- The AI Catalyst
- Posts
- CrowdStike disrupts Microsoft systems
CrowdStike disrupts Microsoft systems
PLUS: Small Language Models Arena
Hello, AI Enthusiasts.
A single update from CrowdStrike took down Microsoft systems globally. How big is the impact?
Small Language Model is on the rise. AI tech giants compete with their SLMs.A new AI breakthrough in healthcare. Let’s explore!
This Week’s AI insights:
CrowdStrike/Microsoft Global Outage
OpenAI’s mini-model
Mistral AI and NVIDIA’s Enterprise Model
Hugging Face’s new small model
Google’s AI Developers features.
Gartner’s IT Spending numbers are out
More interesting AI news..
CrowdStrike brings down Microsoft Systems
Image Source: Forbes
A faulty update from the Cybersecurity company CrowdStrike brought down Microsoft hosts worldwide.
The Essentials:
Yesterday, A single update from Crowd Strike's Falcon content system was targeted towards all the Windows hosts, not Linux or Mac.
This update had a defect and so the Windows machines issued the popular "Blue Screen of Death (BSOD)" and were forced to boot loop which prevented it from turning on.
There does not seem to be a universal fix. The only way to fix this is to apply a manual reset at an earlier point in time to receive the latest update issued by CloudStrike.
This global outage affected systems including IT, 911 and Airlines and also took down the London Stock Exchange.
The Takeaway:
Rolling out updates to massive systems should be thoroughly tested. When the stakes are high, the impact is also significant.
OpenAI's mini-model

Image Source: OpenAI
OpenAI unveiled its cost-efficient small model named GPT-4o mini. With this, OpenAI has joined many other LLM companies to create a small model that is widely accessible.
The Essentials:
The GPT-4o mini is 60% cheaper than the GPT-3.5 turbo.
1 million input tokens for 15 cents
1 million output tokens for 60 cents
128K tokens of context window supports up to 16K output tokens per request.
GPT-4o mini scores 82% on MMLU and outperforms GPT-4 on chat preferences in the LMSYS leaderboard
Currently supports text and vision in the API, soon it will include support for text, image, video and audio inputs.
GPT-4o mini knowledge is up to October 2023.
Built-in safety features include :
Pre-training: Undesirable contents are filtered out
Post-training: Accuracy and reliability are improved using Reinforcement learning with human feedback (RLHF)
The Takeaway:
OpenAI's GPT-4o mini is a small model with built-in safety features which outperforms other small models in "Reasoning tasks", "Math and Coding proficiency" and "MultiModal reasoning", a solid contender in this small model arena.
Mistral AI and NVIDIA's Enterprise AI model
Image Source: Mistral
Mistral and NVIDIA together have released a new language model for enterprise applications. This collaboration brings Mistral AI's expertise in training data on NVIDIA's ecosystem.
The Essentials:
Mistral Nemo specifications:
12-billion parameter model
128K context length
Uses FP8 data format to reduce memory size and speed deployments.
Mistral Nemo 12B is released under an Apache 2.0 license.
Excels in multi-turn conversations, math, common sense, coding, etc.
Mistral Nemo 12B can be packaged as an NVIDIA NIM MicroService, which can be deployed on the Cloud, data centre or even in an RTX Workstation.
Mistral Nemo is available as NVIDIA NIM via ai.nvidia.com
The Takeaway:
Mistral Nemo 12B is an open model with unprecedented accuracy, high efficiency and enterprise-grade support MicroService that allows enterprises to easily integrate into their commercial applications.
Hugging Face's new small model

Image Source: HuggingFace
Hugging Face released a series of state-of-the-art small language models called SmolLM. Unlike other small language models, data curation and training of SmolLM models are made publicly available.
The Essentials:
SmolLM is available in three sizes:
135M and 360M models are trained on 600 billion tokens
1.7 billion model is trained on 1 Trillion tokens.
These models are trained on a high-quality training dataset or corpus called SmolLM-Corpus.
Significant improvements were made in data curation by including a large synthetic dataset consisting of over 30 million textbooks, blog posts and stories.
The design of SmolLM is similar to MobileLLM where the focus is on the depth over the width of layers in the neural network.
SmolLM shows strong performance in various evaluation benchmarks including coding tasks.
SmolLM is designed to run locally on smartphones and laptops.
The Takeaway:
SmolLM's high performance in various benchmarks demonstrates that this type of Small language model trained with high-quality datasets can conveniently run on handheld devices like Smartphones and laptops which is a game-changer.
Google’s AI Developer features

Image Source: Google
Google has announced a lot of new features for AI developers at its I/O Connect Bengaluru event in India.
The Essentials:
Project IDX integration - Developers can build apps using a web browser on any device.
Firebase AI monitoring dashboard - Real-time insights into Generative AI performance and behaviour.
Project Vaani - To capture the diversity of spoken Indian languages, the Indian Institute of Science and ARTPARK with Google funding, developed a benchmark named "IndicGenBench" to evaluate the performance of LLMs in Indian languages.
Google Open-sourced the CALM (Composition of Language Models) framework.
AI-Generated UI for Flutter - Dynamically compose and personalize UI elements with natural language inputs.
The Takeaway:
Google's announcements are aimed towards providing developers with the tools for enhancing AI developments across various platforms and languages in India.
Interesting AI news :
Apple's new AI technique MUSCLE
Apple has introduced a new AI technique called MUSCLE ( (Model Update Strategy for Compatible LLM Evolution) to enhance user experience when the LLMs are updated.
It is mainly focused on addressing pain points like unexpected behaviour changes and inconsistencies caused by new LLM updates.
ChatGPT's Strawberry
OpenAI is working on a project codenamed "Strawberry", with which ChatGPT is expected to perform "deep research" on the internet autonomously.
While the specifics of project "Strawberry" are undisclosed, as per Reuters it is expected to improve ChatGPT's ability to reduce the inaccuracies in its responses.
AI breakthrough in Alzheimer's prediction
Cambridge University researchers have developed a new AI tool which can predict patients with mild cognitive impairment that could lead to Alzheimer's disease via analyzing MRI scans and cognitive assessments found to be 80% accurate.
AI giants steal YouTube videos
Proof news revealed that AI tech giants including Apple, Anthropic and Nvidia used YouTube videos to train their AI models.
The dataset "YouTube Subtitles" contains transcripts from Khan Academy, MIT, WSJ, BBC etc, "Jimmy Kimmel Live". Also includes YouTube mega channels like MrBeast, MKBHD and PewDiePie videos were used for AI training.
Chart:
Gartner predicts IT spending reaches $5.26 Trillion

Image Source: Gartner
Key 2024 Forecast in IT and its growth compared to 2023.
Total IT spending: $5.26 trillion, a 7.5% increase
Software: $1.09 trillion, a 12.6% increase
IT services: $1.6 trillion, a 7.1% increase
Data Center: $293 billion, a 24.1% increase
The growth is majorly due to investments in Generative AI.
Magic Prompt :
A delightful photograph of a playful Jack Russell Terrier, eagerly chasing and nipping at a vibrant beach ball. The dog's fur is slightly damp, hinting at its recent splash in the ocean. The idyllic beach setting features golden sand, swaying palm trees, and an azure sky. The bold, glittering gold text on the black background exclaims "Weekend fun!" The image is framed like a captivating poster, highlighting the dog and the beach ball, creating a lively and joyful atmosphere that captures the essence of a fun-filled weekend., portrait photography, typography, photo, posterCredits - aviramb
FEEDBACK
Your feedback is precious and helps us tailor future newsletters to better suit your needs!
We welcome your feedback and insights. Please reply to this email with any comments or suggestions you'd like to share.
Appreciate your time.
AI-ways forward,Naveen Krishnan,Founder and Writer,The AI Catalyst Solutions Inc.
© 2024 The AI Catalyst Solutions Inc. All rights reserved. This newsletter and its content are protected by copyright law. No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of The AI Catalyst Solutions Inc., except in the case of brief quotations embodied in critical reviews and certain other noncommercial uses permitted by copyright law.