Small Models are the next BIG thing in AI?

Apple, Microsoft launched smaller models that are capable of huge things. Microsoft's Phi-3 is even better than LLama-3. Let's take a look.

April 25, 2024

Smaller models or SLMs (Small Language Models) are talk of the town this week. Phi-3, OpenELM announced by Microsoft and Apple have been killing the benchmarks so far. Let’s talk a bit about both and why these can be huge for smartphone devices.

Apple has been in talk with Google for integrating Gemini Nano into iPhones, but this is unlikely to go through after they acquired Datakalabs last week and subsequently launched open source OpenELM models today. Even Microsoft launched PHI-3 which seems to be beating LLama-3 (8bn parameter variant) on a few benchmarks. Note this is just a 3.8bn parameter model. So are small models the next “BIG” thing? no pun intended.

Stuff you should know

Apple announces OpenELM

Apple has introduced OpenELM, a family of open-source large language models designed to operate locally on devices, marking a shift from reliance on cloud-based processing. OpenELM utilizes a layer-wise scaling strategy to efficiently allocate parameters, improving accuracy while reducing computational load. The project includes open-source availability, a comprehensive training framework, and enhanced privacy and speed through on-device processing. Apple plans to incorporate OpenELM into the upcoming iOS 18 release, which is expected to introduce new AI features and potentially power more advanced versions of Siri and other AI-driven applications.

Huggingface links: https://huggingface.co/apple/OpenELM

Microsoft launches Phi-3 Models

Microsoft has introduced the Phi-3 family of small language models, including Phi-3-mini, Phi-3-small, and Phi-3-medium, with 3.8 billion, 7 billion, and 14 billion parameters, respectively. These models are designed for specific tasks and are more accessible to organizations with limited resources. They outperform models of similar and larger sizes on various benchmarks, and their capabilities are comparable to models with significantly more parameters, such as GPT-3.5. The Phi-3 models are trained using high-quality, meticulously curated data and are suitable for simpler tasks, custom applications, education, coding, and reasoning. Phi-3-mini can run locally on consumer devices, and the smaller size of these models may help reduce the environmental impact of AI.

AI will soon take over Government documentation

Axon AI's Draft One is an AI-powered report-writing software designed to assist public safety professionals, announced in November 2023. It aims to streamline the report-writing process, making it more efficient and accurate by using advanced AI algorithms to analyze and summarize information. Axon describes Draft One as a "first-of-its-kind AI-powered force multiplier for public safety." The software can help generate comprehensive reports quickly and easily, improve the accuracy of information analysis, enhance working conditions for public safety professionals, and ultimately contribute to public safety enhancement. Axon's founder, Rick Smith, believes that AI-powered solutions like Draft One can help scale police work, making it more efficient and effective.

Handpicked video

I built a LLama-3 powered AI Agent, take a look:

Top Products launched this week 

  • 150+ Prompt Templates on GAI Courses: GenAI Courses has added 150+ Prompt Templates to their repository for all disciplines.

  • Insights by Ayraa -  you can query your workplace for Insights - not just summaries! What should I work on? What did I accomplish last week? Any important Slack threads for me? How many hours did I spend in meetings last week? And so much more! Your workplace AGI assistant.

  • Intrvu Space - Intrvu SPACE is an end-to-end interviewing solution, that helps you automate conducting Interviews, Generating Reports and automating approvals. 

  • Voxal.AI - Drive growth and enhance engagement, boost branding, and maximize conversions effortlessly. More than just a standard chatbot.

  • LangWatch - LangWatch provides an easy, open-source platform to improve and iterate on your current LLM pipelines, as well as mitigating risks such as jailbreaking, sensitive data leaks and hallucinations.jailbreaking, sensitive data.

  • Wizad - Wizad is your go-to app for effortlessly creating stunning social media posters that perfectly match your brand's identity. Say goodbye to the hassle of hiring designers or spending hours tweaking templates.

  • MarketerGrad by Pangea- Created a vetted network of top marketers and designers who have experience taking products from 0 to 1. Our AI will instantly recommend relevant talent so you can start reviewing profiles in seconds.

  • Dart - Dart is an intelligent project management tool that automates and enhances many standard PM functions. Dart's integrated AI can generate reports, break tasks into subtasks, detect duplicate tasks, develop roadmaps, and even execute basic tasks for you.

This week in AI

  • New Ray-Ban Meta Smart Glasses -  Meta has introduced new Ray-Ban Meta Smart Glasses styles and AI updates, including the Skyler frames with a cat eye design and a low bridge option for Headliner frames. These glasses offer hands-free control, real-time information, and a multimodal AI update for visual context awareness. The new Ray-Ban Meta smart glasses collection is available in various styles and colors, with Meta AI integration for enhanced user experience.

  • Nvidia buys Run:ai for $700m  – Nvidia has acquired Israeli startup Run:ai for $700 million, which specializes in artificial intelligence (AI) computing resources management. The acquisition is part of Nvidia's strategic moves to strengthen its position in the AI sector and optimize infrastructure management for AI workloads.

  • Google Axion Processors - Google's Axion Processor is a custom Arm-based chip for data centers, offering high performance and energy efficiency for various workloads like web servers, databases, AI, and more.

  • How to Stop ChatGPT’s Voice Feature From Interrupting You - President Biden signed a bill that could lead to a ban on TikTok unless its Chinese parent company, ByteDance, sells the app within nine months. TikTok plans to challenge the ban in court, citing concerns over free speech and privacy. The situation raises legal and constitutional issues, with potential implications for the app's future in the U.S.