Today:

Microsoft's new "AI Agent Foundation Model" SHOCKS the Entire Industry! | "Agent AI toward... AGI"

Microsoft introduces an interactive agent Foundation model, suggesting a path toward artificial general intelligence (AGI). This model examines human cognition and agent AI, covering skills like decision-making and language processing. 

The proposed model unifies various training strategies, demonstrating its versatility across robotics, gaming, and healthcare domains. It leverages diverse data sources for effective learning, marking a significant step toward AGI development.

WATCH THE VIDEO ON YOUTUBE

Google pledges 25 million euros to boost AI skills in Europe

Google is investing 25 million euros in Europe to teach people about artificial intelligence (AI). The tech giant aims to bridge knowledge gaps and support businesses using AI. By funding nonprofits and offering free training courses in 18 languages, Google hopes to prevent AI from widening existing inequalities. 

This initiative aligns with their recent $1 billion investment in a data center near London to meet growing internet demands. Adrian Brown from the Centre for Public Impact emphasizes the importance of empowering Europeans with AI skills to ensure inclusive growth.

REUTERS

OpenAI hits $2 bln revenue milestone - FT

OpenAI, backed by Microsoft, raked in a whopping $2 billion in revenue in December, according to the Financial Times. Predictions suggest this figure could double by 2025 due to hefty business interest in their AI tools. The company's success is largely attributed to its ChatGPT product. Investors are bullish, valuing the San Francisco startup at over $80 billion. 

CEO Sam Altman is eyeing further growth, including a chip-building initiative, with potential investors like the UAE. Despite requests for comment, neither Microsoft nor OpenAI responded to Reuters.

REUTERS

Microsoft Copilot AI climbs Google, Apple app store charts after Super Bowl ad (despite errors)

Microsoft's Copilot AI app gained traction after a Super Bowl ad, ranking high in app stores. The ad pitched Copilot as more than a search tool, leveraging AI for various tasks. However, users reported errors, like outdated info about NFL teams. 

Despite this, Copilot's popularity surged. Yet, its success hinges on accuracy and reliability. Microsoft must seize this momentum to secure Copilot's future. In essence, the ad worked, but delivering accurate results will determine its lasting impact.

VENTUREBEAT

Bugcrowd snaps up $102M for a ‘bug bounty’ security platform that taps 500K+ hackers

Bugcrowd, a startup connecting organizations with over 500,000 hackers for bug bounty programs, secures $102M in funding led by General Catalyst. With a total of $180M raised, the valuation has significantly increased since 2020. The investment aims to expand operations, potentially through mergers, and enhance platform functionality, including penetration testing and training for hackers. 

Bugcrowd's model capitalizes on the increasing need for comprehensive security measures in a tech-driven world, tapping into a pool of skilled individuals who can identify vulnerabilities. CEO Dave Gerry notes a growth rate of over 40%, nearing $100M in annual revenues, with plans for further expansion.

TECHCRUNCH

LeoLabs Raises $29 Million to Deliver Enhanced AI-powered Insights for Space Operations

LeoLabs, a space safety and security company, secured $29 million to boost its AI-powered insights for space operations. The investment aims to enhance data delivery for both commercial and government players by improving end-user applications and partnerships. 

LeoLabs, known for its extensive radar network and the largest commercial orbital catalog, utilizes AI to process data for over 20,000 objects in low Earth orbit. The funding will accelerate efforts to ensure space safety amidst the increasing satellite presence. Investors recognize LeoLabs' critical role in managing space expansion responsibly, emphasizing its importance in addressing environmental concerns. The company's expansion reflects the growing demand for space safety solutions.

PR NEWSWIRE

🧠RESEARCH

Animated Stickers: Bringing Stickers to Life with Video Diffusion

"Animated Stickers" by Meta introduces a cool way to bring stickers to life using video diffusion. By adding motion to static images based on text prompts, this tech bridges the gap between still stickers and dynamic videos. It uses a smart two-step fine-tuning method to nail the motion while keeping the sticker's original style. Plus, it's fast, creating eight-frame videos in less than a second, making stickers more fun and engaging.

InternLM-Math: Open Math Large Language Models Toward Verifiable Reasoning

InternLM-Math is a high-performing, open-source math reasoning large language model that enhances abstract reasoning abilities. It integrates various techniques like chain-of-thought reasoning and code interpretation into a unified approach, achieving top results in math benchmarks. The model not only solves and verifies math problems but also helps in developing future math models. It showcases potential in multi-task learning with LEAN for both solving and proving math problems, with resources available on GitHub.

Keyframer: Empowering Animation Design using Large Language Models

Keyframer by Apple is a tool that lets folks make animations from static images using plain talk. It was created after talking to pro animators and engineers, and it's all about mixing writing commands with hands-on tweaks to get animations just right. They also let you ask for different designs to compare and think up new ideas. After testing it with 13 people, they figured out the best ways to ask the tool to do stuff and showed how tweaking and asking can help people make better animations without needing to be a pro from the get-go.

ViGoR: Improving Visual Grounding of Large Vision Language Models with Fine-Grained Reward Modeling

ViGoR by Amazon tackles the issue of large vision language models (LVLMs) often misinterpreting visual data, like imagining things not there or missing key scene parts. It introduces a new framework that boosts how well these models understand visuals by using detailed reward modeling. This approach, cheaper than full supervision and including human and automated evaluations, improves LVLMs' performance on several benchmarks. The team also made a tough dataset for testing LVLMs and will share their data, featuring around 16,000 image-text pairs, to support further research.

SubGen: Token Generation in Sublinear Time and Memory

SubGen is a new method that makes large language models (LLMs) smarter about using memory during token generation, especially for long texts. It introduces a cool trick to group similar bits of data (keys) and sample (values) in a way that uses less space and time, without losing accuracy. This technique, which is both fast and doesn't need much memory, beats other methods in tests, making it a big deal for question-answering tasks that need to remember a lot of information.

DeAL: Decoding-time Alignment for Large Language Models

"DeAL: Decoding-time Alignment for Large Language Models" tackles the challenge of ensuring large language models generate content that aligns with human preferences, not just during training but also at decoding time. It introduces DeAL, a framework that allows for the customization of reward functions to guide model outputs according to specific alignment objectives, such as harmlessness and helpfulness. This approach addresses limitations of current alignment techniques like RLHF, by enabling fine-grained control over the trade-offs between different objectives and improving model adherence to these goals. However, it notes the trade-off in decoding speed as an area for future optimization.

Model Editing with Canonical Examples

The paper presents a new method called "Model Editing with Canonical Examples," which uses single learning examples to guide language models towards desired behaviors. This method is tested in various scenarios, including improving knowledge, reducing social bias, and handling syntactic nuances. It showcases how "sense finetuning," focusing on adjusting specific sense vectors for each example, outperforms traditional finetuning methods, offering significant improvements in model performance. The technique is notably effective, even when applied to a smaller model and used in conjunction with a larger one like GPT-J-6B, showcasing its efficiency and potential for targeted model improvements.

🛠️TOP TOOLS

MGIE - uses PyTorch and multimodal large language models for instruction-based image editing, integrating techniques like IPr2Pr and MagicBrush. 

Alvea - utilizing AI for various applications. It includes instructions for setup, usage, and contributions, aiming to facilitate AI integration into projects.

Warmbox - email warm-up tool designed to improve your email deliverability and ensure that your emails reach the inbox of your leads, rather than getting marked as spam. 

Openart - AI-powered platform that helps users unleash their creativity and realize their artistic vision. It offers various tools and features to assist creators in generating images and exploring their creative potential. 

Shortwave - AI Autocomplete feature saves you time by finishing your emails for you, learning from past messages to mimic your style and provide accurate info. 

🗞️MORE NEWS

Travel startup Layla acquires AI itinerary building bot Roam Around

Travel startup Layla, endorsed by notable investors, acquires Roam Around, an AI itinerary builder. Layla aims to enhance trip planning with Roam Around's expertise, integrating their services and phasing out the Roam Around brand gradually. The move aligns with Layla's strategy of merging strengths for improved user experience. TECHCRUNCH

Fan wiki hosting site Fandom rolls out controversial AI features

Fandom, a popular fan wiki hosting site, introduces controversial AI features like Quick Answers, which previously faced criticism for inaccuracies. Quick Answers provides concise responses to fan queries but was paused due to errors. Now, users can opt out, and admins can review answers to ensure accuracy. Fandom aims to improve user experience based on creator feedback. THE VERGE

The AI chat app being trialed in NSW schools which makes students work for the answers

NSW schools trial NSWEduChat, an AI app tailored to the state's curriculum. Unlike ChatGPT, it prompts critical thinking by asking students for clarification, reducing dependency on external resources. Designed to assist, not provide direct answers, it aims to enhance learning while addressing privacy concerns and maintaining curriculum alignment. THE GUARDIAN

Nvidia CEO Huang says countries must build sovereign AI infrastructure

Nvidia CEO Jensen Huang emphasizes the importance of countries developing their own AI infrastructure to harness economic potential while preserving cultural values. He dismisses fears about AI, comparing it to successfully regulated industries like cars and aviation. Huang's remarks come amid Nvidia's efforts to comply with U.S. restrictions on AI chip exports and its ongoing role in democratizing AI access. REUTERS

AI Scores in Super Bowl Commercials: You Can Watch Them Here

Super Bowl 2024 ads showcased AI's diverse roles. Microsoft's Copilot portrayed AI replacing creatives, Google's Pixel 8 aided visually impaired with Guided Frame, CrowdStrike wielded AI for cybersecurity in the Wild West, while Etsy comically utilized AI for gifting decisions. However, BodyArmor spoofed AI's flaws with absurdity. CNET

How to use ImageFX, Google's new AI image generator

Google's latest AI marvel, ImageFX, competes with other image generators like DALL-E 3 and Microsoft's Image Creator. Utilizing Imagen 2 technology, ImageFX creates watermarked images seamlessly. Users simply log in with their Google account, provide detailed prompts, and download generated images for free. It's intuitive and surprisingly effective. ZDNET