Last Week in AI #256: A truly open LLM, Deepfake scams $34 million, FCC cracks down on AI robocalls, how AI is used in music, and more!
AI2 releases LLM with open data, weights, and code, deepfake zoom call scams $34 million from Hong Kong firm, FCC to ban AI-voice robocalls, music industry is carefully exploring Gen AI applications
Top News
Hello OLMo: A truly open LLM
The Allen Institute for AI (AI2) has launched OLMo 7B, a fully open, state-of-the-art large language model (LLM) that includes pre-training data and training code. The OLMo framework is designed to assist researchers in training and experimenting with LLMs, and is available for direct download on Hugging Face and GitHub. The framework includes full pretraining data, training code and model weights, and an evaluation suite. By making OLMo and its training data fully accessible to the public, AI2 aims to foster collaborative development of the best open language model in the world. The OLMo framework is expected to increase precision in AI research, reduce carbon emissions associated with AI development, and provide lasting results by keeping models and their datasets open.
HK firm scammed of $34 million after employee duped by video call with deepfake of CFO
A multinational company in Hong Kong was defrauded of HK$200 million ($34 million) after an employee was deceived by a deepfake video conference call featuring the company's CFO and other staff members. The scammers used publicly available footage to digitally recreate each individual, convincing the employee to transfer the large sum across five bank accounts in 15 transactions. The employee, who works in the finance department, only realized the deception a week after the initial contact. The Hong Kong police are investigating the case, which is the first of its kind in the region, but no arrests have been made so far.
FCC moves to criminalize most AI-generated robocalls
The Federal Communications Commission (FCC) is planning to criminalize unsolicited robocalls that use artificial intelligence (AI) generated voices, following a recent incident where a fake message mimicking President Joe Biden's voice was used to discourage voting in New Hampshire's primary election. The proposed change, which is expected to pass in the coming weeks, would outlaw such robocalls under the Telephone Consumer Protection Act (TCPA), a law that regulates automated political and marketing calls made without the receivers' consent. The FCC has previously used the TCPA to impose hefty fines on illegal robocall activities. This move will empower state attorneys general to take legal action against spammers who use AI, and is welcomed by organizations like AARP, who warn that AI can be used to enhance scams targeting vulnerable groups like seniors.
Inside the Music Industry’s High-Stakes A.I. Experiments
The music industry is diving headfirst into experiments with AI, led by Lucian Grainge, chairman of Universal Music Group (UMG). UMG has used AI to generate music, detect copyright infringement, and engage with fans. Grainge sees the potential in AI but wants to ensure it is deployed thoughtfully and legally. He is collaborating closely with tech companies like YouTube to establish guidelines, while also pushing for fair compensation as AI utilizes UMG's catalog. The industry is hopeful AI can drive new revenue streams, but risks like deepfakes have raised concerns about undermining creativity and IP rights. Overall, the music business views AI as the next frontier, but wants to shape its development to protect artists and labels.
Other News
Tools
Meta’s free Code Llama AI programming tool closes the gap with GPT-4 - Meta's latest update to its code generation AI model, Code Llama 70B, is the largest and best-performing model yet, offering improved accuracy and the ability to handle more queries, closing the gap with GPT-4.
ChatGPT finally has competition — Google Bard with Gemini just matched it with a huge upgrade - Google Bard with Gemini has matched ChatGPT's performance in a chatbot arena, coming second on the leaderboard just behind GPT-4-Turbo, OpenAI’s most advanced model, thanks to a new version of the Gemini Pro-scale model.
Bard generates photos now, finally - Google's Bard chatbot now has AI image generation using Google’s Imagen 2 text-to-image model, positioning it as a competitor to OpenAI’s ChatGPT Plus and offering a free alternative with responsible design features.
Amazon announces AI shopping assistant called Rufus - Amazon introduces AI shopping assistant Rufus to help users search and shop for products by answering conversational questions and using Amazon's product catalog, customer reviews, and Q&As.
Shopify’s ‘Magic’ AI image editor can make any product pics look professional - Shopify's AI image editor uses generative technology to help merchants easily enhance product photos, offering various background styles and conversational search powered by AI.
This robot can tidy a room without any help - A robot equipped with AI successfully tidies rooms by identifying and moving objects, utilizing open-source AI models and tools.
Google Maps experiments with generative AI to improve discovery - Google Maps introduces generative AI feature to provide personalized recommendations based on user queries and preferences, aiming to enhance the discovery of new places.
LLaVA-1.6: Improved reasoning, OCR, and world knowledge - LLaVA-1.6 introduces improved reasoning, OCR, and world knowledge, surpassing Gemini Pro on benchmarks and achieving the best performance among open-source LMMs, with a low training cost and zero-shot Chinese capability.
Microsoft Makes Swift Changes to AI Tool - Microsoft introduces protections to AI tool Designer after reports of nonconsensual use to create nude images of celebrities, including Taylor Swift.
Business
Can This A.I.-Powered Search Engine Replace Google? It Has for Me. - A.I.-powered search engine Perplexity is gaining traction as a potential replacement for Google, with tech insiders and investors praising its effectiveness and potential to challenge Google's dominance in the search engine market.
AI Chip Startup Rebellions Snags Funding to Challenge Nvidia - Rebellions Inc. secures $124 million in funding to develop a next-generation AI chip, joining the competitive market of AI hardware.
AI companies lose $190 billion in market cap after Alphabet and Microsoft report - AI-related companies lost $190 billion in stock market value after disappointing quarterly results from tech giants like Microsoft and Alphabet, highlighting investors' high expectations for AI technology.
Mark Zuckerberg explained how Meta will crush Google and Microsoft at AI—and Meta warned it could cost more than $30 billion a year - Meta's gameplan for AI dominance against Google and Microsoft involves leveraging its walled garden of data, aiming for "general intelligence" and investing billions in infrastructure, despite potential privacy concerns and competition from Google's vast corpus of web data.
DataSnipper, startup that uses AI to eliminate some of the ‘dread’ in accounting, is valued at $1 billion in latest funding round - DataSnipper, a startup valued at $1 billion, uses AI to automate critical tasks for accountants and auditors, helping them extract and link data from various documents and databases, ultimately aiming to alleviate the shortage of trained accountants and make auditing work less onerous.
Mastercard jumps into generative AI race with model it says can boost fraud detection by up to 300% - Mastercard has developed a proprietary generative AI model to enhance fraud detection for thousands of banks in its network, using transformer models and transaction data to assess suspicious transactions in real-time.
India’s population speaks over 100 languages. Microsoft thinks AI can bridge its linguistic gaps - Microsoft's AI for Good initiative aims to use AI to bridge India's linguistic gaps, with projects like Jugalbandi chatbot and VeLLM, while also considering participatory design and potential business opportunities in Asia.
Volkswagen sets up its own AI lab as car industry looks to embrace the tech - Volkswagen establishes its own AI lab to develop AI innovations for its vehicles and collaborate with technology companies.
OpenAI is working on AI education and safety initiative with Common Sense media - OpenAI partners with Common Sense Media to develop AI guidelines and educational materials for teens and educators, aiming to ensure safe and responsible use of AI technology.
A.I. Fuels a New Era of Product Placement - A.I. technology is revolutionizing product placement in videos on platforms like YouTube and TikTok, creating new opportunities for creators and advertisers to generate additional revenue.
Twin Labs automates repetitive tasks by letting AI take over your mouse cursor - Paris-based startup Twin Labs is developing an automation product using AI to replicate human tasks, such as onboarding employees and reordering stock, by training an AI agent to perform these tasks.
Meta to deploy in-house custom chips this year to power AI drive - Meta Platforms plans to deploy a new version of a custom chip to reduce its dependence on Nvidia chips and control costs associated with running AI workloads.
AI Startup ElevenLabs Bans Account Blamed for Biden Audio Deepfake - ElevenLabs bans account responsible for creating a deepfake of Biden's audio.
Amazon terminates iRobot deal, Roomba maker to lay off 31% of staff - Amazon terminates planned acquisition of iRobot, leading to layoffs and regulatory concerns.
Research
This baby with a head camera helped teach an AI how kids learn language - A baby wearing a head camera provided unique data that helped train an AI model to learn language, offering insights into early language learning and the potential for AI to mimic human learning.
From GPT-4 to Gemini and Beyond: Assessing the Landscape of MLLMs on Generalizability, Trustworthiness and Causality through Four Modalities - Assessing the landscape of MLLMs on generalizability, trustworthiness, and causality through four modalities, from GPT-4 to Gemini and beyond.
Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - A new open corpus called Dolma, containing three trillion tokens, has been created for language model pretraining research.
DeepMind's robot chef cooks up 'novel' materials with a side of controversy - AI-driven robot at A-Lab produces supposedly novel materials, but chemists dispute the claim, arguing that the materials are not actually new.
LongAlign: A Recipe for Long Context Alignment of Large Language Models - Recipe for LongAlign: Long context alignment of large language models is crucial for individuals and organizations working with arXivLabs, embracing values of openness, community, excellence, and user data privacy.
Anything in Any Scene: Photorealistic Video Object Insertion - A framework for photorealistic video object insertion is proposed, addressing challenges in generating diverse and high-quality visual content through realistic image and video simulation.
ReplaceAnything3D:Text-Guided 3D Scene Editing with Compositional Neural Radiance Fields - A new AI model, ReplaceAnything3D, allows for text-guided 3D scene editing by erasing and replacing specific objects within a scene, demonstrating high-resolution, multi-stage, and multi-view consistent results.
SliceGPT: Compress Large Language Models by Deleting Rows and Columns - Compressing large language models using SliceGPT by deleting rows and columns to reduce size and improve efficiency.
Corrective Retrieval Augmented Generation - AI technology is being developed to correct and improve the generation of content.
WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models - Building an end-to-end web agent with large multimodal models for individuals and organizations, embracing values of openness, community, excellence, and user data privacy.
Mobile-Agent: Autonomous Multi-Modal Mobile Device Agent with Visual Perception - A new autonomous mobile device agent, Mobile-Agent, is introduced, utilizing visual perception tools for operation localization, self-planning, and self-reflection, and achieving high task completion rates without relying on system code.
YOLO-World: Real-Time Open-Vocabulary Object Detection - Real-time open-vocabulary object detection using YOLO-World has been embraced by individuals and organizations working with arXivLabs.
MobileDiffusion: Subsecond Text-to-Image Generation on Mobile Devices - MobileDiffusion introduces a highly efficient text-to-image diffusion model with fewer than 400 million parameters, enabling sub-second generation of high-quality images on mobile devices.
BootPIG: Bootstrapping Zero-shot Personalized Image Generation Capabilities in Pretrained Diffusion Models - BootPIG introduces a novel architecture for zero-shot personalized image generation, utilizing a bootstrapped learning procedure to train the model in just 1 hour and outperforming existing methods.
Learning Universal Predictors - Neural networks trained on UTM data can learn universal prediction strategies for meta-learning.
Guiding Instruction-based Image Editing via Multimodal Large Language Models - Multimodal large language models facilitate edit instructions and guide image editing.
Concerns
Can Taylor Swift Save Humanity From AI’s Dark Side? - AI-powered image generators are creating illicit deepfake pornography, leading to a broader problem of harmful effects and the need for genuine solutions.
Three ways we can fight deepfake porn - Combatting nonconsensual deepfake porn can be achieved through the use of watermarks, protective shields, and legal measures to hold perpetrators accountable.
As Tech CEOs Are Grilled Over Child Safety Online, AI Is Complicating the Issue - Tech CEOs are grilled by Senators over child safety online, with a focus on the increasing issue of AI-generated child sexual abuse material and the challenges in preventing its spread.
Universal Music Group expected to pull music from TikTok over concerns with AI and artist pay - Universal Music Group is expected to pull its music from TikTok due to concerns about AI-generated content and the platform's treatment of artists.
The New Luddites Aren’t Backing Down - Activists are organizing to combat generative AI and other technologies, reclaiming the misunderstood label of Luddite and seeking to widen the scope of who gets to participate in technological development.
Unions plan pushback on proposed driverless taxi expansion in L.A. - Unions plan to rally against Waymo's driverless taxi expansion in L.A., calling for stricter regulation and expressing concerns about job loss and safety.
Microsoft AI engineer says company thwarted attempt to expose DALL-E 3 safety problems - Microsoft AI engineer discovered vulnerabilities in OpenAI’s DALL-E 3 image generator, urged its removal from public use due to potential for abuse, and faced obstacles from both companies in addressing the issue.
ChatGPT accused of violating EU data privacy rules by Italian regulators - Italian regulators accuse OpenAI's ChatGPT of violating EU data privacy rules, prompting an investigation and a response from the company.
Following lawsuit, rep admits “AI” George Carlin was human-written - AI-generated George Carlin comedy special was actually written by a human, leading to a lawsuit from Carlin's estate for unauthorized use of his name and likeness.
OpenAI Says GPT-4 Poses Little Risk of Helping Create Bioweapons - OpenAI's GPT-4 is deemed to pose minimal risk in contributing to the development of bioweapons.
Policy
AI companies will need to start reporting their safety tests to the US government - AI companies will be required to disclose their safety test results to the US government, as part of a new mandate under the Biden administration to ensure AI systems are safe before release.
China Ups Approvals for Public AI Models in Race to Rival US - China is rapidly approving public release of AI models to catch up with the US in AI technology development and become a world leader by 2030.
Lawmakers propose anti-nonconsensual AI porn bill after Taylor Swift controversy - Lawmakers propose anti-nonconsensual AI porn bill to allow people to sue over faked pornographic images of themselves, following the spread of AI-generated explicit photographs of Taylor Swift.
Analysis
Where do LLMs spend their FLOPS? - The article discusses the allocation of FLOPS in LLMs, the impact of attention mechanisms, the KV cache size, performance changes, and empirical analysis of Llama2 models.
Copyright © 2024 Skynet Today, All rights reserved.