'Pixel Sense' with the Pixel 10, an on-device assistant that uses data from various Google apps to provide contextual predictions, operating entirely offline to ensure user privacy. [Android Authority].

AI models make precise copies of cuneiform characters - Researchers have developed an AI approach called ProtoSnap, enabling precise replication of ancient cuneiform characters from tablet images, facilitating the automated copying and deciphering of these scripts. [Phys.org].

🤳 Social Media

Digital Services Act: Our fourth transparency report on content moderation in Europe - TikTok has published its fourth transparency report, detailing content moderation efforts in Europe, including the removal of approximately 18 million violative content pieces and the introduction of new metrics for enhanced transparency. [TikTok Newsroom].

YouTube will soon restrict creators from mentioning certain online gambling sites - Starting March 19th, YouTube will prohibit creators from mentioning, displaying logos, or linking to non-Google-approved gambling services in videos, aiming to protect the community, especially younger viewers. [The Verge].

🔬 Research

Aya Vision: Expanding the Worlds AI Can See - Cohere's research lab, Cohere For AI, has developed Aya Vision, a dataset created with contributions from over 3,000 global researchers, aiming to support diverse languages and address data scarcity in non-Western languages. [Cohere].

Introducing NextGenAI: A consortium to advance research and education with AI - OpenAI has launched NextGenAI, a consortium with 15 leading research institutions, committing $50 million in funding and tools to accelerate research breakthroughs and transform education using AI. [OpenAI].

DiffRhythm: Blazingly Fast and Embarrassingly Simple End-to-End Full-Length Song Generation with Latent Diffusion - Researchers have introduced DiffRhythm, a latent diffusion-based model capable of synthesizing complete songs with both vocals and accompaniment in under ten seconds, maintaining high musicality and intelligibility. [ASLP Lab].

⚖ Legal

Alphabet's Google urges US government to avoid breaking up firm, source says - Google has met with U.S. government officials to persuade them to halt efforts to break up the company amid ongoing antitrust cases related to search and advertising technology, arguing that such measures would harm the economy and national security. [Reuters].

🎱 Random

Nothing's Phone 3A and 3A Pro use AI to organize all your stuff - The Nothing Phone 3A and 3A Pro feature a new 'Essential Space' powered by AI, designed to help users organize important information efficiently, with prices starting at $379. [The Verge].

🔌 Plug-Into-This

On GPT-4.5

Zvi Mowshowitz analyzes the release of GPT-4.5, questioning whether it represents a meaningful leap in intelligence, a costly misstep, or something entirely different. He ultimately argues that GPT-4.5 occupies a distinct niche with strengths in verbal intelligence, contextual adaptation, and aesthetics but does not excel at reasoning or code generation. OpenAI’s positioning of the model, pricing strategy, and the debate over its classification as a "frontier model" also draw significant discussion.

• A Different Kind of Intelligence: GPT-4.5 is described as having a unique form of intelligence, excelling in writing quality, aesthetics, and contextual awareness but lacking in traditional reasoning power. Sam Altman himself warns that it is "not a reasoning model" and will not dominate benchmarks.

• Pricing and Access Issues: The model is expensive, and OpenAI’s uncertainty over whether to continue offering it via API suggests that its computational demands may be unsustainable. Mowshowitz points out speculation that GPT-4.5 may have been trained much earlier but was delayed due to cost concerns.

• Benchmarks and Evaluation Challenges: Standard benchmarks fail to capture GPT-4.5’s strengths. While it performs well on tasks requiring nuance and "taste," its performance on technical or mathematical evaluations is underwhelming. Critics argue that the AI industry lacks proper tools to measure models like this.

• Mixed Reception: Some experts, like Tyler Cowen, praise its aesthetic and conversational qualities, calling it "the first model that feels like talking to a thoughtful person." Others criticize it for being slow, expensive, and underwhelming in hard capability areas, with some even labeling it OpenAI’s first major failure.

• Implications for Future AI Models: The release of GPT-4.5 raises questions about OpenAI’s strategic direction. Mowshowitz suggests that its strengths may hint at a future in which multiple specialized models are integrated, rather than a single dominant AI system.

🧩 GPT-4.5 might not be the groundbreaking leap some expected, but its ability to generate high-quality, contextually aware text suggests that AI models may increasingly be judged not just on logic and reasoning but on their ability to engage, persuade, and "think" in more human-like ways.

Amazon is reportedly developing its own AI reasoning model

Amazon is developing a new AI model, codenamed "Nova," aiming to enhance reasoning capabilities and compete with AI leaders like OpenAI and Anthropic. The model employs a "hybrid reasoning" approach, designed to provide both quick responses and complex, extended thinking. Nova is expected to launch by June 2025.

• Hybrid Reasoning Approach: Nova's architecture enables it to deliver rapid answers for straightforward queries while engaging in deeper, more thoughtful analysis for complex problems.

• Competitive Positioning: Amazon aims for Nova to rank among the top five performers in benchmarks assessing coding, function calling, and mathematical reasoning, positioning it as a direct competitor to models like OpenAI's GPT series and Anthropic's Claude 3.7 Sonnet.

• Cost Efficiency: The development team is focused on making Nova more cost-effective than existing AI models, potentially offering a more affordable alternative for businesses and developers.

• Leadership and Development: The project is led by Amazon's AGI team, under the guidance of head scientist Rohit Prasad, reflecting Amazon's commitment to advancing its in-house AI capabilities.

🚀 Nova is expected to be integrated into Amazon Web Services (AWS), enhancing the platform's AI offerings and providing customers with advanced reasoning capabilities.

World's first "Synthetic Biological Intelligence" runs on living human cells

Cortical Labs has unveiled the CL1, the first commercial "biological computer" integrating human brain cells with silicon hardware to create dynamic neural networks. This Synthetic Biological Intelligence (SBI) system offers a new form of computing that is more adaptive, sustainable, and energy-efficient than traditional AI models. The CL1 is expected to become available to users in the coming months.

• Rapid Learning and Adaptation: The human-cell neural networks within the CL1 learn quickly and flexibly, outperforming traditional silicon-based AI chips used in training large language models like ChatGPT.

• Integrated Design: The CL1 features stable hardware and optimized "wetware"—human brain cells—cultivated on a planar electrode array. This design allows for high control in activating the neural network, facilitating real-time operation.

• Server Stack Deployment: Cortical Labs is assembling these units into a biological neural network server stack, comprising 30 individual units, each containing living cells on their electrode arrays. This server stack is anticipated to go online in the coming months.

• Ethical Considerations: The development of the CL1 has prompted discussions about the ethics of integrating human cells into computing systems, particularly concerning consciousness and sentience.

🧠 Yikes. Clearly these guys never watched The Matrix. For all our sakes, let’s hope that computing based on human cells is less efficient than silicon 🤣😳🦠

 🆕 Updates

📽️ Daily Demo

🗣️ Discourse