The Daily Newsletter for Intellectually Curious Readers

• We scour 100+ sources daily

• Read by CEOs, scientists, business owners and more

• 3.5 million subscribers

Sign up today!

Today:

• OpenAI Launches Advanced ChatGPT Search

• Recraft V3 Revealed as red_panda

• Gemini AI integrates Google Search

• Anthropic brings Claude to computers

• Meta Enhances Robots’ Touch Capabilities

OpenAI Launches Advanced ChatGPT Search

OpenAI has launched ChatGPT Search, a new feature that allows ChatGPT to search the internet and provide quick, accurate answers. Instead of using separate search engines, users can ask questions naturally, and ChatGPT will find relevant information and link to reliable websites. This includes up-to-date details like weather, news, sports scores, and stock prices. 

Currently available to paying users, it will soon be accessible to more users. The feature aims to make finding information easier and faster by combining conversation with live web data. Partnerships with major news and data providers ensure trustworthy sources are used.

Recraft V3 Revealed as red_panda

Recraft has unveiled Recraft V3, its latest AI model that sets a new standard in image generation quality, outperforming competitors according to industry benchmarks. The model excels in generating images with accurate text placement, correct anatomy, and high aesthetic value. New features allow users to specify the exact size and position of text in images, combine text with other images to create complex designs, and finely control styles for brand consistency. 

Recraft V3 supports both raster and vector images and offers AI editing tools like erasing, inpainting, and background removal. Available on desktop, mobile, and via API, it empowers designers with greater control over their creations.

Gemini AI integrates Google Search

Google has introduced "Grounding with Google Search" for its Gemini AI platform, enabling AI models to access current information from Google Search. Aimed at developers creating AI applications, this feature allows them to enhance their apps with up-to-date search data, complete with citations and sources, at a cost of $35 per 1,000 queries. 

Launched just hours before OpenAI's consumer-focused ChatGPT Search, Google's move intensifies competition in AI-powered search services. This development could change how people find information online, shifting from browsing through links to relying on AI systems that synthesize answers from multiple sources.

Anthropic brings Claude to computers

Anthropic has released desktop apps for its AI chatbot, Claude, now available for Mac and Windows users in a public beta test. This allows people to use Claude directly on their computers. Additionally, they've introduced a dictation feature for mobile devices like iPhones and Android phones. 

Users can record voice messages up to 10 minutes long, which Claude will transcribe (convert speech to text) and respond to—though it's not a real-time conversation tool. These updates aim to make Claude more competitive with rivals like OpenAI's ChatGPT, as Anthropic reportedly seeks significant investment to enhance its AI offerings.

Meta Enhances Robots’ Touch Capabilities

Meta's AI research team is making robots better at understanding and interacting with the physical world. They introduced Meta Sparsh, a system that helps robots feel touch across different sensors and tasks. They also unveiled Meta Digit 360, an artificial fingertip that senses touch with human-like precision, capturing detailed information when it touches something. 

Another innovation is Meta Digit Plexus, which allows various touch sensors to work together in a single robot hand. They're partnering with companies to make these technologies available to others. Additionally, they launched PARTNR, a tool to improve collaboration between humans and robots.

🧠RESEARCH

CORAL: Benchmarking Multi-turn Conversational Retrieval-Augmentation Generation

CORAL, a benchmark for evaluating Retrieval-Augmented Generation (RAG) in multi-turn conversational settings, aiming to improve language models with real-world, multi-step conversations. CORAL includes Wikipedia-based conversations, addressing challenges like open-domain coverage, complex topics, and response generation, offering insights for refining RAG methods in conversational AI.

ReferEverything: Towards Segmenting Everything We Can Speak of in Videos

REM, a video segmentation framework that identifies objects and actions described in natural language, using video diffusion models trained on large datasets. By preserving model representations and fine-tuning on specific segmentation tasks, REM accurately segments rare items and dynamic scenes, outperforming existing methods in unfamiliar domains.

AutoMIR: Effective Zero-Shot Medical Information Retrieval without Relevance Labels

 AutoMIR, a method for zero-shot medical information retrieval without labeled relevance data. Using Self-Learning Hypothetical Document Embeddings (SL-HyDE), it generates context-rich documents to improve retrieval accuracy. The Chinese Medical Information Retrieval Benchmark (CMIRB) further evaluates AutoMIR's effectiveness across diverse medical tasks, surpassing traditional methods in accuracy and adaptability.

SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation

SlowFast-VGen, a video generation model that mimics human learning by blending slow and fast learning processes. Using a novel dual-speed approach, it addresses inconsistencies in long video generation by storing short-term memory for local updates while learning broader dynamics. With better scene consistency and long-term planning, SlowFast-VGen surpasses traditional models on action-driven video tasks.

A Large Recurrent Action Model: xLSTM enables Fast Inference for Robotics Tasks

LRAM, a large action model for robotics that uses the xLSTM recurrent architecture, providing faster inference than Transformer models. Designed for real-time applications, LRAM excels in handling extended sequences efficiently, outperforming Transformers across multiple robotics tasks with quicker response times and high performance.

🛠️TOP TOOLS

Recraft AI - Premium image generation and editing tool

Sourcely - Find Academic Sources in Seconds

Averi - Craft marketing strategy, create endless content, and build a perfect-fit team

Readkidz - makes children's multimedia content creation dramatically easier and faster

Lasso AI - Find the revenue you’re missing with AI

📲SOCIAL MEDIA

🗞️MORE NEWS

• Meta released MobileLLM, a set of efficient AI models optimized for mobile devices. Available under a research-only license, MobileLLM’s compact design tackles on-device memory constraints, making advanced AI accessible to researchers on platforms like Hugging Face.

• Google is adding generative AI tools to Google Maps Platform and Google Earth, enhancing developers' and urban planners' access to real-time place data and city insights. These updates will support applications for up-to-date location information and quick, in-depth urban planning analysis.

• OpenAI has expanded its Realtime API, adding five new voices for speech-to-speech applications and significantly reducing costs with prompt caching. This update aims to enhance real-time voice assistant performance while making API access more affordable for developer

• Toyota and NTT will invest $3.3 billion by 2030 to develop an AI platform aimed at reducing traffic accidents. This mobility AI initiative, targeting urban areas and automated driving, aims for industry-wide adoption by 2030.

• Microsoft’s OmniParser, an open-source AI tool, translates screenshots into data usable by AI agents, making it the top trending model on Hugging Face. OmniParser uses OCR and object detection to help AI interact with various screen elements, enhancing automated interface navigation.

• Startup Ataraxis has launched Ataraxis Breast, an AI-powered test that predicts breast cancer severity up to 30% more accurately than current methods, potentially sparing thousands from unnecessary treatments. The model was trained using images from over 4,500 patients.