🍎 Apple's new machine learning framework

Welcome, humans.☕

What to decode:

• 😉 How to animate anything with AI

• 🤯 DeepMind creates AI that can learns from humans

• 🍎 Apple’s new machine learning framework for Apple silicon

• 🤖 Google postpones launch of GPT-4 rival Gemini

• 📦 Amazon's Q suffering hallucinations, privacy issues

• ⛏️ New open-source AI tool

Read time: 4 minutes

🎬 How to animate anything with AI

A new AI model has recently gone viral, introducing an innovative technique for animating characters, images, and everything in between. To try a similar demo yourself, follow these steps:

• Visit the MagicAnimate demo on HuggingFace.

• Upload your chosen subject for animation. Note that most example motion sequences are full-body and vertical; consider this when generating or selecting your reference image.

• Click 'animate' and be patient! This demo may take longer than others, particularly due to the high volume of current users testing it. We're on the brink of an era where anyone can create and transform anything, constrained only by their imagination…

🧠 DeepMind creates AI that can learn from humans

Researchers from Google DeepMind have recently developed a new way for AI agents to acquire knowledge from human demonstrations in real-time. This breakthrough allows for "cultural transmission" without the need for large datasets.

Here are the details:

• The AI agents can learn from humans in a rich 3D simulation by observing movements and reproducing behaviors.

• The system combines deep reinforcement learning with memory, attention mechanisms, and automatic curriculum learning to achieve impressive performance.

• Tests have shown that the AI can generalize across tasks, recall demonstrations even when the expert is not present, and closely match human trajectories with goals.

This development is significant because cultural transmission in AI enables feedback loops that can greatly enhance learning over time. This method serves as a stepping stone towards systems that can accumulate knowledge over generations, similar to how humans do.

🍎A Machine Learning Framework Optimized for Apple Silicon

In an exciting development for the tech and AI communities, Apple has open-sourced MLX, a novel machine learning framework tailor-made for its Apple silicon chips. This initiative, spearheaded by Apple's machine learning research team, promises to revolutionize how researchers work on Mac, iPad, and iPhone.

Key Highlights of MLX:

• Intuitive APIs: MLX integrates Python and C++ APIs with familiar frameworks like NumPy and PyTorch, offering a seamless learning curve for experienced researchers.

• Optimized Performance: Leveraging composable function transformations, MLX is finely tuned to maximize Apple silicon's performance.

• Efficient Computation: The framework employs a 'lazy computation' strategy, materializing arrays only when necessary, thereby enhancing resource efficiency.

• Dynamic Flexibility: Adapting computation graphs to changes in input shape, MLX simplifies debugging and experimentation processes.

• Multi-Device Integration: Seamlessly utilizing the CPU and GPU of Apple devices, MLX ensures optimal hardware utilization.

• Shared Memory Utilization: Unique to MLX, it uses unified memory for storing arrays, cutting down on data movement between devices and speeding up operations.

• Researcher-Centric Design: With a clean and extensible codebase, MLX encourages contributions from the research community.

Apple has showcased MLX's prowess, particularly in enhancing natural language processing with efficient Transformer model training. Incorporating technologies like LLaMA and LoRA, MLX enables extensive text generation on Apple devices. Moreover, it supports stunning image generation with Stable Diffusion and offers accurate, efficient speech recognition using OpenAI's Whisper technology directly on Apple devices.

Given its focus on optimizing for Apple silicon and familiar APIs, MLX stands out as a potentially dominant framework for researchers exploring the frontiers of machine learning on Apple devices.

🤖 Google postpones launch of GPT-4 rival Gemini

Google recently postponed the release of its multimodal ChatGPT rival, Gemini, to January. This decision came after recognizing its shortcomings in processing non-English languages.

Here's what we know:

• At its I/O conference in May, Google showcased the upcoming model, initially set for a December release. Gemini, rumored to incorporate extensive training data, including YouTube transcripts, is said to match GPT-4's capabilities in certain aspects.

• However, difficulties with non-English language processing have been a significant factor in the delay.

• Google's CEO, Sundar Pichai, has suggested that Gemini marks the beginning of a series of advanced models scheduled for 2024.

The significance: Despite the high expectations placed on Gemini, a one-month postponement likely won't have a long-term impact on Google. Yet, the ongoing challenges faced by OpenAI's competitors highlight the difficulty of rivaling the market leader.

📦 Amazon's Q suffering hallucinations, privacy issues

Only a few days following its debut, internal documents are said to reveal warnings from employees about "severe hallucinations" and leaks of confidential data linked to Q, Amazon's latest enterprise AI assistant. While Amazon has refuted any confirmed problems, the company is looking into these concerns.

🛠️ COOL TOOLS

(Product launches, updates and demos)

• Zinbo - Use Local and GPT-4 LLMs to intelligently clean you Gmail inbox. (link) (Made by me!)

That's all for now!

As always, thanks for reading, and see you next time. 🫡