Today:

Claude 3 "Self-Portrait" Goes Viral | Beats GPT-4 Benchmarks | Why does it appears SELF-AWARE?

In the ongoing discussion about Claude 3's abilities, researchers debate its self-awareness while it competes closely with GPT 4. Claude 3 impresses with its performance, even generating a self-portrait and showing potential meta-awareness in test scenarios.

Amidst concerns of ideological bias, diverse testing methods become crucial. Claude 3's translation prowess in less-covered languages showcases its potential beyond scripted tasks.

WATCH THE VIDEO ON YOUTUBE

Hugging Face is launching an open source robotics project led by former Tesla scientist

Hugging Face announces a robotics venture led by former Tesla scientist Remi Cadene. This move, a departure from software, aims to integrate AI into low-cost, open-source robots. Cadene hints at recruitment in Paris, emphasizing the project's open-source ethos. 

The initiative reflects a broader trend in robotics investment and interest, exemplified by Tesla's Optimus and Figure's $675 million funding. With a focus on "embodied" AI, the industry seeks autonomous machines for various tasks. Cadene's tenure at Tesla underscores his expertise in this domain. Further details are awaited as Hugging Face ventures into hardware.

VENTUREBEAT

AI2 Incubator scores $200M in compute to feed needy AI startups

The AI2 Incubator, born from the Allen Institute for AI, snagged a hefty $200M to aid fledgling AI startups. Managed by Jacob Colker, it's a boon for AI enthusiasts facing compute shortages. The program offers up to $1M in dedicated compute from undisclosed data centers, a rare perk for early-stage ventures. 

This resource, unlike mere cloud credits, includes custom silicon and dedicated machines. AI2 has a track record of aiding pre-seed startups like WellSaid Labs and Xnor.ai. With its recent funding, it aims to propel more startups into the AI sphere with gusto.

TECHCRUNCH

Inflection AI launches new model for Pi chatbot, nearly matches GPT-4 

Inflection AI, founded by Mustafa Suleyman and Reid Hoffman, launched its new model, Inflection-2.5, which rivals OpenAI's GPT-4, particularly in STEM subjects. The model powers the Pi assistant, aiming to outperform ChatGPT and Gemini. It emphasizes empathy and safety, with a distinct personality and high EQ. 

Despite improvements, it still trails GPT-4 in benchmarks, though achieves 94% of its performance with more efficient training. Inflection-2.5 integrates real-time web search, enhancing user experience. Currently available for testing, it's driving user growth and engagement for Pi, with millions of active users and billions of messages exchanged.

VENTUREBEAT

Adobe adds Firefly generative AI features to new Express mobile app

Adobe has launched a new mobile app for Adobe Express, letting iOS and Android users access the same creative features as desktop users. The beta version is free and includes premium Firefly generative AI tools for editing images and text. Collaboration between desktop and mobile users is seamless, with historical data migrating to the new app upon its official release. 

While the AI features rely on cloud processing, not all smartphones support them yet. This move puts Adobe ahead of rivals like Canva, which lacks similar mobile AI integration. The app is available on Google Play, with iOS users needing to sign up separately.

THE VERGE

Salesforce announces new AI tools for doctors

Salesforce drops new AI tools to lighten doctors' admin loads. Einstein Copilot: Health Actions lets docs book appts, sum up patient info, and send referrals using conversational AI. Assessment Generation digitizes health assessments sans manual input. These tools aim to cut down on tedious paperwork, a major cause of physician burnout. 

Salesforce's platform consolidates medical data from various sources, easing data tracking for clinicians. Expected rollout: Assessment Generation this summer, Health Actions by year-end, both HIPAA-compliant. With over 90% of physicians reporting burnout, easing admin tasks could be a game-changer for healthcare.

CNBC

Google experiments with a tool to enable on-device AI

Google introduces the MediaPipe LLM Inference API, allowing developers to run AI models on devices with limited computing power. This innovation enables Large Language Models (LLMs) to operate fully on-device, despite their significant memory and compute requirements. Optimizations such as new ops, quantization, caching, and weight sharing facilitate this capability. 

MediaPipe currently supports four models - Gemma, Phi 2, Falcon, and Stable LM - across web, Android, and iOS platforms. Google aims to broaden support to additional models and platforms in the near future.

THE VERGE

🧠RESEARCH

GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection

"GaLore” is a smart way to train big computer brains (large language models) without needing a ton of memory. It tweaks training to use less memory—up to 65.5% less for some parts—without losing out on how smart or efficient the model is. They even got a big 7-billion-neuron model to learn on regular gaming graphics cards, which is pretty groundbreaking. 

ShortGPT: Layers in Large Language Models are More Redundant Than You Expect

"ShortGPT" finds that large language models (LLMs) have a lot of unnecessary layers. By examining how similar layers are, they introduce a new way to trim these models down without losing performance. They've come up with a simple yet effective approach to remove these extra layers, proving it can even do better than the latest complex methods. This makes LLMs leaner, meaner, and more efficient, opening doors to further optimize them for different tasks. 

SaulLM-7B: A pioneering Large Language Model for Law

SaulLM-7B is the first large language model (LLM) designed specifically for the legal field. It's built on the Mistral 7B architecture and trained on over 30 billion tokens from English legal documents. SaulLM-7B shines in understanding and generating legal texts, offering top-notch performance. The paper also details a new method to fine-tune this model on legal tasks, further boosting its abilities. SaulLM-7B is available under the CC-BY-SA-4.0 License. 

Enhancing Vision-Language Pre-training with Rich Supervisions

The paper presents a new way to teach vision-language models called Strongly Supervised pre-training with ScreenShots (S4). It uses web screenshots to grab both visual and text clues not found in usual image-text pairs. By tapping into the structured nature of HTML and spatial details, they create 10 unique training tasks that mirror real-world applications. This method significantly boosts image-to-text model performance across a variety of tasks, showing remarkable improvements, like a 76.1% jump in Table Detection. 

3D Diffusion Policy

"3D Diffusion Policy (DP3)" is a new visual imitation learning method that uses 3D visual representations to teach robots complex skills with fewer human demonstrations. By incorporating diffusion policies with a compact 3D visual representation, DP3 shows significant improvements in learning efficiency and generalization. In simulations and real robot tasks, DP3 outperforms baselines, achieving high success rates and demonstrating excellent safety and adaptability. The approach underlines the importance of 3D representations in robotic learning. 

🛠️TOP TOOLS

MarketMuse - AI to automate content strategy, offering tools for competitive analysis, content planning, and optimization

OctoEverywhere - AI Print Failure Detection Assistant

Castmagic - Turn Conversations Into Content 

Vercel -  introduces AI integrations, enabling developers to incorporate AI models and services into their projects easily. 

GeneGPT  - offers code and data aimed at enhancing genetic research.

Plus - AI writing assistant that enhances content creation within Google Docs. 

📲SOCIAL MEDIA

🗞️MORE NEWS

Microsoft confirms Surface and Windows AI event for March 21st

Microsoft is hosting a digital event on March 21st, focusing on new Surface Pro 10 and Surface Laptop 6 releases. These devices target business customers and feature AI improvements for Windows 11. The Surface Laptop 6 may sport a sleek design with thinner bezels, rounded corners, and multiple ports. Microsoft aims to ship models with Intel Core Ultra and Snapdragon X Elite processors. Additionally, Windows 11 will introduce an advanced AI Explorer feature, enhancing user experience. THE VERGE

Zapier Central debuts as no-code tool for building enterprise AI bots

Zapier launches Zapier Central, simplifying AI bot creation for its 2.2 million users. This "AI workspace" enables natural language commands to customize bots for various tasks, without coding. Unlike competitors, it offers more customization, signaling a new era in AI automation. Additionally, Zapier acquires Vowel's team to enhance its AI capabilities, aiming to democratize AI and automation for businesses. VENTUREBEAT

VC firm Andreessen Horowitz plans to raise billions for AI, with portions devoted to two funds

Andreessen Horowitz plans to raise $6.9 billion, with a focus on AI, gaming, and American Dynamism sectors. Half will fund growth, while the rest will be divided between two AI funds and others. The move reflects the firm's strategy shift towards emerging tech sectors. Critics, however, speculate over potential market impacts. FORTUNE

India announces $1.2 bln investment in AI projects

India announces a $1.25 billion investment in AI projects, focusing on computing infrastructure, language models, and funding startups. The move aims to capitalize on India's growing AI market, projected to reach $17 billion by 2027. This initiative underscores India's commitment to advancing AI technology and fostering innovation in the sector. REUTERS

SK Hynix Invests $1 Billion in Key AI Memory Chip Technology

SK Hynix invests over $1 billion in advanced chip packaging for high-bandwidth memory crucial for AI development. The investment aims to enhance chip manufacturing processes, reduce power consumption, and drive performance. SK Hynix's focus on advanced packaging reflects the industry's shift towards back-end innovations to meet AI demands. YAHOO!

AWS launches Generative AI Competency to grade AI offerings

Amazon Web Services (AWS) launches the Generative AI Competency program to validate AI models, software, and hardware for businesses. With over 40 partners, including Accenture and NVIDIA, the program aims to help customers identify and adopt the best AI solutions, accelerating generative AI strategy and growth. ZDNET

Researchers Develop New Technique to Wipe Dangerous Knowledge From AI Systems

Researchers developed a method to detect and remove dangerous knowledge from AI systems, aiming to prevent misuse in cyberattacks and bioweapon deployment. Through a questionnaire designed by experts, they identified hazardous knowledge areas and applied a "mind wipe" technique, showing promise in enhancing AI safety. This advancement responds to growing concerns about AI misuse and aligns with governmental efforts to mitigate risks. TIME

Healthy Together Acquires Kinsa Health To Build AI Illness Forecasting & Expand Into New Markets

Healthy Together, a health tech firm, acquires Kinsa Health, integrating its AI illness forecasting into their platform. The move aims to enhance public health management by predicting illness outbreaks. Kinsa's technology, powered by millions of smart thermometers, aids in early detection and proactive measures. The acquisition strengthens Healthy Together's commitment to positive health outcomes. PR NEWSWIRE

Scientists predict the extinction risk for all the world's plants with AI

Scientists at Royal Botanic Gardens, Kew, using AI, predict extinction risk for all known flowering plants. Their research aids both casual plant lovers and biodiversity experts, empowering them to protect threatened species. By freely sharing their data, they aim to accelerate conservation efforts and raise awareness about the importance of preserving plant life. PHYS ORG