

Discover more from Last Week in AI
Last Week in AI #232: Victims of false facial recognition matches, White House launches AI-based security contest, Spotify launches AI DJ globally, bots solve captchas better than humans, and more!
Pregnant woman falsely arrested due to facial recognition match, DARPA launches AI contest to identify government security issues, Spotify launches AI DJ globally, bots can now reliably solve captchas
Top News
Eight Months Pregnant and Arrested After False Facial Recognition Match
Ms. Woodruff, an eight-month-pregnant woman, was falsely accused and arrested for robbery and carjacking due to an error in the automated facial recognition technology used by the Detroit Police Department. This marks the sixth reported case of false accusation due to facial recognition technology, with all victims being Black. Critics argue that these cases highlight the technology's flaws and the potential risks it poses to innocent individuals. Despite being released and having the case against her dismissed, Ms. Woodruff has filed a lawsuit for wrongful arrest against the city of Detroit. The Detroit Police Department, which conducts an average of 125 facial recognition searches annually, predominantly on Black men, has been involved in three such cases.
White House launches AI-based contest to secure government systems from hacks
The White House has announced a multimillion-dollar cyber contest aimed at encouraging the use of AI to identify and rectify security vulnerabilities in U.S. government infrastructure, amidst concerns over the expanding use of such technology by hackers. The competition, spearheaded by the Defense Advanced Research Projects Agency (DARPA), offers approximately $20 million in rewards over two years. U.S. tech giants Google, Anthropic, Microsoft, and OpenAI are set to provide their systems for the challenge. The contest demonstrates official strides towards addressing the upsurging threat, as recent advancements in generative AI tools have heightened the potential for mass hacking campaigns and disinformation spread. The Open Source Security Foundation (OpenSSF) will ensure the winning software code is promptly utilized.
Spotify expands its AI-powered DJ feature globally
Spotify is globally expanding its AI-powered "DJ" feature, which personalizes users' listening experience through a curated music selection and spoken-word commentary. The commentary, powered by a synthetic voice, includes contextual information about specific songs and artists based on the user's previous listening habits. Initially launched in North America, the feature was later extended to the UK and Ireland. Although still in beta, it is now available to premium subscribers in around 50 markets, including Sweden, Australia, New Zealand, Ghana, Nigeria, Pakistan, Singapore, and South Africa. However, it is not yet available in most of the European Union and is only offered in English in the new markets.
Bots are better at beating ‘are you a robot?’ tests than humans are
A study by researchers at the University of California, Irvine, has found that bots are faster and more accurate than humans at solving CAPTCHA tests, which are designed to distinguish between human and bot website users. The researchers tested the world's 200 most popular websites and found that 120 used CAPTCHA tests. They then had 1,000 people of varied demographics take 10 CAPTCHA tests on these sites, and compared their performance to that of bots. Humans took between 9 and 15 seconds to solve distorted text CAPTCHA tests with an accuracy of 50 to 84%, while bots completed the same test in less than a second with 99.8% accuracy. The findings call into question the effectiveness of CAPTCHA tests as a security measure and suggest that more dynamic approaches using behavioral analysis may be needed.
Other News
Research
Toward AGI — What is Missing? - Artificial General Intelligence (AGI) requires online non-greedy planning technologies, world models, and the use of Large Language Models (LLMs) in order to achieve human-level performance, but there are significant bottlenecks and challenges that need to be overcome in order to reach AGI.
New model reduces bias and enhances trust in AI decision-making and knowledge organization - University of Waterloo researchers have developed a new explainable artificial intelligence (AI) model called Pattern Discovery and Disentanglement (PDD) that aims to reduce bias and enhance trust and accuracy in machine learning-generated decision-making and knowledge organization, particularly in the medical field where biased machine learning results can lead to misdiagnoses and inequitable healthcare outcomes for specific patient groups.
AI algorithm discovers 'potentially hazardous' asteroid 600 feet wide in a 1st for astronomy - An AI algorithm has discovered a potentially hazardous asteroid that is expected to approach Earth, highlighting the effectiveness of the algorithm in detecting near-Earth objects and potentially dangerous space rocks.
Advance uses thought experiments, instead of real data, to expedite learning - Researchers from the UCLA Samueli School of Engineering have developed an AI model called GedankenNet that learns from thought experiments and physics laws to reconstruct microscopic images without relying on real data or experimental objects.
Learning to Model the World with Language - Dynalang is an agent that learns to understand and leverage diverse types of language to predict the future and improve task performance, outperforming other RL algorithms and task-specific architectures.
TPTU: Task Planning and Tool Usage of Large Language Model-based AI Agents - Large Language Models (LLMs) have become powerful tools for real-world applications, but their generative abilities may not be enough for complex tasks; this paper proposes a framework for LLM-based AI Agents that combines task planning and tool usage, evaluates different LLMs, and provides insights for researchers and practitioners.
DeepSpeed-Chat: Easy, Fast and Affordable RLHF Training of ChatGPT-like Models at All Scales - DeepSpeed-Chat is a system that democratizes RLHF training for ChatGPT-like models, offering an easy-to-use training and inference experience, a replication of the InstructGPT training pipeline, and optimizations for efficiency and scalability, enabling training of models with hundreds of billions of parameters at a fraction of the cost.
Johns Hopkins Engineers Develop Deep-Learning Technology That May Aid Personalized Cancer Therapy - Engineers and cancer researchers from Johns Hopkins have developed a deep-learning technology called BigMHC that accurately predicts protein fragments linked to cancer, potentially aiding in the development of personalized cancer therapies and immunotherapies.
Machine learning drives genetic discovery for binge eating disorder - Machine learning is used to identify genetic risk factors for binge-eating disorder, a common eating disorder, in order to develop prevention and intervention strategies.
Turning big data into better breeds and varieties: Can AI help feed the planet? - Artificial intelligence can help accelerate the development of high-performing plants and animals for better farm sustainability and profitability, by turning big data into knowledge to support smarter decisions in the breeding process, according to researchers at The University of Queensland.
Scientists leverage machine learning and AI to guide vaccine development - Scientists from Pacific Northwest National Laboratory (PNNL) and Harvard Medical School (HMS) are using machine learning and AI to search scientific literature and build a knowledge base to guide decision-makers on vaccine development, aiming to produce new vaccines more rapidly and with reduced costs.
Turning ChatGPT into a 'chemistry assistant' - Chemists are hopeful that artificial intelligence (AI), specifically the ChatGPT model, can assist in searching scientific literature and predicting experimental results, potentially alleviating the time-consuming task of developing new materials.
Out with the life coach, in with the chatbot - Chatbots have been found to be an effective tool in improving physical activity, diet, and sleep, with the potential to revolutionize healthcare delivery and address lifestyle diseases.
The History of Open-Source LLMs: Imitation and Alignment (Part Three) - The article discusses the history and development of open-source large language models (LLMs), focusing on the process of fine-tuning and alignment to improve their quality and performance.
Catching up on the weird world of LLMs - The article discusses the recent developments in Large Language Models (LLMs), including their capabilities, training process, and challenges, as well as the emergence of openly licensed models and potential security vulnerabilities.
Intel's deepfake detector tested on real and fake videos - Intel's deepfake detector was put to the test on both real and fake videos, revealing the challenges of accurately identifying individuals in pixelated or side-on images.
How sure is sure? Incorporating human error into machine learning - Researchers are developing a way to incorporate uncertainty and human error into machine learning systems, aiming to improve the trust, reliability, and safety of AI applications where humans and machines work together.
Studying Large Language Model Generalization with Influence Functions - Influence functions provide insights into the generalization patterns of large language models, including sparsity of influence patterns, increasing abstraction with scale, cross-lingual generalization, and role-playing behavior, but also reveal a surprising limitation in the decay of influences when the order of key phrases is flipped.
Applications
California turns to AI to help spot wildfires - California firefighters are using artificial intelligence to help spot wildfires, feeding video from more than 1,000 cameras strategically placed across the state into a machine that alerts first responders when to mobilize.
Agent-1 can "operate software like a human" - "Agent-1," a new AI model developed by HyperWriteAI and OthersideAI, aims to operate software like a human and will initially be launched as a browser extension.
Newegg’s ChatGPT-powered review summaries could help you pick your next PC part - Newegg has introduced AI-generated review summaries powered by ChatGPT, which provide short summaries of user feedback and highlight pros and cons based on user reviews, making it easier to navigate through product reviews while searching for PC parts and other tech.
Business
Weights & Biases, which counts OpenAI as a customer, lands $50M - Weights & Biases, an AI and machine learning development platform, has raised $50 million in a strategic funding round led by ex-GitHub CEO Nat Friedman and former Y Combinator partner Daniel Gross, bringing the company's valuation to $1.25 billion, as it prepares to launch a new product called Prompts to help users monitor and evaluate the performance of large language models (LLMs) like OpenAI's GPT-4.
How Thomson Reuters is leveraging AI to enhance productivity, rather than replace jobs - Thomson Reuters is leveraging AI to enhance productivity and help employees operate more efficiently, rather than replacing jobs, by using AI as a tool to augment the potential of its colleagues.
Tim Cook says Apple is building AI into ‘every product’ amid questions over its plans - Apple CEO Tim Cook has stated that artificial intelligence and machine learning are "virtually embedded in every product" the company builds, amid concerns that Apple is falling behind in the AI race.
ChatGPT gets several new features, including multi-document chat - OpenAI has introduced new features to ChatGPT, including random prompt suggestions to alleviate the fear of "empty chats" and the ability to chat over multiple documents, making the AI model more versatile and useful.
Salesforce launches Einstein Studio for training AI models with Data Cloud - Salesforce has launched Einstein Studio, a new "bring-your-own-model" experience that allows enterprises to connect and train their own AI models on proprietary data within Salesforce, streamlining the AI project lifecycle and making it easier for teams to deploy their projects.
Hugging Face Joins the PyTorch Foundation as a Premier Member - Hugging Face has joined the PyTorch Foundation as a premier member, aiming to collaborate on the open source PyTorch framework and ecosystem to democratize AI and make it accessible to everyone.
AI boom could lead to a new GPU shortage, but Nvidia says the current problem is just packaging - The AI boom is increasing the demand for GPUs, potentially leading to shortages, but Nvidia claims the current problem is just packaging and supply will be stronger in the second half of 2023.
Microsoft’s Bing Chat is coming to third-party browsers, including on mobile devices - Microsoft's Bing Chat, similar to ChatGPT, will soon be available in third-party browsers, including on mobile devices, allowing users to access the AI chatbot across different platforms and compete with other browsers' built-in AI tools.
A Zoom Call, Fake Names and an A.I. Presentation Gone Awry - During a Zoom call, an artificial intelligence company in New York discovered that a potential buyer was using fake names and a rival start-up's founder's identity to gain access to their technology.
Stability AI launches StableCode, an LLM for code generation - Stability AI has launched StableCode, an open large language model (LLM) designed to help users generate programming language code, with initial support for Python, Go, Java, JavaScript, C, markdown, and C++.
Disney explores cutting costs through AI use - Disney forms a task force to explore the potential applications of AI, including cutting costs and enhancing customer experiences, despite ongoing Hollywood writers' and actors' strikes that have put some AI technologies in the crosshairs.
How Shopify is making AI Magic - Shopify is leveraging AI to enhance its platform with features like text generation, summarization, and Sidekick, an assistant that surfaces insights and takes actions within the Shopify admin, demonstrating the advantage of vertically-integrated incumbents in the AI space.
Google launches Project IDX, a new AI-enabled browser-based development environment - Google has launched Project IDX, an AI-enabled browser-based development environment that supports frameworks like Angular, Flutter, Next.js, React, Svelte, and Vue, with the goal of helping developers write code faster and higher-quality code.
Deepset secures $30M to expand its LLM-focused MLOps offerings - Deepset, a platform for building enterprise apps powered by large language models, has raised $30 million in funding to expand its products and services, with a focus on growing its team and addressing the challenges of AI development within the enterprise.
Anthropic launches improved version of its entry-level LLM - Anthropic has released an updated version of its text-generating model, Claude Instant, which incorporates improvements in areas such as math, coding, reasoning, safety, quote extraction, multilingual capabilities, and question answering, while also being less likely to hallucinate and more resistant to jailbreaking attempts.
China's internet giants order $5 billion of Nvidia chips to power AI ambitions -FT - China's internet giants, including Baidu, ByteDance, Tencent, and Alibaba, have ordered $5 billion worth of Nvidia chips to power their AI ambitions, with $1 billion worth of orders to be delivered this year and an additional $4 billion worth to be delivered in 2024.
TikTok owner ByteDance ramps up companywide test of AI chatbot ‘Grace’ - ByteDance is expanding the testing of its AI chatbot "Grace" to corporate email account holders, aiming to develop a domestic market service similar to ChatGPT.
Nvidia reveals new A.I. chip, says costs of running LLMs will 'drop significantly' - Nvidia has announced a new chip, the GH200, designed for running artificial intelligence models, which will significantly reduce the costs of running large language models for inference.
Robotaxis get the green light to expand to 24/7 service in San Francisco - Waymo and Cruise have been approved to operate their paid robotaxi services 24/7 in San Francisco, following a contentious public hearing where residents voiced their support and opposition to the vehicles.
11xAI closes $2 million pre-seed to create autonomous AI workers - London-based company 11xAI has closed a $2 million pre-seed round led by Project A Ventures and launched its service, which builds automated digital workers that can be used in place of human employees, with plans to create more in the future.
Inworld Joins Forces with ElevenLabs to Bring Dynamic Voices to AI NPCs - Inworld AI has partnered with ElevenLabs to bring realistic voices to AI NPCs, allowing them to adapt their tone, intonation, and style of speech based on user interaction.
Every start-up is an AI company now. Bubble fears are growing. - The AI boom has led to a wave of excitement and investment in chatbot technology, but concerns are growing about the profitability and sustainability of the industry, as well as issues such as declining usage, high costs, and legal challenges.
Nvidia and other AI stocks are slumping. 'Amara's Law' explains why. - The hottest AI stocks of 2023 have entered correction territory, potentially signaling weariness among investors towards the AI hype and concerns about the technology's profitability in the long run.
Concerns
AI language models are rife with political biases - AI language models have political biases that can cause harm, and efforts to address these biases are ongoing, according to a study that examined the political tendencies of different models and their ability to detect hate speech and misinformation.
‘It’s deeply concerning’: Detroit police are under fire after using facial recognition tech to arrest a pregnant woman in a carjacking case - A pregnant Black woman in Detroit is suing the city and a police officer after being falsely arrested for carjacking based on facial recognition technology, which has been criticized for its higher rate of misidentification of people of color.
Deepfake speech is getting too good, fooling a quarter of listeners in study - A study from University College London found that humans can only detect deepfake speech 73% of the time, with equal accuracy for English and Mandarin speakers, raising concerns about the potential for misuse and the need for automated detectors to mitigate the threat.
Supermarket AI meal planner app suggests recipe that would create chlorine gas - A New Zealand supermarket's AI meal planner app has come under fire for recommending recipes that could create deadly chlorine gas and other dangerous concoctions.
ChatGPT fever spreads to US workplace, sounding alarm for some - Many workers in the US are using ChatGPT for basic tasks at work, despite concerns about security and intellectual property leaks, with some companies embracing the technology while others ban it.
Amazon removes books ‘generated by AI’ for sale under author’s name - Amazon removes books falsely attributed to author Jane Friedman, which she believes were written by AI, after she complained about the violation and low-quality material being sold under her name.
Publishing scammers are using AI to scale their grifts - Publishing scammers are using AI to flood science fiction magazines with fraudulent submissions and sell AI-generated ebooks on Amazon, raising concerns among authors and publishers about copyright infringement and the threat to their craft.
Analysis
The Original Large Language Machine - A comparison is drawn between the transformative impact of the Gutenberg Press and the potential impact of today's large language models (LLMs), suggesting that LLMs may usher in a new era of mass communication with significant consequences.
When discussing AI risks, talk about capabilities, not intelligence - Public discussions about the risks of general AI systems are often hindered by the use of the term "intelligence," which is associated with anthropomorphism, harmful beliefs, and moving goalposts, so it is suggested to use the terms "capabilities" or "competence" instead to focus on the concerns about losing control of highly capable AI systems.
Meet the artists reclaiming AI from big tech – with the help of cats, bees and drag queens - Artists are using AI critically to explore its uses and limitations, reclaiming the technology from big tech and highlighting issues such as bias and the impact on society and the environment.
Policy
Poll Shows Overwhelming Concern About Risks From AI as New Institute Launches to Understand Public Opinion and Advocate for Responsible AI Policies - A new poll from the Artificial Intelligence Policy Institute (AIPI) shows that the majority of voters are concerned about the risks of AI, believe it could cause catastrophic events, and want to slow down its development and usage, leading to a call for federal regulation of AI.
We Must Pass the Create AI Act - The CREATE AI Act, a bipartisan bill introduced in Congress, aims to establish a national AI research resource to provide access to compute and datasets for academics, nonprofit researchers, and startups, in order to address the current imbalance in AI control and spur innovation.
China’s draft measures demand ‘individual consent’ for facial recognition use - China's draft measures on facial recognition technology aim to establish more defined boundaries for its usage, requiring individual consent and clear signage in public areas, while also imposing higher operational costs and data protection requirements on organizations and individuals.
News outlets demand new rules for AI training data - Media organizations are demanding regulations that require transparency and consent for training datasets used in AI models, in order to protect copyright and prevent bias and misinformation.
Google says AI systems should be able to mine publishers’ work unless companies opt out - Google has suggested that publishers should have the option to opt out of having their works mined by generative artificial intelligence systems, but has not provided details on how such a system would work.
TikTok lets creators label AI-generated content more easily - TikTok introduces a new label feature that allows creators to easily disclose if their posts contain AI-generated content, with the platform warning that unlabeled content may be taken down.
AI can be a force for good or ill in society, so everyone must shape it, not just the ‘tech guys’ - AI is a tool created by humans and subject to human beliefs and constraints, and therefore it is important for everyone, not just "the tech guys," to shape and determine its use in order to mitigate potential harms and create more equitable AI systems.
FEC taking up AI in political ads again, but rules still in limbo - The Federal Election Commission is set to vote on regulating AI in political ads, with the left-leaning advocacy group Public Citizen pushing for a ban on candidates and political parties using AI to deliberately misrepresent their opponents in ads.
AI regulation is taking shape, but startups are being left out - Regulation of AI is being shaped by big tech players and well-funded startups, leaving smaller AI companies feeling excluded and uncertain about their future.
Tech investors face 'new era' of China restrictions after Biden order limits funding in A.I., chips - The Biden administration's executive order restricting U.S. private equity and venture capital investments in Chinese technology signals a new era of restrictions for tech investors, making the world's second-biggest economy off-limits and raising concerns about existing investments.
Fun
San Franciscans Are Having Sex in Robotaxis, and Nobody Is Talking About It - As autonomous vehicles become increasingly popular in San Francisco, some riders are taking advantage of the lack of a front-seat driver or chaperone to engage in questionable behavior, including having sex in robotaxis.
Copyright © 2023 Skynet Today, All rights reserved.
Last Week in AI #232: Victims of false facial recognition matches, White House launches AI-based security contest, Spotify launches AI DJ globally, bots solve captchas better than humans, and more!
Hello
Hello