Last Week in AI #255: AI voice scams, flood of bad AI translations, self-rewarding LLMs, FTC probes AI partnerships of tech giants, and more!
Scammers using AI to mimic police voice, web is inundated with low-quality machine translations, new method for LLM self improvement, FTC investigates tech giants' investments in AI startups
Top News
‘‘Where is my son?’ Scammers using AI put Bay Area family through terrifying ordeal
Amy and Andy Trapp almost became the victims of an AI-generated phone scam that convincingly mimicked their son's voice to trick them into handing over cash for his supposed bail from jail. The Trapp family received a phone call from people claiming to be their son and a police officer, narrating a story about how their son had been in a car accident, severely injuring a pregnant woman and landing himself in jail. A fake public defender later called asking for more than $15,000 for their son's bail. The ruse was detected when the Trapp's called the local police station directly, discovering that their son was safe at home. This type of scam is becoming increasingly common due to the proliferation of AI technology that enables highly convincing voice cloning. It calls into question how previously trusted communication mediums, like voice and video, are becoming increasingly compromised by generative AI.
A ‘Shocking’ Amount of the Web Is Already AI-Translated Trash, Scientists Determine
A recent study by researchers at the Amazon Web Services AI lab reveals that over half of the sentences on the internet have been translated into two or more languages, often with deteriorating quality due to poor machine translation (MT). The study, which analyzed a corpus of 6.38 billion sentences, found that 57.1% of the sentences were translated into at least three languages. The quality of translations varies significantly, with "low-resource" languages, particularly those spoken in Africa and the Global South, suffering from insufficient training data, resulting in inaccurate text. The study also found a selection bias towards shorter, "more predictable" sentences from low-quality articles, suggesting that a large portion of the internet in lower-resource languages is poorly machine-translated, raising concerns for the development of large language models in these languages.
Self-Rewarding Language Models
Researchers from Meta propose Self-Rewarding Language Models (SRLMs), a new approach to improving the performance of Large Language Models (LLMs) by incorporating a self-improving reward model. Unlike traditional methods such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), which rely on human preference data and often face limitations due to the quality and size of this data, SRLMs continually update the reward model during LLM alignment, eliminating these bottlenecks. The SRLMs act as instruction-following models, generating responses for prompts and evaluating new instruction-following examples to add to their training set. Experiments show how the proposed technique can improve the performance of a Llama 2 70B model to match or exceed the performance of superior LLMs on AlpacaEval, all without human supervision.
Alphabet, Amazon, Microsoft Face FTC Inquiry on AI Partners
The US Federal Trade Commission (FTC) is investigating Alphabet, Amazon, and Microsoft over their partnerships with AI startups Anthropic and OpenAI, aiming to understand the impact these relationships have on market competition. The probe involves a review of over $19 billion in investment by the tech giants. Concern is growing worldwide that the most promising AI startups are highly dependent on established tech companies for financial and infrastructural support. Critics have accused the corporations of structuring transactions to avoid US merger law. Despite this, Microsoft defends its partnerships as beneficial for competition and innovation. The FTC could use any information discovered in the course of the investigation to begin or support other probes.
Message from a partner:
AI Webby Award - Win Big, Reward Your Team and Celebrate your Clients
Have an exciting AI product or experience? Enter The Webby Awards by their Extended Entry Deadline - Friday, February 9th!
As the leading international honor for excellence on the Internet, The Webby Awards recognizes innovation from world-class brands, agencies, creative studios, media, and tech companies, creators, and more. To keep up with the recent explosion of AI innovation they have introduced a new suite of AI Apps and Experience categories including: Best Use of AI & Machine Learning in Social, Video, Advertising, Media & PR, Responsible AI, and Apps and Experiences categories across Media & Entertainment, Work & Productivity, Financial Services/Banking and many more.
Join past Webby Winners like Shopify, Square, Canva, Microsoft, and Google, and don't miss your last chance to enter the 28th Annual Webby Awards by the Extended Entry Deadline on February 9th, 2024.
Other News
Tools
Tesla finally releases FSD v12 - Tesla releases FSD v12, its last hope for self-driving, introducing end-to-end neural nets to power vehicle controls, with the update being rolled out to customers after being used in the internal test fleet.
Google is using AI to organize and customize your Chrome browser - Google is using AI to enhance Chrome browser with features like tab organization, automatic theme generation, and a "Help me write" tool, aiming to integrate AI into web interaction and creation.
Opera to launch new AI-powered browser for iOS in Europe following Apple’s DMA changes - Opera is launching a new AI-powered browser for iOS in Europe following Apple's DMA changes, allowing developers to offer non-WebKit-based browsers and providing iPhone users with an alternative to Safari.
Introducing Stable LM 2 1.6B - Introducing Stable LM 2 1.6B, a state-of-the-art 1.6 billion parameter small language model trained on multilingual data, with a compact size and speed to lower hardware barriers for developers, and the release of the last pre-training checkpoint and optimizer states for fine-tuning.
OpenAI drops prices and fixes ‘lazy’ GPT-4 that refused to work - OpenAI drops prices for API access and introduces new models, including a fix for the "lazy" GPT-4, while also releasing new text embedding models and a free moderation API.
Business
Waymo looks to launch full fleet of robotaxis in LA - Waymo plans to expand its driverless robotaxi service in Los Angeles, facing potential challenges due to the fallout from Cruise and concerns from regulators, despite its success in San Francisco and claims of safety.
Nvidia, Microsoft, Google, and others partner with US government on AI research program - US government partners with tech giants to launch National Artificial Intelligence Research Resource (NAIRR) pilot program, aiming to provide researchers and educators across the country with access to high-powered AI technologies.
Voice cloning startup ElevenLabs lands $80M, achieves unicorn status - ElevenLabs, a voice cloning startup, has raised $80 million in funding, achieved unicorn status, and faced criticism for misuse of its AI-powered tools, while also attempting to address concerns from voice actors and compete with other synthetic voice startups and Big Tech companies.
Baidu's Ernie AI chatbot to power Samsung's new Galaxy S24 smartphones - Baidu's Ernie AI chatbot will be integrated into Samsung's Galaxy S24 smartphones, enabling real-time call translation and other advanced features.
Ola Founder’s Krutrim Becomes First $1 Billion Indian AI Startup - Ola Founder's Krutrim achieves the milestone of becoming the first $1 billion Indian AI startup.
Alphabet Shares Flirt With Record High on AI Hype - Alphabet's shares are approaching a record high due to the excitement surrounding AI.
Research
Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data - On the power of large-scale unlabeled data for monocular depth estimation. This discusses the benefits and challenges of using massive, diverse, and cheap unlabeled images, as well as the approach of jointly training large-scale labeled and unlabeled images to enhance the model's performance.
New Texas Center Will Create Generative AI Computing Cluster Among Largest of Its Kind - University of Texas at Austin is creating a powerful artificial intelligence hub with a new GPU computing cluster to lead in research and offer world-class AI infrastructure to a wide range of partners, focusing on biosciences, health care, computer vision, and natural language processing.
ChatQA: Building GPT-4 Level Conversational QA Models - Building on the success of ChatGPT, this article introduces ChatQA-70B, a white-box conversational QA model with GPT-4 level accuracy, achieved through a two-stage instruction tuning recipe, an enhanced retriever for retrieval-augmented generation, and careful data curation.
DiffusionGPT: LLM-Driven Text-to-Image Generation System - DiffusionGPT is an all-in-one text-to-image generation system that leverages a Large Language Model (LLM) to seamlessly integrate various generative models, addressing challenges faced by existing stable diffusion models and offering a training-free, efficient, and pioneering solution.
LEGO:Language Enhanced Multi-modal Grounding Model - Advancements in large language models have led to the development of LEGO, a multi-modal grounding model that comprehends inputs across various modalities and addresses the issue of limited data through a diverse and high-quality multi-modal training dataset.
Deep Learning Tackles Deep Uncertainty - Deep learning using neural networks is being used to emulate melt rates at the base of Antarctic ice shelves, offering a faster and potentially more accurate method for modeling future sea level rise.
VMamba: Visual State Space Model - VMamba is a novel visual state space model with global receptive fields and dynamic weights, addressing the computational complexity issue of attention mechanism in visual tasks and achieving promising results across various visual tasks.
WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models - WebVoyager outperforms GPT-4 and text-only setups with a 55.7% task success rate, showcasing the capabilities of large multimodal models in building an end-to-end web agent.
Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models - A unifying framework called Patchscopes allows for inspecting hidden representations of language models, aligning with values of openness, community, excellence, and user data privacy.
New Theory Suggests Chatbots Can Understand Text - AI chatbots like Bard and ChatGPT may have the ability to understand and generate humanlike text, as new research suggests that the largest language models can develop new skills and combine them in a way that hints at understanding, challenging the notion that they are just "stochastic parrots."
Using artificial intelligence and satellites, U of M helps farmers detect aphid infestations - University of Minnesota is using artificial intelligence and satellites to help farmers detect aphid infestations, aiming to create a website or app for farmers to use.
Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text - Detecting machine-generated text using zero-shot detection methods, such as spotting LLMs with binoculars, is a key focus for individuals and organizations working with arXivLabs.
Humans Still Cheaper Than AI in Vast Majority of Jobs, MIT Finds - MIT study finds that humans are still more cost-effective than AI in the majority of jobs, countering fears of widespread job displacement.
Concerns
Akutagawa Prize draws controversy after win for work that used ChatGPT - AI-generated novel wins controversial Akutagawa Prize, sparking debate over the use of ChatGPT in literature.
Taylor Swift Is Living Every Woman’s AI Porn Nightmare - AI-generated nudes of Taylor Swift are spreading across social media platforms, and tech companies are struggling to crack down on the abuse, highlighting the consequences of the rise of AI-generated content and the challenges in mitigating the production of harmful content.
George Carlin Estate Sues Creators of AI-Generated Comedy Special: ‘Computer-Generated Click-Bait’ - George Carlin's estate sues creators of AI-generated comedy special for unauthorized use of the comedian's copyrighted works, denouncing the special as "computer-generated click-bait" that detracts from Carlin's comedic works and harms his reputation.
Man sues Macy’s, saying false facial recognition match led to jail assault - Faulty facial recognition match leads to wrongful arrest and jail assault, highlighting the dangers of technology's use by law enforcement.
San Francisco takes legal action over ‘unsafe,’ ‘disruptive’ self-driving cars - San Francisco is suing the state over the expansion of autonomous car companies in the city, citing serious safety incidents and public nuisance caused by the vehicles.
DOJ and SEC investigate GM-owned self-driving car company Cruise - DOJ and SEC investigate GM-owned self-driving car company Cruise following an October incident where one of its cars hit a pedestrian and dragged her 20 feet, leading to a federal probe and criticism of the company's response and transparency.
Cruise wasn’t hiding the pedestrian-dragging video from regulators — it just had bad internet - Cruise's attempt to send a video of a pedestrian-dragging incident to regulators was hindered by internet connectivity issues, leading to accusations of misleading behavior and a subsequent investigation by the Department of Justice and the Securities and Exchange Commission.
New Hampshire Officials to Investigate A.I. Robocalls Mimicking Biden - AI-generated robocalls impersonating President Biden urged New Hampshire voters not to participate in the primary election, prompting an investigation by state officials.
Policy
Iceland Has Its Own AI George Carlin Moment, Considers Law Against Deepfaking the Dead - Iceland considers restrictions on using AI to reanimate dead people after national broadcaster reanimates beloved comedian for New Year's Eve celebration.
Fun
Guns N’ Roses share AI-generated video for ‘The General’ - Guns N’ Roses released an AI-generated video for their track ‘The General’, combining live footage with animated sequences to create a trippy and bold visual experience.
Copyright © 2024 Skynet Today, All rights reserved.