2024 is the Year of the AI AGENT

PLUS: Roblox introduces real-time chat translator, AI Discovers Copper Deposit and more…


2024 is the Year of the AI AGENT

2024 is shaping up to be the breakthrough year for AI agents, which will be capable of performing a wide array of tasks autonomously, from web surfing to playing games and even carrying out physical tasks.

This movement is underpinned by advances in multimodal models and the use of neuro-symbolic algorithms for more intuitive, human-like interaction with software. As AI agents become more sophisticated, they promise to revolutionize how we interact with the digital world, raising questions about the future of work, the economy, and our online lives.

Microsoft partners with Semafor for AI-assisted news content

Microsoft is teaming up with media startup Semafor to leverage its AI chatbot for crafting news stories, a move happening amidst a massive lawsuit from the New York Times over AI use. Microsoft's funding a project called "Signals" at Semafor, aiming to provide diverse global perspectives in news with a hefty daily dose of breaking stories. 

Despite the financial details being under wraps, it's hinted that Microsoft's contribution is significant for Semafor. This collaboration highlights the growing intrigue and concern over AI in news production, balancing innovation with the fear of losing audience to AI-generated content. 

Semafor insists on journalist-driven content, using AI as a tool for enhancing research and breadth of coverage. Amidst industry challenges, this partnership underscores a forward-looking approach to journalism and technology synergy.

Apple’s new AI research promises high performance without the high price tag

Apple has unveiled a groundbreaking AI research that promises high performance at a lower cost. Their paper, titled "Specialized Language Models with Cheap Inference from Limited Domain Data," offers an economical approach to AI development. They identify four key cost areas: pre-training, specialization, inference, and training set size. 

By managing these expenses efficiently, businesses can create affordable and effective AI models. The research introduces two pathways: hyper-networks and mixtures of experts for generous budgets, and smaller selectively trained models for tighter budgets. Practical guidelines for choosing the best method for a specific domain and budget are provided, making language models more accessible.

Roblox releases real-time AI chat translator

Roblox has introduced a real-time AI chat translator that swiftly translates text chats in 16 languages, including English, French, Japanese, Thai, Polish, and Vietnamese. The goal is to enhance user engagement by eliminating language barriers, making it easier for players to communicate. Users can see the original message by clicking an icon. 

Roblox created this feature using a transformer-based large language model (LLM) within a mixture of experts (MoE) architecture, saving resources. The company monitors chats for safety and privacy, ensuring banned words are still blocked. Roblox plans to expand the translator's capabilities to include voice translation in the future. Other companies like Meta and Google are also developing AI translator models.

AI-Powered Miner Says It Found Huge Copper Deposit in Zambia

KoBold Metals, an AI-driven mining startup, has unearthed a massive copper deposit in Zambia. This San Francisco-based company, supported by Bill Gates, Jeff Bezos, and others, has been drilling in Zambia for over a year. 

The discovery, known as Mingomba, is hailed as remarkable by KoBold President Josh Goldman. This development highlights how AI is transforming the quest for materials crucial to the green-energy transition.


Specialized Language Models with Cheap Inference from Limited Domain Data by Apple tackles the challenge of applying large language models (LLMs) to tasks with limited inference budgets and specific domain data. The authors identify key variables affecting model application: pretraining budget, specialization budget, inference budget, and in-domain training set size. They explore alternatives to large, costly transformer models, finding hyper-networks and mixture of experts perform better with ample pretraining budgets, while smaller models trained on carefully selected datasets excel when specialization budgets are larger. This approach offers a more efficient path for specialized LLM application.

StepCoder is a fresh take on making computers write code better, using a smart twist with reinforcement learning—kinda like learning from trial and error, but with code. It breaks down the big task into smaller, manageable bits and fine-tunes using detailed feedback, making it outshine older methods. Plus, they whipped up a new, reliable dataset to train it. 

TravelPlanner introduces a benchmark for evaluating AI in real-world planning, like organizing trips. It's a challenge for language models, including GPT-4, which only succeeded 0.6% of the time. This benchmark, packed with nearly four million records and over 1,200 planning scenarios, reveals AI's current struggle with complex tasks, signaling both progress and the need for improvement in AI planning capabilities. 

PokéLLMon introduces an AI that hits human-level skills in Pokémon battles, mixing in-context learning, knowledge augmentation, and consistent action strategies to dodge common pitfalls and make smart moves. It's impressing folks with a win rate of 49% in competitive ladders and 56% in invited matches. They've shared their code and battle logs for anyone interested in how they're shaking up game AI. 

Boximator introduces a novel method for creating detailed and adjustable motions in video synthesis. It allows users to precisely control an object's position, shape, or movement in videos by using "hard" or "soft" boxes. This tool enhances existing video models without altering their original structure, thanks to a unique self-tracking technique that simplifies the association between boxes and objects. Boximator not only boosts video quality significantly but also provides robust motion control, as shown by improved metrics and positive user feedback.

This paper highlights a clear advantage of transformer models over generalized state space models (GSSMs) in tasks requiring the ability to copy from input context. Through theoretical analysis and empirical tests, the authors demonstrate that transformers can handle exponentially long strings due to their architecture, unlike GSSMs limited by fixed-size latent states. This fundamental difference showcases transformers' superior efficiency and generalization in context copying tasks, underlining a significant gap in performance between the two types of models for practical applications. 


Brainner - AI-Driven Resume Screening

Keep It Shot - utilizes AI to automatically provide descriptive names for your screenshots.

Findr - AI-powered search platform that allows users to search across all their apps at once, streamlining workflow and enhancing productivity with features like real-time search, AI search assistant, and granular access controls for security.

Taped - AI tool that helps organize thoughts and notes. It converts spoken words into structured, easily navigable notes, simplifying the process of managing ideas and discussions for better productivity and clarity.

Meco - a newsletter aggregator designed to free your inbox from clutter and make reading newsletters a more enjoyable experience. 

Checkget AI - a versatile smart assistant accessible through a simple keyboard shortcut, helping users across websites. It adapts to various tasks, such as email responses and code explanations.



Bumble’s new AI tool identifies and blocks scam accounts, fake profiles

Bumble's rolling out a new AI tool, the Deception Detector, nipping scams and fake profiles in the bud by automatically blocking 95% of them. This move has slashed spam complaints by almost half in just two months. It's part of Bumble's push to make online dating safer, especially for women, who've been voicing concerns over match authenticity. Amidst soaring romance scam losses, hitting $1.3 billion last year, this tech aims to bolster trust in the digital love hunt. TECHCRUNCH

Huawei just retasked a factory to prioritize AI over its bestselling phone

Huawei shifts gears, reprioritizing a factory from churning out its hot-selling Mate 60 phones to pumping out more of its AI chip, the Ascend 910B. This move, sparked by booming demand for AI capabilities domestically, sidelines the Mate 60's production despite its edge over Apple in China's market last year. With the U.S. squeezing chip exports, Chinese AI players are turning to homegrown solutions like Huawei's Ascend to keep their AI ambitions alive amidst a global race where they're playing catch-up. THE VERGE 

AI helps scholars read scroll buried when Vesuvius erupted in AD79

AI breakthrough enables scholars to decode text from a nearly 2,000-year-old charred scroll buried by Mount Vesuvius in AD79, marking a pivotal moment in ancient studies. This scroll, part of a collection from Herculaneum, was previously unreadable due to damage. The $1m Vesuvius Challenge, initiated by Brent Seales and tech sponsors, awarded a team for revealing over 2,000 Greek letters from the scroll using advanced CT scans and AI. This technology, capable of digitally unwrapping and detecting ink on the scrolls, may revolutionize our understanding of Roman era texts, offering insights into ancient Greek philosophy and potentially leading to the discovery of more scrolls at the site. THE GUARDIAN

Human resources gets AI upgrade with ADP Assist

ADP Assist, a new AI-powered chatbot from HR and employee benefits software provider ADP, aims to revolutionize human resources management by offering support to HR managers and employees alike. It's designed to handle common inquiries and identify potential issues, easing the workload of HR departments by automating responses to repetitive questions. Features include a payroll inspector for accuracy and smart analytics for strategic HR focus. ADP leverages its vast human capital management dataset to train the AI, ensuring it meets the specific needs of users. The tool, currently in early release to select ADP clients, will be widely available without additional charges, emphasizing ADP's commitment to integrating AI responsibly while safeguarding data and ethical standards. VENTUREBEAT

Inside the Underground Site Where ‘Neural Networks’ Churn Out Fake IDs

OnlyFake, a new underground site, is shaking up cybersecurity by using AI to crank out realistic fake IDs for a mere $15, setting the stage for easier bank fraud and money laundering. 404 Media's tests show these IDs, looking legit enough to fool some systems, can be whipped up on the fly with custom info. The site boasts a high output, claiming to use "neural networks" and "generators" for mass production. This move could spell the end for the old-school, labor-intensive fake ID game. 404 MEDIA

Scientists use AI to investigate structure and long-term behavior of galaxies

Scientists at Bayreuth University are leveraging artificial intelligence, specifically a deep neural network, to explore the structure and longevity of galaxies. Their research, grounded in Einstein's general theory of relativity, aims to predict the stability of galaxy models, which is critical given the black holes at most galaxy centers. This AI method allows for the quick verification or falsification of astrophysical hypotheses, making it a significant advancement in the field. The deep neural network, described as a breakthrough, can process vast amounts of data to determine the feasibility of galaxy models in reality, offering predictions much faster than traditional numerical simulations. This innovative approach, published in the journal Classical and Quantum Gravity, represents a new era in understanding galaxies and could potentially accelerate further discoveries in astrophysics. PHYS ORG

The UK has decided to shelve the development of a crucial code that would set rules for the use of copyrighted material in training artificial intelligence (AI) models, marking a significant setback for the creative industries. This decision emerged after unsuccessful attempts to reach a consensus among AI companies and rights holders on a voluntary code of practice, under the guidance of the Intellectual Property Office. The initiative aimed to address concerns in the creative sector about their work being used without fair compensation to train AI models, highlighting a significant conflict between the need for access to data for AI development and copyright protection. The forthcoming government white paper on AI regulation is expected to acknowledge the need for industry agreement on this issue but will not present a concrete policy, underscoring the challenges in balancing innovation in AI with copyright protection. FINANCIAL TIMES

IBM tops list of AI-related patent applications

IBM leads in AI-related U.S. patent applications over the past five years, outpacing giants like Microsoft, Google, and OpenAI, highlighting its significant investment in AI research. With 1,591 AI-related patent applications, IBM's dominance suggests a strong focus on innovation beyond its traditional business areas. The surge in generative AI patents, growing at an annual rate of 31%, underscores the tech industry's rapid advancement in this field. Despite the prolific filing, the practical application of these patents remains to be seen, as not all will translate into marketable products or features. This patent activity reflects the broader industry trend towards AI development, with companies seeking to protect their innovations while navigating the complexities of intellectual property in the era of AI-generated content. AXIOS

Nvidia CEO Says Nations Seeking Own AI Systems Will Raise Demand

Nvidia's CEO, Jensen Huang, anticipates a surge in demand for Nvidia products as nations across the globe, including India, Japan, France, and Canada, aim to develop and maintain their own artificial intelligence (AI) infrastructure domestically. Huang highlights the growing recognition of the importance of "sovereign AI capabilities," with countries viewing their data as a natural resource that should be processed and utilized for national benefits. This global shift towards building in-house AI systems underscores the strategic value placed on AI technology and data sovereignty, potentially leading to increased investments in Nvidia's AI-focused hardware and software solutions. BLOOMBERG

AI Revealed as Surprising Secret to 4-Day Workweek Adoption

The integration of Artificial Intelligence (AI) in businesses is showing promising signs of facilitating a more flexible work environment, particularly the adoption of a 4-day workweek. A survey of 1,000 US-based business leaders revealed a significant correlation between the use of AI technologies and the willingness to consider shorter workweeks. Specifically, companies with extensive AI experience are more than twice as likely to be open to a 4-day workweek compared to those that have not utilized AI. This trend suggests that AI is not just transforming operational efficiencies but also reshaping traditional work paradigms. By automating tasks and streamlining processes, AI is creating opportunities for companies to experiment with and potentially adopt more worker-friendly policies, like reduced workweeks, without sacrificing productivity. TECH.CO


What'd you think of today's edition?

Login or Subscribe to participate in polls.

What are MOST interested in learning about AI?

What stories or resources will be most interesting for you to hear about?

Login or Subscribe to participate in polls.


or to participate.