Skip to main content
Back

The First AI PC and Lightest LLMs: Microsoft is Back

READ ANNOUNCEMENTS

May 22, 2024 5 min read

Hey,

Welcome to today's edition of AlphaSignal. 


Whether you are a researcher, engineer, developer, or data scientist, our summaries are there to keep you up-to-date with the latest breakthroughs in AI. 


Let's get into it!


Lior

.

IN TODAY'S SIGNAL

Top News

Lambda 

  • Get on-demand instances in seconds and accelerate AI workloads with NVIDIA H100 on Latitude.sh.

Trending Signals

Top Repos

PyTorch Tip

  • Lambda Layers enable dynamic, function-based model modifications based on inputs.

Read Time: 5 min 41 sec

Enjoying this newsletter?
Please forward it to a friend or colleague. It helps us keep this content free.

TOP NEWS

Industry

Microsoft unveils new Phi-3 models, GitHub Copilot updates, AI PCs, and more

⇧ 5821 Likes

What's New

The AlphaSignal team was invited to the Microsoft Build conference where new AI tools and breakthroughs were released at every layer of the tech stack.


Let's go through each of the releases!


Small Language Models

Microsoft wants to take the lead on the Small Language Model race. After their tiny but mighty Phi-3 released a few months ago, they just added new models to the Phi-3 family:

  • Phi-3-vision: the first 4.2B multimodal model of the Phi family

  • Phi-3-small: a new 7B parameters language model that rivals with models twice it size. 

  • New Phi-3 model: 14B parameter trained on 4.8T tokens achieving MMLU of 78 (comparable with Llama3 70B)

  • Phi-Sliica: included in every Copilot+PC designed for NPUs with 650 tokens/sec prompt processing.


Microsoft Copilots and GitHub Copilot 

Many updates and new features were added to the Copilots family.

  • GitHub Copilot Extensions: Microsoft is empowering any developer to customize the Copilot experience with services like Azure, Docker, or Sentry directly within GitHub Copilot Chat.
  • Team Copilot: An expansion of Copilot from a personal AI assistant to a collaborative team member. Team Copilot can act as a meeting facilitator, note-taker, project manager, and collaborator across Microsoft 365 apps like Teams, Loop, and Planner. This moves Copilot capabilities from just assisting individuals to also assisting teams.

  • Copilot Studio Agents: New agent capabilities were added in Copilot Studio to allow developers to build proactive Copilots that can independently manage complex business processes. These Copilots can leverage memory, learn from feedback, and ask for help when needed.


New Copilot+ PCs

Microsoft introduced the world to a new category of Windows PCs designed for AI: Copilot+ PCs. Copilot+ PCs are the fastest, most intelligent Windows PCs ever built.

  • Windows Copilot library and new Phi-Silica, a SOTA Small Language Model, included in every Copilot+PC.

  • PyTorch and a new Web Neural Network - a web native machine learning framework - now native on Windows

  • Powerful new silicon capable of 40+ TOPS

  • AI tools like Recall (to easily find and remember what you have seen in your PC), Cocreator (for AI image generation), Live Captions (for real-time audio translation), and more

  • 20x more powerful and up to 100x as efficient for running AI workloads and deliver industry-leading AI acceleration


Other announcements

Microsoft also announced Real Time intelligence within their Fabric app the Azure AI-powered analytics platform, now in preview, as well as a long list of partnerships from AMD to Khan Academy or Cognition AI. 

Access

  • You can try Phi-3-vision now on the Azure AI platform. 

  • Phi-3 is also available on HuggingFace

  • GitHub Copilot Extensions are supported in GitHub Copilot Chat on GitHub.com, Visual Studio, as well as VS Code.

Community Response

Bindu Reddy
The new Phi-3 14B model scores phenomenally on all benchmarks. On key numbers, it seems to be pretty close to Llama-3-Instruct 🤯🤯 As small models become more and more powerful, we will see 7b-sized GPT-4 class models, which will be instantly useful for simple tasks Open-source will not only catch up but will do better in both cost and efficiency.
Tom Warren
Microsoft is making a big bet on building a new type of AI computer. AI chips are at the core of Microsoft’s new Copilot Plus PCs, and Microsoft thinks they could change the way we use computers.
Read announcements

Get On-demand Instances in Seconds

Accelerate your training and inference workloads with dedicated GPU instances powered by NVIDIA’s H100 Tensor Core available at Latitude.sh's platform.


With pre-installed AI software, high-performance networking, and enterprise-grade hardware, you get from zero to training your LLMs with just a few clicks.

GET MY GPUs

partner with us

TRENDING SIGNALS

Google

Google releases a new Gemini 1.5 Technical Report claiming it is the most capable LLM on the market

⇧ 1942 Likes

Object Detection

DINO-1.5 is the new state-of-the art model for object detection

⇧ 981 Likes

ChatGPT

The EU AI Act gets its official green light, the world's first comprehensive regulations for AI

⇧ 357 Likes

Industry

OpenAI takes Sky voice down after Scarlett Johansson's accusations of using her voice

⇧ 2381 Likes

Theory

New paper shows AI outperforms humans in Theory-of-Mind test

⇧ 621 Likes

TOP REPOS

Crowd-Sourcing

Fabric

☆ 10,701

fabric is an open-source framework for augmenting humans using AI. It provides a modular framework for solving specific problems using a crowdsourced set of AI prompts that can be used anywhere.

Multimodal Models

Lumina-T2X

☆ 1170

Lumina-T2X is a series of text-conditioned Diffusion Transformers (DiT) capable of transforming textual descriptions into images, videos, detailed multi-view 3D images, and synthesized speech. 


This family of models allow generation of output in any modalities, resolution, and duration within one framework, and requires low training resources. 

RAG

Verba

☆ 3715

Verba is Weaviate's open-source application designed to offer an end-to-end, streamlined, and user-friendly interface for Retrieval-Augmented Generation (RAG) out of the box. 

You can run Verba either locally with HuggingFace and Ollama or through LLM providers such as OpenAI, Cohere, and Google. You can also choose between different RAG frameworks, data types, chunking & retrieving techniques in a very customizable way. 

PYTORCH TIP

Model Architecture

Using Lambda Layers for Dynamic Adjustments in Network Behavior

Lambda Layers in PyTorch can be used to incorporate function-based modifications within a model dynamically, which can adjust to varying inputs or conditions without restructuring the entire network.


Here’s a practical example demonstrating how to use a Lambda Layer to conditionally apply different operations based on input characteristics:


import torch
import torch.nn as nn

class DynamicLambdaLayer(nn.Module):
    def __init__(self):
        super(DynamicLambdaLayer, self).__init__()

    def forward(self, x):
        # Apply a different function based on input
        if x.mean() > 0.5:
            return x * 2  # Scale up feature values if mean > 0.5
        else:
            return x * 0.5  # Scale down otherwise

# Example usage in a model
model = nn.Sequential(
    nn.Linear(10, 20),
    DynamicLambdaLayer(), # Inserting the dynamic lambda layer
    nn.ReLU(),
    nn.Linear(20, 5)
)

# Test input
input = torch.randn(1, 10)
output = model(input)
print(output)

Ghita is an editor at AlphaSignal and AI Engineer at Clear Ventures. Experienced in Computer Vision and Language models, she holds 2 masters in Applied Mathematics.

Thank You.

Want to promote your company, product, job, or event to 180,000+ AI researchers and engineers?

Sponsor