World’s first 16 Hour LIVE Training to become an AI-Powered human in 2025

The world of AI is evolving at lightning speed, and the only way to stay relevant is to MASTER AI before it masters you.

Join the World’s first 2-Day Mastermind Challenge to learn the Tools, Tactics, and Strategies to Automate Your Work Like Never Before!

Best part? It is usually for $395, but the first 100 of you get in for FREE!

Welcome to AI_Distilled. Today, we’ll talk about:

Techwave

Copilot for all: Introducing Microsoft 365 Copilot Chat

Scheduled tasks in ChatGPT

Andrew Ng announces AI-powered Climate Simulator

GitHub Next | Copilot Workspace

Codestral 25.01 | Mistral AI | Frontier AI in your hands

Awesome AI:

GitPodcast

Scrape anything with AI - FetchFox

AISmartCube - Low Code AI Tools

STORM

Whisk by Google Labs

Masterclass

Titans: Learning to Memorize at Test Time

Agents

HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs

AutoGen v0.4: Reimagining the foundation of agentic AI for scale, extensibility, and robustness

Agent Laboratory: Using LLM Agents as Research Assistants

Hackhub

facebookresearch/coconut: Training Large Language Model to Reason in a Continuous Latent Space

Efficient-Large-Model/Sana

vikhyatk/moondream2

hexgrad/Kokoro-82M

Sky-T1: Train your own O1 preview model within $450

Cheers,

Shreyans Singh

Editor-in-Chief, Packt

Cloud Conversations: A Fireside Chat with Forrest Brazeal and Rubrik

Join us on Jan. 28th @ 10 AM PST for a captivating fireside chat where storytelling meets cloud innovation. Forrest Brazeal—acclaimed cloud architect, author, and the creative mind behind cloud computing's most beloved cartoons—teams up with Rubrik’s Chief Business Officer, Mike Tornincasa to explore the evolving challenges of data protection in a multi-cloud world.

Copilot for all: Introducing Microsoft 365 Copilot Chat

Microsoft has launched Microsoft 365 Copilot Chat, a new AI-powered tool for businesses, combining GPT-4o chat capabilities with agents to automate tasks and enhance productivity. Available in free and pay-as-you-go versions, it allows users to perform tasks like summarizing documents, analyzing data, and generating content while enabling businesses to create custom agents for workflows like customer service and field operations.

Scheduled tasks in ChatGPT

OpenAI has introduced Scheduled Tasks in ChatGPT, now available in beta for Plus, Pro, and Team users on Web, iOS, Android, and macOS (Windows support coming later). This feature lets users automate tasks by scheduling prompts for specific times or intervals. Tasks run independently of user activity, with notifications sent upon completion. Examples include daily reminders, news briefings, or language practice. Users can manage, edit, or delete tasks through a dedicated "Tasks" menu and customize notification preferences. Limited to 10 active tasks, this beta feature supports GPT-4o capabilities while expanding automation and proactive engagement in ChatGPT workflows.

Andrew Ng announces AI-powered Climate Simulator

Andrew Ng recently announced the release of an AI-powered Climate Simulator to explore how geoengineering, specifically Stratospheric Aerosol Injection (SAI), could help mitigate global warming. SAI involves injecting aerosols into the stratosphere to reflect a small portion of sunlight, potentially cooling the planet and opening pathways to limit global warming to 1.5°C. The simulator allows users, including policymakers and the public, to experiment with SAI deployment scenarios and compare their effects against continued warming.

GitHub Next | Copilot Workspace

Copilot Workspace is a developer environment powered by AI, designed to simplify everyday coding tasks. It allows users to describe their goals in natural language, with AI agents proposing and implementing plans, troubleshooting errors, and brainstorming ideas. Features like an integrated terminal, repair suggestions, and easy collaboration make development seamless, while secure versioning and one-click PR creation streamline workflows.

Codestral 25.01 | Mistral AI | Frontier AI in your hands

Codestral 25.01 is a cutting-edge coding model from Mistral AI, designed to make software development faster and more efficient. Optimized for tasks like code completion, correction, and test generation, it supports over 80 programming languages and excels in fill-in-the-middle (FIM) scenarios. The latest update offers twice the speed of its predecessor, a more efficient architecture, and better tokenizer performance, making it a leader among lightweight coding models.

GitPodcast

GitPodcast is a tool that transforms GitHub repositories into quick, engaging podcasts, making it easier to understand projects on the go. Simply replace "hub" with "podcast" in a GitHub URL to generate a podcast summarizing the repository. It offers short (~5-minute) podcasts for quick insights and longer (~10-minute) versions with a sign-in. This is especially useful for developers and teams who want a convenient way to grasp project details without diving into the code directly.

Scrape anything with AI - FetchFox

FetchFox is an AI-powered web scraping tool that lets users extract data from any website by simply describing what they want in plain English. Available as a Chrome extension or npm library, it enables tasks like collecting leads, market research, or analyzing directories.

AISmartCube - Low Code AI Tools

AISmartCube is a no-code platform that allows you to build and deploy AI tools easily using drag-and-drop functionality, much like assembling Lego blocks. It offers a wide range of features, including access to large language models like ChatGPT and Claude, integration with plugins for tasks like data scraping, SEO, and image or voice processing, and a real-time shared knowledge base to keep your tools updated. You can automate tasks with ready-to-use templates for social media, copywriting, and e-commerce, or customize AI assistants to handle specific workflows.

STORM

The STORM website, developed by Stanford's OVAL lab, is a research preview tool that generates Wikipedia-like reports using AI. Users must agree to terms stating that STORM has limited safety measures, may generate offensive or incorrect content, and should not be used for illegal, harmful, or inappropriate purposes.

Whisk by Google Labs

Whisk is a new experimental tool from Google Labs that allows users to create and remix images by inputting other images instead of using lengthy text prompts. You can provide a subject image, a scene image, and a style image, and Whisk will combine them into unique creations, such as digital art or merchandise designs. The AI behind Whisk uses the Gemini and Imagen models to process the images and generate new combinations, but it is designed for creative exploration rather than precise edits. The tool is meant to quickly experiment with visual ideas, and users can tweak the results if needed.

Titans: Learning to Memorize at Test Time

The paper introduces "Titans," a new family of neural architectures designed to improve memory handling in machine learning models, addressing challenges of scalability and long-term dependency modeling. Traditional Transformers excel at capturing short-term dependencies but face efficiency issues due to quadratic memory complexity. Titans incorporate a novel neural long-term memory module, inspired by human memory, to memorize past data effectively and complement the short-term memory of attention mechanisms. This architecture integrates three key components: short-term memory for immediate context, long-term memory for persistent historical information, and persistent memory for task-specific knowledge.

Agents

Intelligent AI agents are systems designed to perceive and act upon their environment to accomplish tasks, from creating websites to analyzing data. These agents, powered by foundation models, gain enhanced capabilities through tools like knowledge retrievers, web browsers, and code interpreters, allowing them to adapt and perform complex tasks in varied environments. While tools significantly boost their performance, agents face challenges like compounding errors over multiple steps and higher risks due to their ability to perform impactful tasks. Effective agents rely on strong planning capabilities, careful tool selection, and robust security measures to minimize failure modes and ensure reliable, beneficial operation.

HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs

HuatuoGPT-o1 is a medical large language model (LLM) designed to excel in complex medical reasoning by leveraging a novel two-stage training process. The approach starts by using a verifier to guide the model in constructing and refining reasoning trajectories for verifiable medical problems, which are derived from challenging medical exam questions. These refined trajectories are used to fine-tune the model. In the second stage, reinforcement learning (RL) with verifier-based feedback further enhances reasoning abilities. This method enables HuatuoGPT-o1 to iteratively analyze and correct its reasoning, achieving superior performance on medical benchmarks compared to general and medical-specific models, all while using only 40,000 training problems.

AutoGen v0.4: Reimagining the foundation of agentic AI for scale, extensibility, and robustness

AutoGen v0.4 is a major update to Microsoft's agentic AI framework, enhancing scalability, extensibility, and robustness for multi-agent systems. It introduces an asynchronous, event-driven architecture with modular components, enabling seamless communication, debugging, and observability. The framework supports cross-language compatibility (Python and .NET), robust type enforcement, and distributed agent networks. Key tools include AutoGen Bench for benchmarking and AutoGen Studio, a low-code interface for rapid prototyping with real-time updates, interactive feedback, and visual message flow mapping. Additionally, a new multi-agent application, Magentic-One, tackles complex web and file-based tasks.

Agent Laboratory: Using LLM Agents as Research Assistants

Agent Laboratory is an open-source framework that uses large language models (LLMs) to assist researchers in executing machine learning projects efficiently and cost-effectively. It automates key research stages—literature review, experimentation, and report writing—producing comprehensive outputs like research reports and code repositories. Users can provide feedback at each stage, significantly improving output quality. The framework supports various compute levels, making it accessible to different users, and offers a "co-pilot" mode for collaborative research.

facebookresearch/coconut: Training Large Language Model to Reason in a Continuous Latent Space

Coconut is an open-source framework developed by Facebook Research for training large language models (LLMs) to reason in a continuous latent space. It supports end-to-end workflows for research, from preprocessing datasets to training and evaluating models. The framework includes configurations for various reasoning models, like CoT (Chain-of-Thought) and Coconut, with flexible settings for training stages, batch sizes, and checkpoints. Users can customize runs using YAML files and log experiments with wandb. Coconut is designed to reproduce state-of-the-art results on reasoning tasks like GSM8K and ProntoQA, enabling scalable and efficient experimentation with detailed documentation for setup and usage.

Efficient-Large-Model/Sana

Sana is a cutting-edge text-to-image framework developed by NVIDIA that generates high-resolution images up to 4096 × 4096 pixels with remarkable speed and text-image alignment. Based on a Linear Diffusion Transformer architecture with 1648M parameters, it leverages pretrained encoders and advanced diffusion techniques for efficient synthesis. Designed for research and artistic applications, Sana supports creative workflows, educational tools, and the exploration of generative models. While capable of producing stunning visuals, it has limitations in photorealism and handling complex text or detailed features.

vikhyatk/moondream2

Moondream2 is a compact vision-language model optimized for efficient operation on edge devices, enabling tasks like image captioning, visual querying, object detection, and more. With 1.93 billion parameters and FP16 tensors, it offers advanced features such as streaming caption generation and fine-grained visual understanding. Users can easily integrate it via Hugging Face's Transformers library, with options for GPU acceleration.

hexgrad/Kokoro-82M

Kokoro-82M is a lightweight text-to-speech (TTS) model designed for efficient and high-quality audio generation, featuring just 82 million parameters. Despite its compact size, it has achieved top rankings in the TTS Spaces Arena for single-voice settings, outperforming much larger models in Elo ratings. Kokoro supports American and British English, utilizes an Apache 2.0 license, and offers voice customization through multiple pre-trained voicepacks. Trained on less than 100 hours of permissive audio, Kokoro is optimized for edge devices and is easy to use via ONNX or Python-based workflows. Its design is based on StyleTTS 2 and ISTFTNet architectures, prioritizing accessibility and efficiency.

Sky-T1: Train your own O1 preview model within $450

NovaSky, a team from UC Berkeley's Sky Computing Lab, developed Sky-T1-32B-Preview, an open-source reasoning model trained for under $450. This model rivals proprietary reasoning models like o1-preview in tasks like math and coding, while being fully transparent with its data, code, and weights. By refining training methods, balancing diverse datasets, and leveraging efficient infrastructure, NovaSky enables the academic and open-source community to replicate and improve upon their results.

📢 If your company is interested in reaching an audience of developers and, technical professionals, and decision makers, you may want toadvertise with us.

If you have any comments or feedback, just reply back to this email.

Thanks for reading and have a great day!