Good morning! This is your daily ☕️ Techpresso.
On this day in 1976, the Apple 1 was released.
In today's Techpresso:
💥 Apple hit with ‘mercenary spyware attacks’
🧠 Humane AI has only one problem: it just doesn’t work
🔍 MistralAI unveils groundbreaking open model Mixtral 8x22B
🙃 Microsoft proposed using DALL-E to US military last year
🎵 New AI music generator Udio synthesizes realistic music on demand
🎬 Adobe is purchasing video content to train its AI model
🎁 + 7 other news you might like
🔮 + 8 handpicked research papers and tools
💥 Apple hit with ‘mercenary spyware attacks’LINK
Apple has issued a warning to iPhone users in 92 countries about a potential "mercenary spyware attack" that aims to compromise their devices.
The spyware, associated with state actors and created by private companies like the NSO Group, can extract data and monitor devices remotely.
This widespread alert follows Apple's history of issuing threat notifications, highlighting the continuing threat of Pegasus spyware infections globally.
🧠 Humane AI has only one problem: it just doesn’t workLINK
The Humane AI Pin, retailing for $699 plus a $24 monthly fee, is designed as a wearable alternative to smartphones, promising users freedom from their screens through AI-assisted tasks. However, its functionality falls significantly short of expectations. Throughout testing, the AI Pin struggled with basic requests and operations, demonstrating unreliability and slow processing times, leading to the conclusion that it fails to deliver on its core promise of a seamless, smartphone-free experience. Despite its well-intentioned vision for a post-smartphone future and the integration of innovative features like a screenless interface and ambient computing, the device's current state of performance and high cost make it a poor investment for consumers.
🔍 MistralAI unveils groundbreaking open model Mixtral 8x22BLINK
Paris-based AI startup Mistral has unveiled the Mixtral-8x22B MoE, a new open language model, available via torrent.
The Mixtral-8x22B model features a 64,000 token context window and demands 258 gigabytes of VRAM for operation.
Similar to its predecessor, the Mixtral-8x7B, the new model utilizes a mixture-of-experts approach to machine learning.
🙃 Microsoft proposed using DALL-E to US military last yearLINK
Microsoft proposed to the U.S. Department of Defense in 2023 to use OpenAI's DALL-E AI for software development in military operations.
The proposal included using OpenAI tools like ChatGPT and DALL-E for document analysis, machine maintenance, and potentially training battlefield management systems with synthetic data.
Microsoft had not implemented the use of DALL-E in military projects, and OpenAI, which did not participate in Microsoft's presentation, restricts its technology from being used to develop weapons or harm humans.
🎵 New AI music generator Udio synthesizes realistic music on demandLINK
Uncharted Labs has officially launched its music generator, Udio, which can transform text prompts into professional-quality music tracks, challenging the leading AI music generator, Suno V3.
Udio has impressed users and reviewers alike with its ability to generate songs that feature coherent lyrics, well-structured compositions, and competitive rhythms, some even considering it superior to Suno V3.
Despite facing initial server overload due to high user demand, Udio's user-friendly interface and strong backing from notable investors suggest a promising future for AI-assisted music creation, though it remains free during its beta testing phase.
🎬 Adobe is purchasing video content to train its AI modelLINK
Adobe is developing a text-to-video AI model, offering artists around $3 per minute for video footage to train the new tool, as reported by Bloomberg.
The software company has requested over 100 video clips from artists, aiming for content that showcases various emotions and activities, but has set a low budget for acquisitions.
Despite the potential for AI to impact artists' future job opportunities and the lack of credit or royalties for the contributed footage, Adobe is pushing forward with the AI model development.
Other news you might like
Amazon adds AI heavyweight Andrew Ng to board of directors as Jassy says it may be biggest tech breakthrough ‘perhaps since the internet’.LINK
Meta unveils second-gen AI training and inference chip.LINK
Adobe executive says its new AI tech is aimed at making people more creative – not produce horror stories.LINK
Google Photos is making its AI photo editing tools available for all users.LINK
Google global affairs chief says global alignment on AI ‘going surprisingly well’.LINK
Apple will let you repair your iPhone using genuine used parts, starting with iPhone 15.LINK
Hopes rise for mRNA cancer vaccine after Moderna trial shows promise.LINK
Latest research and tools
Aider: a command line tool that enables AI-driven pair programming by allowing GPT-3.5/GPT-4 to directly edit and commit code in your local git repository, supporting a variety of programming languages and operations on both new and existing projects.LINK
RecurrentGemma: a family of open-weights Language Models by Google DeepMind that leverages the Griffin architecture for efficient long sequence generation, with Flax and PyTorch implementations available.LINK
Transformers.js: enables state-of-the-art machine learning in the browser by running 🤗 Transformers directly without a server, using ONNX Runtime for model execution.LINK
Lens sort: a tool that utilizes facial recognition and pixel sorting to apply glitch effect on images, creating Snapchat-like lenses.LINK
Next-token prediction in JavaScript: enables the creation of language models for high-quality predictions and can serve as an alternative to decoder-only models with sufficient training.LINK
Full forward pass of GPT-2 in one file of pure CUDA: a tool that simplifies the execution of a full forward pass of the GPT-2 model using CUDA for improved efficiency.
Saved searches: a feature that allows users to save specific search criteria for quicker access to relevant results in the future.
Feedback system: a platform mechanism that collects users' opinions and suggestions to enhance the service or product quality.LINK
World's Smallest CSV Parser (C#): a compact, 100-line CSV (RFC4180) parser that eschews complex abstractions, allowing for straightforward customization like DictReader creation or Excel column group support.LINK
Toward Inference-Optimal Mixture-of-Expert Large Language Models: the paper explores designing large language models that efficiently combine expert knowledge for better performance.LINK
Want to get the latest news differently? Find us on:
See you tomorrow for a new dose of ☕️ Techpresso!
-min.png){kind=small}
.png){kind=small}
