Google Deepmind's SIMA

PLUS: Microsoft Launches Copilot Pro Worldwide, Snowflake and Landing AI Partnership and more.

Today:

Google Deepmind's SIMA - the GOAT of AI Videogame Agents? [BIG progress towards 'human-like' play]

Google DeepMind introduces a versatile AI agent, SIMA, for 3D virtual environments, essentially video games. SIMA follows natural language instructions to perform tasks across various game settings, aiming to create a universal AI capable of real-world applications like robot piloting. By training on diverse game data, it learns to generalize skills and adapt to different environments. 

Unlike previous AI models, SIMA operates in real-time using keyboard and mouse inputs, resembling human gameplay. Although still imperfect, its progress signifies a significant breakthrough in AI capabilities with implications for remote work and automation. 

Apple Buys Canadian AI Startup as It Races to Add Features

Apple Inc. has quietly acquired Canadian AI startup DarwinAI, augmenting its capabilities ahead of a significant push into generative AI in 2024. The acquisition, made earlier this year, sees numerous DarwinAI employees joining Apple's AI division. 

This move signals Apple's commitment to advancing its AI technology, potentially leading to significant developments in the near future. The details of the acquisition were disclosed by sources familiar with the matter, who preferred anonymity as the deal has not been publicly announced.

Databricks invests in Mistral and brings its AI models to data intelligence platform

Databricks teams up with Mistral, investing in the Paris-based startup known for its high-performing language models. This partnership brings Mistral's models, including Mistral 7B and Mixtral 8x7B, to Databricks' Data Intelligence Platform. Users can access these models seamlessly for generative AI applications while maintaining data security and governance. 

The move expands Mistral's distribution network, following similar partnerships with Snowflake and Microsoft. While Databricks integrates select Mistral models, there's no mention of Mistral Large yet. Mistral continues to secure partnerships, aiming to drive AI use cases across various industries.

OpenAI Inks Deals With European Publishers Le Monde, Prisa

OpenAI has secured licensing agreements with two prominent European publishers, Le Monde from France and Prisa from Spain. These deals enable ChatGPT to access French and Spanish language news content, enhancing its training data for language models. 

The agreements reflect OpenAI's strategy to collaborate with media organizations rather than face disputes over content usage. While the terms of the deals were not disclosed, they mark another step in OpenAI's efforts to expand its partnerships with the media industry.

Invoke rolls out Workflows, AI tools for game developers

Invoke debuts Workflows, AI tools tailored for game developers to implement custom AI on a wide scale. Aimed at streamlining AI integration, it caters to non-tech team members with intuitive interfaces and robust security measures. Crafted for expansive studios, it empowers artists with precise control over image creation, aligning with their unique style and process. 

CEO Kent Keirsey emphasizes the adaptability of the tools, catering to users' varying skill levels with both complex and straightforward interfaces. Industry feedback applauds Invoke's focus on artists, enhancing creativity and workflow efficiency. Workflows signals a shift towards AI democratization within the gaming industry.

Microsoft launches Copilot Pro worldwide with a one-month free trial

Microsoft rolls out Copilot Pro globally, offering a free month trial. Subscribers gain access to advanced AI models and can integrate Copilot into Office web apps without a Microsoft 365 subscription. The expansion covers 222 countries and promises forthcoming mobile app integration. The service targets businesses, with availability expanding to more Microsoft 365 plans. 

Microsoft plans further enhancements, including Windows 11 updates, to be unveiled at the Build developer conference in May. To try Copilot Pro, users must download the app on iOS or Android. The move underscores Microsoft's push to integrate AI into everyday tasks, aiming to redefine productivity.

Snowflake and Landing AI combine forces to tackle unstructured data challenges with computer vision

Snowflake teams up with Landing AI, founded by Andrew Ng, to tackle unstructured data challenges using computer vision. Their partnership aims to integrate Landing AI’s advanced platform into Snowflake’s Data Cloud, enabling enterprises to harness visual data efficiently. With 90% of the world's data being unstructured, this collaboration targets various industries like manufacturing and healthcare. 

Snowflake's VP Stefan Williams highlights the ability to train accurate computer vision models with limited datasets, enhancing data security and governance. Despite Snowflake facing challenges like CEO transition, analysts see potential in the partnership's global expansion, initially focusing on North America and EMEA.

Apple supplier Foxconn turns upbeat on 2024, cites huge demand for AI servers

Foxconn, a key Apple supplier, anticipates a substantial revenue surge in 2024, driven by soaring demand for artificial intelligence (AI) servers. The optimistic outlook follows a notable 33% increase in fourth-quarter net profit. Foxconn's chairman cites robust client demand for AI servers, projecting over 40% revenue growth this year. The company foresees growth aligning with or surpassing the expected 30% annual increase in the AI server market until 2025. 

Despite expecting healthy growth, Foxconn notes lower first-quarter revenue compared to last year due to pandemic-related surges. The company plans to maintain robust capital expenditure growth, focusing on global diversification and expanding into sectors like electric vehicles and semiconductors.

OpenAI’s Sora: The devil is in the ‘details of the data’

OpenAI's Sora, showcased in a Wall Street Journal interview, left questions unanswered about its training data source. Despite Murati's vague responses, hints of Shutterstock and public data were dropped. The lack of transparency fuels copyright concerns and erodes trust. This isn't just OpenAI's problem; it's an industry-wide issue affecting trust and transparency. 

While tech giants like Google and Meta likely use similar data, public awareness remains low. The debate over AI training data's origins and ethics persists, potentially challenging its commercial viability. While the short-term benefits are clear, long-term consequences may haunt these companies.

🧠RESEARCH

Gemma is a new, lightweight set of models that pack a punch in language understanding, reasoning, and safety. Built on the tech behind Gemini models, Gemma comes in two sizes and outshines similar models in 11 out of 18 text-based challenges. It emphasizes safety and responsible model release, hoping to pave the way for future innovations in large language models (LLMs). 

This paper discusses efficient methods for updating large language models (LLMs) without starting from scratch every time new data is available. It suggests a combo of learning rate adjustments and data replay to avoid performance loss with new or previous data, proving effective for both minor and significant data shifts. These strategies allow LLMs to update with less computing power, proposing alternatives to prevent forgetting during updates. It's a step towards more sustainable and scalable LLM training processes. 

"VLOGGER" introduces a groundbreaking technique for generating videos of humans from just a single image, synchronized with audio. It utilizes advanced diffusion models for realistic movements and expressions across diverse scenarios without specific training for each individual. The method generates full-body images, not just faces, improving upon prior work. A new dataset, MENTOR, enhances training for diverse and dynamic gestures. VLOGGER sets new benchmarks in video quality, consistency, and inclusivity, showing potential in video editing and personalization. 

"SOTOPIA-π" dives into how humans pick up social skills through watching and interacting, areas where current language agent research is lacking. The paper introduces a new way to teach language agents to be more socially aware, using a mix of copying behaviors and self-reinforcement learning from selected social data. This approach not only boosts the agents' social skills to expert levels but also keeps their general knowledge sharp and safe for use. It points out the challenge of accurately measuring social intelligence in these agents. 

TRUMANS is a groundbreaking dataset for human-scene interaction (HSI) modeling, boasting over 15 hours of detailed motion-captured interactions in 100 indoor scenes. It uniquely captures whole-body human motions and dynamic object interactions with a focus on realistic contact. To scale, TRUMANS combines real physical and virtual environments, enhancing realism through motion and appearance augmentations. A novel diffusion-based autoregressive model leverages this dataset to generate lifelike HSI sequences, demonstrating exceptional zero-shot generalization across various 3D scene datasets. This approach achieves high fidelity in motion synthesis, confirmed by quantitative and human evaluations. 

🛠️TOP TOOLS

Smartly - AI-powered advertising platform that simplifies media buying, creative automation, reporting, and optimization

Kittl - AI-powered design tools

Reword - a platform that empowers users to train their own AI co-writer, enabling the creation of unique content that matches their writing style

Synthesia - Turn text to video in minutes

StickerBaker - Open Source AI Sticker Maker

Chesski - AI chess tutor  that coaches you in plain English

Saner - Instantly capture, find, and develop ideas without manual organizing

Octomind - find bugs before your users do

📲SOCIAL MEDIA

🗞️MORE NEWS

Oracle adds generative AI features to finance, supply chain software

Oracle announced the integration of generative AI into its business software, aiming to streamline tasks like report generation and data summarization. Competing with Microsoft's AI features, Oracle leverages AI for finance, supply chain, and HR functions. The company emphasizes human oversight to ensure accuracy and productivity enhancements without extra charges. REUTERS

Kolena debuts platform for testing AI models and fine-tuned variants

Kolena launches AI Quality Platform, aiding businesses in testing and validating AI models for safe deployment. The platform offers extensive testing capabilities, including data quality monitoring and debugging, crucial for avoiding mishaps and maintaining reputation. Developed by ex-Amazon and Rakuten engineers, it targets enterprise adoption and AI model providers, priced on a scalable SaaS model. VENTUREBEAT

Cerebras Unveils Its Next Waferscale AI Chip

Cerebras introduces WSE-3, its latest waferscale AI chip boasting 4 trillion transistors, doubling performance while maintaining power consumption. It's set to power an 8-exaflop supercomputer in Dallas. Cerebras also partners with Qualcomm to enhance AI inference performance. The CS-3 can train models with 24-trillion parameters, revolutionizing AI capabilities. IEEE SPECTRUM

CBP wants to use AI to scan for fentanyl at the border

Customs and Border Protection (CBP) aims to enhance border security by employing AI technology to detect fentanyl smuggling. The proposed system involves using high-resolution X-rays to scan vehicles at ports of entry, with plans to achieve 100 percent coverage. Despite having advanced fentanyl-detecting machines, CBP faces funding constraints hindering their installation. The majority of fentanyl seizures occur at official border crossings, emphasizing the significance of screening vehicles. CBP's goal is to scan 40 percent of passenger vehicles and 70 percent of commercial trucks by 2025, utilizing AI to expedite processing. THE VERGE

UK researchers seek to slash AI computing costs by a factor of 1,000

The UK's Advanced Research and Invention Agency (Aria) has launched a £42mn programme called Scaling Compute, aiming to reduce the cost of computing for artificial intelligence (AI) by over 1,000 times. The programme seeks to find cheaper and less energy-intensive alternatives to the current costly silicon-based digital infrastructure on which AI runs. Aria aims to leapfrog the limits of computing power and efficiency, paving the way for globally accessible and transformative AI. The agency will consider bids based on any type of hardware, including silicon, other semiconductors, and even biological materials. FINANCIAL TIMES

Smartschool's AI assistant tutor bags $1.5M Pre-Seed

Smartschool, an edtech company, secures $1.5M for their AI tutor, Yoko, aiding US students with SAT prep. Co-founders cite educational disparities, highlighting Yoko's personalized approach, akin to human tutoring but more affordable. Backed by Inovo VC and The Explorer Fund, funds target US expansion and product enhancement for equitable learning access. TECH.EU

Hackers can read private AI-assistant chats even though they’re encrypted

Hackers can decrypt supposedly encrypted AI-assistant chats, exposing sensitive info. Despite encryption efforts by providers like OpenAI, a side-channel attack reveals 55% accuracy in deciphering responses. This flaw affects all major AI assistants except Google Gemini, compromising privacy and security. The attack exploits real-time token transmission, revealing users' private conversations. ARS TECHNICA

AI Maps Brain Tissues to Disease Symptoms

A groundbreaking AI language model analyzes medical summaries and brain tissue data from the Netherlands Brain Bank, linking clinical symptoms to brain abnormalities. It aims to reduce misdiagnoses, potentially revolutionizing brain disease diagnosis and treatment by creating a molecular atlas of symptoms and identifying biomarkers for accurate diagnoses and targeted therapies. NEUROSCIENCE NEWS

As Generative AI Takes Off, Researchers Warn of Data Poisoning

Generative AI's rise brings hacking risks, with researchers warning of data poisoning. Malicious actors could manipulate training data to spread misinformation, compromise chatbot functionality, or steal sensitive information. As AI models ingest extensive data, vulnerabilities increase, highlighting the need for robust security measures. THE WALL STREET JOURNAL

Forget Chatbots. AI Agents Are the Future

Tech industry shifting from text-based chatbots to AI agents like Devin, capable of coding tasks beyond simple assistance. Startups and big players demo AI agents like Devin, a coder that plans, writes, and implements code. These advancements hint at a future where AI agents tackle complex tasks efficiently. WIRED

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.

Reply

or to participate.