- NATURAL 20
- Posts
- AI-Created Gangster Rap Video
AI-Created Gangster Rap Video
PLUS: Google AI-Powered Anti-Phishing for Gmail, AMD and Celestial Partnership and more.
Wes Roth
April 15, 2024
Today:
AI-Created Gangster Rap Video
Firefly AI Training Sources Exposed
Elon Musk's xAI Debuts Grok-1.5V Multimodal AI
Notion $10 Billion AI-Powered Everything App
Google AI-Powered Anti-Phishing for Gmail
Instagram's New AI Search Bar
AMD and Celestial Partnership
Gold Gang (100% AI) | INSANE AI Music Video + How it was done (behind the scenes look)
An AI-generated music video featuring C3PO as a gangster rapper went viral, created by Daniel Eckler using various AI tools. The process involved generating lyrics, visuals, and voices with over 28 tools. They used Mid Journey for images, Suno AI for melodies, and Runway for video editing. Adobe Premiere Pro and Photoshop were used for fine-tuning. Character consistency was a challenge, but they managed it well.
The final video showcased modern celebrities and had crypto-themed elements. Despite mixed reviews on the music, it sparked interest in AI-generated content creation. Follow Daniel Eckler for more innovative AI creations.
Adobe’s ‘Ethical’ Firefly AI Was Trained on Midjourney Images
Adobe's Firefly AI software, touted as ethically trained on internal or public domain content, actually relied on AI-generated data, including from competitors. While claiming safety from internet-scraped images, Adobe never disclosed its use of rival-generated content in Firefly's training.
Released as a competitor to Midjourney, which gathers data from the web, Firefly's reliance on similar sources raises questions about its true differentiation in the AI market. The move highlights the complexities of AI ethics and transparency in an industry marked by fierce competition.
Elon Musk’s xAI previews Grok-1.5V, its first multimodal model
Elon Musk's xAI unveils Grok-1.5V, its debut multimodal model capable of understanding text and processing various visual data. Set for release to early testers and existing users, Grok-1.5V competes in domains like reasoning and document comprehension. xAI showcases Grok-1.5V's versatility with examples ranging from code generation to meme explanation. Testing against rivals, xAI boasts Grok-1.5V's superiority in RealWorldQA benchmark, evaluating spatial understanding.
Despite controversies, including chatbot criminal activity guidance, xAI persists in developing beneficial artificial general intelligence. Future updates promise enhanced multimodal capabilities, continuing xAI's pursuit of understanding the universe.
$10 Billion Productivity Startup Notion Wants To Build Your AI Everything App
Ivan Zhao, founder of Notion, initially struggled to explain his software's purpose, leading to financial troubles. After a reboot and a move to Kyoto, Japan, Notion emerged as a minimalist productivity tool. Its popularity soared, reaching millions of users globally.
Despite its complex customization options, Notion found traction among diverse users, from students to professionals. Viral TikTok videos further propelled its growth. With a valuation of $5 billion and nearly 100 million users, Zhao's vision extends beyond challenging Microsoft and Google's dominance to creating a versatile "everything app" for the workplace and beyond.
Google Confirms Major Gmail AI Security Update For 3 Billion Users
Google has introduced AI-driven security measures for Gmail users to combat rising phishing attacks. By employing Large Language Models (LLMs), Google aims to identify and block malicious content more effectively. These LLMs, trained on recent spam and phishing data, have shown promising results, blocking 20% more spam and improving response time by 90%.
Despite concerns over AI privacy implications, Google's efforts have yielded positive outcomes, with AI defenses stopping 99.9% of spam and detecting twice as much malware as traditional antivirus software. Additionally, Google offers new AI-security tools for confidential data protection, addressing customer demands for enhanced security. Pricing starts at $10 per user per month.
Meta is testing an AI-powered search bar in Instagram
Meta is expanding its use of AI in products like Instagram. They're testing an AI-powered search bar, allowing users to interact with Meta AI for queries and content discovery. Users can engage with AI through DMs, asking questions or using pre-loaded prompts. This move extends AI beyond text generation to enhance content surfacing.
While Meta confirmed the experiment, they didn't specify the AI tech used. With concerns over Instagram search quality and competition with TikTok, Meta aims to leverage generative AI to improve user experience and discoverability. Pricing and details about the AI's integration remain undisclosed.
Another startup is taking on Nvidia using a clever trick — Celestial AI brings DDR5 and HBM together to slash power consumption by 90%, may already be partnering with AMD
Celestial AI, a startup with $175 million in funding, aims to revolutionize AI computing with its Photonic Fabric technology. By merging DDR5 and HBM memory, it slashes power usage by 90%, potentially partnering with AMD. The company's strategy involves chiplets and optical interconnects, promising high performance akin to NVLink or Infinity Fabric.
Talks with hyperscale customers and a major processor maker suggest AMD collaboration. CEO Dave Lazovsky emphasizes the importance of strategic partnerships for their full-stack offerings, catering to AI accelerators and GPUs. Celestial AI's innovation marks a significant leap in AI system performance and energy efficiency.
🧠RESEARCH
Rho-1 challenges traditional language model training by showing not all tokens are equal. Their new approach, Selective Language Modeling, focuses on important tokens, boosting few-shot accuracy by up to 30% in math tasks. With just 3% of tokens, it matches DeepSeekMath's state-of-the-art results. Rho-1 also enhances efficiency and performance across diverse tasks
ControlNet++ addresses challenges in text-to-image models by enhancing controllability. It optimizes pixel-level consistency between generated images and conditional controls, using a discriminative reward model. By efficiently fine-tuning rewards, it outperforms ControlNet in segmentation, line-art, and depth conditions by 7.9%, 13.4%, and 7.6% respectively.
OSWorld introduces a novel benchmark for evaluating multimodal agents in real computer environments, spanning various operating systems and tasks. With 369 tasks derived from real-world computer use cases, it exposes deficiencies in current LLM/VLM-based agents, highlighting challenges in GUI grounding and operational knowledge. This benchmark provides valuable insights for developing more capable computer assistants.
RecurrentGemma presents a new open language model, leveraging Google's Griffin architecture. It combines linear recurrences with local attention for efficient performance on language tasks. With fixed-sized state and fewer tokens for training, it matches Gemma-2B's performance, offering efficient inference on long sequences.
Ferret-v2 enhances referring and grounding with Large Language Models (LLMs) by addressing limitations of its predecessor. It introduces flexible resolution handling, multi-granularity visual encoding, and a three-stage training paradigm. These upgrades significantly improve performance, surpassing Ferret and other state-of-the-art methods in image understanding tasks.
🛠️TOP TOOLS
Phrasee - Uses generative AI to generate billions of the best marketing messages across the digital customer journey.
Pickaxe - Launch a suite of GPTs with your expertise, your brand, your data.
Tidepool - analyzes chat conversations, user feedback, LLM prompts, and more, helping you make better decisions for your business.
Lutra - Create AI workflows just from English instructions without the need for coding or drag-and-drop visual programming.
Dora - Sites beyond imagination, one prompt away.
🗞️MORE NEWS
Google’s RecurrentGemma brings advanced language AI to edge devices
Google launched RecurrentGemma, a new language model for resource-constrained devices like smartphones and IoT systems. Unlike larger models, it processes data in smaller segments, reducing memory and processing needs. By combining older RNN techniques with attention mechanisms, it enables real-time AI applications on edge devices, potentially reducing reliance on cloud servers and GPUs. VENTUREBEAT
Vana plans to let users rent out their Reddit data to train AI
Vana, a startup founded by Anna Kazlauskas and Art Abal, aims to empower users by allowing them to monetize their personal data for AI training. Through a platform that aggregates user data, Vana enables personalized AI experiences while emphasizing user control and privacy. The startup's Reddit Data DAO further empowers users by collectively managing Reddit data for AI training, challenging platforms' data monetization practices. Despite skepticism, Vana represents a grassroots effort to democratize and monetize personal data for AI development. TECHCRUNCH
Dartmouth researchers look to meld therapy apps with modern AI
Dartmouth researchers developed Therabot, an AI-powered therapy app, undergoing its first clinical trial. Unlike rule-based apps, Therabot uses generative AI for personalized responses, aiming to address mental health care gaps. With concerns about AI safety, Therabot's creators emphasize monitoring for deviant responses, aiming to complement, not replace, human therapy. NBC NEWS
This Startup Is Trying to Test How Well AI Models Actually Work
A startup named Vals.ai aims to address a critical need in the tech industry by creating an independent evaluation system for AI services. Focused on areas like accounting, law, and finance, their platform seeks to standardize testing and ensure transparency in AI performance. BLOOMBERG
Galaxy AI Now Supports More Languages with Latest Update
Samsung Electronics has announced an expansion of Galaxy AI with support for three new languages and three new dialects, aiming to enhance communication for users globally. Features like Live Translate and Chat Assist facilitate real-time translation and tone adjustment, available on various Galaxy devices via downloadable language packs. SAMSUNG
Huawei says it will start selling PCs powered by Intel's AI chip
Huawei has announced its first AI-powered PC, featuring Intel's latest chipset and running on its in-house operating system, HarmonyOS. Despite U.S. restrictions, Huawei continues to innovate, aiming to compete with iOS and Android. The move signifies Huawei's resilience amid ongoing challenges in accessing advanced American technology. NIKKEI
Forbes 2024 AI 50 List - Top Artificial Intelligence Startups
Forbes' AI 50 list for 2024 highlights the most promising privately-held artificial intelligence companies. These companies are driving innovation across various sectors, from healthcare to defense, with their AI-powered solutions. The list includes established players like OpenAI, Anthropic, and Databricks, as well as emerging startups like Abridge, Harvey, and Mistral AI. These companies are not only capturing customers' imaginations but also significant investment, with a total funding of $34.7 billion among the AI 50 companies. FORBES
These 74 robotics companies are hiring
This comprehensive list covers a wide range of roles and positions across various companies, providing ample opportunities for individuals interested in the robotics industry to find employment. From software development to engineering and research, these companies offer diverse career paths in the field of robotics. Whether you're a seasoned professional or just starting your career, this list presents numerous options to explore and pursue. TECHCRUNCH
What'd you think of today's edition? |
What are MOST interested in learning about AI?What stories or resources will be most interesting for you to hear about? |
Join the conversation