Last Week in AI - A Weekly Unwind

From 19-May-2024 to 24-May-2024

It was yet another thrilling week in the AI field with advancements that further extend the limits of what can be achieved with AI.

Here are 10 AI breakthroughs that you can’t afford to miss 🧵👇

OpenAI’s Transparency is in Question Again 🔍

Last week OpenAI’s Chief Scientist and Co-founder Ilya Sutskever resigned along with OpenAI’s Head of Superalignment team Jan Leike. Their departures have sparked discussions about OpenAI’s focus, with concerns that the company prioritizes product releases over AI safety.

OpenAI formed the Superalignment team in July 2023 to control superintelligent AI, promising it 20% of the company’s compute resources. However, they never allocated these resources. Several team members resigned following this disagreement. Departing employees faced restrictive off-boarding agreements, prompting criticism that led Sam Altman to change the policy. Despite these changes, OpenAI’s commitment to transparency remains questioned.

Hugging Face offers $10 Million in Free GPUs 🤗

Hugging Face is offering $10 million in free compute power to developers through its new ZeroGPU program, which provides shared GPU resources. This initiative aims to level the playing field by removing financial and technical barriers, enabling startups, researchers, and independent developers to compete with large tech companies. ZeroGPU, powered by Nvidia’s A100 GPUs, allows multiple users to share access to GPUs, making advanced AI development more affordable and accessible.

Microsoft’s Copilot + PCs with Always-Available AI Assistant 💻

Microsoft held a special event to unveil its new PCs designed for AI, the Copilot+ PCs. These are equipped with a powerful NPU - a specialized computer chip for AI-intensive processes - that can perform more than 40 trillion operations per second (TOPS). Microsoft Copilot, now powered by GPT-4o, is deeply integrated in these PCs. Key features are:

1. Recall captures snapshots of your screen every second and saves them to your hard drive, which can be later referred to to find the content you’ve seen before.

2. Cocreator AI image creation tool lets you combine text prompts with your own drawings to generate new images in real-time.

3. Restyle Image lets you transform your personal images into different styles like cyberpunk or claymation.

4. Live Caption translates audio into English in real-time.

5. Adobe’s flagship apps including Photoshop, Lightroom, and Express are also coming to these PCs.

OpenAI Pauses “Her” in ChatGPT’s Voice Mode 🎙️

Hollywood actress Scarlett Johannson has alleged OpenAI of cloning her voice for “Sky” in ChatGPT’s Voice Mode, without her permission. She accused OpenAI of using a voice in ChatGPT that closely mimicked her distinctive tones, despite declining an offer to voice the system. OpenAI claims the voice was from a different professional actress and not an imitation of Johansson. OpenAI has paused the “Sky” voice for now.

Source

Microsoft Build 2024 Event was All About AI 🌟

Microsoft held its Build 2024 event with a host of new announcements to enhance AI and developer tools across their Copilot and Azure AI ecosystem. Here are the top 5 announcements:

1. Teams Copilot extends Copilot to become a new team member. It can manage meeting agendas, take notes, show important information in chats, track action items, and more.

2. The new Copilot Studio now lets you build your own Copilot agents that can independently carry out tasks, automate business processes, have memory and knowledge, learn, and ask for help. You can also connect data from various sources to ground the Copilot’s responses.

3. Microsoft has released new Phi-3 models - Phi-3-vision is Microsoft’s first small vision language model with 4.2B parameters, Phi-3 small model with 7B parameters, and Phi-3 medium model with 14B parameters.

4. Real-Time Intelligence in Microsoft Fabric allows you to act on high-volume, time-sensitive data with tools like the Real-Time hub, event streams, and AI-powered insights for anomaly detection.

5. GitHub’s new Copilot extensions integrate tools and services from third parties within GitHub’s IDE, like Azure, Docker, and Sentry. Copilot Chat will now have context from all these tools to help you through your workflow without switching tools.

Adobe’s New AI Tool to Remove Unwanted Objects 🌁

Adobe has updated Lightroom with new AI-powered features for easier editing and creative control. The Generative Remove tool, powered by Adobe’s Firefly, allows users to remove unwanted objects from photos with a single click. The Lens Blur tool lets users apply various blur effects for a professional look, and expanded tethering support now accommodates more cameras for real-time viewing and seamless workflows.

Anthropic Researchers Peek Inside Black-box LLMs 👀

Researchers at Anthropic have made significant progress in understanding how LLMs work. By mapping millions of concepts within Claude-3 Sonnet using ‘dictionary learning,’ they uncovered complex features beyond simple words, including cities, people, and abstract ideas like gender bias. These features can be manipulated to change the model’s output. This research could significantly enhance AI safety by identifying and mitigating dangerous behaviors and steering LLMs toward safer outcomes.

AI Pin Startup Humane up for Sale 🏷️

Humane, the company behind the AI Pin that was supposed to be the next big thing, is reportedly looking for a buyer. The company founded by ex-Apple employees Imran Chaudhri and Bethany Bongiorno, has faced a lot of criticism after its underwhelming release of the AI Pin and criticism coming from left, right, and centre. Last year it was valued by investors at $850 million. 

ChatGPT Answers Will Come from Wall Street Journal 📰

OpenAI is partnering with News Corp., the company behind big names like Wall Street Journal, New York Post, Barron’s, and MarketWatch, to bring News Corp.’s content as answers to questions by OpenAI’s users. The aim is to provide users with more reliable information and news from trusted sources while using OpenAI’s products.

This comes at a time when misinformation spread by AI has become more prevalent, and Google is facing criticism in this area again. Google’s AI Overview uses top search results to summarize information and present it to user queries. It has been giving inaccurate, misleading, and garbage responses sourced from Reddit. Users are now simply asking Google to take down AI Overview from Search.

Source

Meta’s Mixed-Modal Early-Fusion AI Model 📖

Current multimodal AI models struggle to blend text and images seamlessly. Meta’s research team has developed Chameleon, a new AI model that can understand and generate both text and images together. Chameleon excels in tasks like image captioning, visual QA, and text-based reasoning, outperforming models like GPT-4V and Gemini. Unlike existing models that process text and images separately, Chameleon’s early fusion architecture allows it to integrate information from both sources from the start, enhancing its ability to create complex content.

Which of the above AI development you are most excited about and why?
Tell us in the comments below ⬇️

That’s all for today 👋

Stay tuned for another week of innovation and discovery as AI continues to evolve at a staggering pace. Don’t miss out on the developments – join us next week for more insights into the AI revolution!

Click on the subscribe button and be part of the future, today!

📣 Spread the Word: Think your friends and colleagues should be in the know? Click the ‘Share’ button and let them join this exciting adventure into the world of AI. Sharing knowledge is the first step towards innovation!

🔗 Stay Connected: Follow us for AI updates, sneak peeks, and more. Your journey into the future of AI starts here!

Shubham Saboo - Twitter | LinkedIn ⎸ Unwind AI - Twitter | LinkedIn