Today:

OpenAI Releases SORA 👀 the BEST AI Video Generator | STUNNING visuals, details and physics.

Sora, an AI model by OpenAI, crafts lifelike videos from simple text prompts, enabling users to conjure scenes ranging from humans in cities to distant planets. Developed by the creators of ChatGPT, Sora's capabilities mark a significant leap in AI-generated content. 

Despite its current limitations, Sora holds promise for empowering artists and filmmakers, revolutionizing creative expression. This advancement underscores AI's growing capacity to comprehend and replicate aspects of the physical world, a stepping stone towards artificial general intelligence.

WATCH THE VIDEO ON YOUTUBE

World’s biggest tech companies pledge to fight AI-created election ‘deepfakes’ 

Tech giants like Microsoft, Meta, Google, and others joined forces to tackle deceptive AI in elections. They unveiled an accord to combat AI-generated content that could mislead voters, without calling for an outright ban. The agreement focuses on detecting and labeling such content, emphasizing transparency and education. 

While critics call for more enforceable measures, the industry sees this unity as a step forward. As concerns rise over AI's potential to disrupt elections, regulators worldwide are also considering rules to govern its use. Despite challenges, tech companies are voluntarily taking steps to address the issue amid broader debates about online misinformation.

FINANCIAL TIMES

OpenAI’s Sam Altman Seeks US Blessing to Raise Billions for AI Chips

OpenAI's CEO, Sam Altman, seeks US approval for a major project to ramp up AI chip production globally. This move could raise national security and monopoly concerns in the US. Altman's been in talks with potential investors worldwide but needs Washington's go-ahead to proceed. 

This effort aims to bolster AI chip manufacturing, with potential backers in the US, Middle East, and Asia. However, Altman's plans hinge on securing approval from US authorities, navigating potential antitrust scrutiny and addressing security apprehensions. The venture, if approved, could reshape the AI landscape, but faces regulatory hurdles.

BLOOMBERG

Apple is reportedly working on AI updates to Spotlight and Xcode

Apple is enhancing its software with AI, aiming to introduce features like natural language search in Spotlight and AI-driven code completion in Xcode. Bloomberg reports that Apple is expanding internal testing of these AI features and plans to release them to third-party developers soon. The move follows recent releases of AI tools like MLX framework and Keyframer. Similar to Microsoft’s GitHub Copilot, Apple's tool utilizes large language models to predict and complete code strings. 

While CEO Tim Cook has been cautious about AI, he confirmed more features are coming this year. Expect announcements at Apple's WWDC event and Microsoft's Build 2024.

THE VERGE

Masayoshi Son Seeks to Build a $100 Billion AI Chip Venture

SoftBank's Masayoshi Son eyes a mammoth $100 billion for an AI chip venture dubbed Izanagi, aiming to rival Nvidia and cater to AI's semiconductor needs. Son shifts focus from startups to this ambitious project, envisioning a synergy with Arm Holdings. 

SoftBank may invest $30 billion, with potential Middle Eastern backing for the rest. This move reflects Son's drive to shape the AI landscape, marking a significant pivot for the company.

BLOOMBERG

Sony to partner with Seagate on hard disk drives used for AI -Nikkei

Sony plans to team up with Seagate Technology to produce hard disk drives (HDDs) for AI needs, as per Nikkei. Sony Semiconductor Solutions will commence mass production of drive lasers in May, investing around $33 million in new production lines. 

This collaboration aims to address the increasing demand for AI-driven technology.

REUTERS

🧠RESEARCH

Chain-of-Thought Reasoning Without Prompting

This study busts the myth that language models need fancy prompts to think hard. By tweaking how they pick words, researchers found these brainy bots can naturally reason out problems, kinda like stumbling onto a thought train without needing a ticket. This cool trick not only makes them smarter but also more sure of their answers, outsmarting the old way of doing things.

Generative Representational Instruction Tuning

GRIT blends the best of both worlds for language models, making them ace at generating text and understanding it deeply. By teaching a big brain model to switch gears between creating and digging into texts based on simple cues, GRIT's technique has leapfrogged previous champs across the board. It's like teaching an old dog new tricks, where the tricks make it smarter and faster, especially in handling long-winded documents. Plus, they're sharing their secret sauce for anyone to cook up their own.

A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts

ReadAgent turbocharges language models to digest and recall long texts, mimicking how humans skim through pages and hang onto the juicy bits. It smartly decides what to remember, squishes those memories into neat summaries, and peeks back into texts as needed. On tough reading tests, it blew past old methods, making it a whiz at grasping lengthy documents, stretching its memory span by up to 20 times.

Real-World Fluid Directed Rigid Body Control via Deep Reinforcement Learning

The paper introduces "Box o Flows," a cutting-edge experimental system for testing reinforcement learning (RL) in dynamic, real-world scenarios, especially in fluid dynamics where simulation is tricky. It showcases how state-of-the-art RL algorithms can create complex behaviors through simple rewards and highlights the efficiency of using offline RL for hypothesis testing by reusing data. This work paves the way for applying systematic RL algorithms to complex systems, supported by hands-on experiments and supplementary resources​.

Premier-TACO: Pretraining Multitask Representation via Temporal Action-Driven Contrastive Loss

Premier-TACO is a new multitask feature representation learning method aimed at enhancing few-shot policy learning efficiency in sequential decision-making tasks. By pretraining with a subset of multitask offline datasets, it captures crucial environmental dynamics and is fine-tuned with minimal expert demonstrations. A novel negative example sampling strategy significantly boosts its computational efficiency, making large-scale multitask offline pretraining viable. It's proven effective across various continuous control benchmarks, markedly improving few-shot imitation learning of novel tasks.

🛠️TOP TOOLS

Dittto - offers AI tools to enhance website messaging, focusing on crafting effective hero copy inspired by top SaaS brands. 

Retell AI - offers an API for creating conversational voice AI that feels human-like, aiming to solve the complexity of integrating speech-to-text, LLM, and text-to-speech technologies.

Brick Generator -  is a creative online tool designed as a proof of concept to generate brick models from user-provided text

Journey - a platform that enhances storytelling for businesses by offering responsive slides, video recording, interactive blocks, and embeds from favorite tools. 

Agent Herbie - AI intern for founders, managers and executives.

📲SOCIAL MEDIA

🗞️MORE NEWS

No ‘GPT’ trademark for OpenAI

The US Patent Office rejected OpenAI's trademark application for "GPT," citing it as descriptive rather than distinctive. Despite OpenAI's popularization of the term for their conversational AI model, it's deemed too generic for trademark protection. This decision may lead to diluted branding but doesn't diminish OpenAI's pioneering status in AI. TECHCRUNCH

Google launches a slew of AI initiatives to enhance cybersecurity

Google launched the AI Cyber Defense Initiative to bolster cybersecurity, aiming to tackle the "Defender's Dilemma" with AI tools. The initiative includes open-sourcing Magika, an AI-powered malware detection tool, and fostering partnerships with startups and research institutions. Google aims to empower defenders against evolving cyber threats using AI technologies. CSO

Chipping In: Billionaire-Founded Unicorn Startup Makes Semiconductor ‘Chiplets’ For The AI Boom

Silicon Box, a Singapore-based startup, raised $200 million in a Series B round to advance chiplet technology for semiconductors. Founded by industry veterans, including Marvell Technology's billionaire cofounders, the startup aims to revolutionize chip design by stacking smaller chips for higher performance. This funding brings their total to $410 million, propelling them to unicorn status. FORBES

New chip opens door to AI computing at light speed

The University of Pennsylvania engineers have developed a groundbreaking chip that utilizes light waves instead of electricity to perform complex mathematical computations crucial for AI training. This silicon-photonic (SiPh) chip combines nanoscale manipulation of materials with silicon-based technology to achieve unprecedented processing speeds while consuming less energy. The chip's design facilitates vector-matrix multiplication, a fundamental operation in neural networks, the backbone of modern AI systems. PHYS ORG

Behold, a giant AI-generated rat dick

The scientific journal Frontiers in Cell and Developmental Biology recently published a paper titled "Cellular functions of spermatogonial stem cells in relation to JAK/STAT signaling pathway." However, what caught attention was not the content but rather the inclusion of AI-generated images, one of which depicted an egregiously inaccurate representation of a rat's anatomy. The image, created by an AI art generator called Midjourney, depicted a white rat with an excessively large frontal appendage, labeled with nonsensical terms such as "testtomcels" and "iollotte sserotgomar cell." MASHABLE