24 | Microsoft Build | LLMs Are not Black Boxes Anymore

Our Menu :)

• Copilot: A New Era for PC Interaction

• Anthropic Unveils Breakthrough AI Research

• Around the table  

Actionable Tips:

• The first step into GenAI

• AI REV HUB - Learning Gen AI path

Jobs Opportunities

•  Job Listings this week

Read time: 6 Minutes

GEN AI AT WORK
Copilot: A New Era for PC Interaction 

May 22-23, 2024. Microsoft kicked off its annual Build developer conference with a bang, making a slew of artificial intelligence (AI) announcements that position the company as a frontrunner in the AI race.

Here are some of its standout features:

1. Screen Monitoring

Copilot can observe your screen while you work, similar to a feature recently introduced by OpenAI with ChatGPT-4o. Running on GPT-4o, Copilot's vision capabilities are poised to enhance tasks such as tech support and image analysis. This integration with Office applications makes it particularly powerful. For example, you can ask, "Does this Word document match our company’s template?"

2. Enhanced Search with "Recall"

Recall is a feature that keeps a continuous record of your screen activity, enabling you to locate anything you've ever searched for or viewed on your computer. For instance, you can ask it to find the “The AI Rev Hub article about GPT-4o from last week” or the “document sent by [x client] last week.” Despite the potential privacy concerns, Microsoft assures that all data seen by Recall remains private as the laptop running the feature uses a local model on-devise.

3. Real-time Translation and Transcription

This new feature offers real-time translation and transcription for meetings across any call platform (Zoom, Teams, Meet) in over 40 languages.

4. AI-Optimized Copilot+ PCs

Microsoft introduced a new category of AI hardware called Copilot+ PCs. These laptops feature specialized chips, all-day battery life, and advanced AI capabilities like real-time image creation and live language translation. Prices start at $999 when they launch on June 18.

5. Phi-3 Family

Microsoft is heavily focused on developing Small Language Models (SLMs) which are efficient AI models designed for use on mobile devices. They see two main reasons for SLM adoption:

• The desire for local inference on mobile devices for privacy and reducing cloud dependency.

• Microsoft's success in building AI-powered mobile apps, transitioning from apps to AI agents.

They have introduced a new multimodal AI model called Phi-3-vision that can handle both text and image inputs, allowing it to analyze things like charts and graphs. It is currently in preview and has 4.2 billion parameters.

• Phi-3-mini is a 3.8B parameter language model, available in two context lengths (128K and 4K).

• Phi-3-small is a 7B parameter language model, available in two context lengths (128K and 8K).

• Phi-3-medium is a 14B parameter language model, available in two context lengths (128K and 4K).

You can find all Phi-3 models on Azure AI and Hugging Face.

6. Models-as-a-Service

Microsoft's MaaS on Azure lets developers use and customize top AI models from various companies on a pay-as-you-go basis, without needing dedicated virtual machines.

New additions include Nixtla TimeGen-1 and Core42 JAIS, among others. MaaS simplifies AI integration and monitoring, promoting innovation while reducing costs.

Microsoft has also expanded its partnership with Hugging Face to make open AI models more accessible across hardware

Conclusion

Microsoft excels at partnerships and collaborations, which help them differentiate between what's impossible and what's just challenging. Kevin Scott emphasizes that the complexity of their systems means no single team can uncover everything, and they rely on others for new ideas.

Satya Nadella's focus on cloud computing and AI has positioned Microsoft as a leader in these fields, with strategic acquisitions and partnerships further boosting their capabilities. The integration of AI into hardware through the Copilot Stack has revitalized the appeal of Windows PCs.

GEN AI AT WORK
LLMs Are not Black Boxes Anymore

Anthropic has released a groundbreaking paper exploring the inner workings of a Large Language Model (LLM) for the first time.

The Challenge: Understanding Neural Networks

• Black Box Problem: Traditionally, LLMs functioned like black boxes, producing outputs without clear reasons.

• Neurons and Features: LLMs consist of neurons activated by certain inputs, forming features that create the model's internal state.

• Interpreting Neurons: Most neurons in LLMs are hard to interpret, hindering understanding.

New Technique: Dictionary Learning

• Method: Anthropic used "dictionary learning," which identifies recurring neuron activation patterns. This helps represent the model's internal state with a few active features.

• Application: Previously successful on tiny models, this technique was applied to Claude 3 Sonnet with remarkable results.

Findings

• Millions of Features: Researchers mapped millions of features, creating a conceptual map of the model's internal states.

• Conceptual Mapping: Similar features are located close to each other, reflecting conceptual similarities.

• Feature Manipulation: Emphasizing or deemphasizing features can change the model's behavior. For instance, activating a feature for blindly agreeing with the user altered the model's responses.

Why It Matters

• Safety Implications: This technique could help ensure AI safety by controlling and manipulating features.

• Future Work: Mapping all features is computationally intensive, and more research is needed to fully understand the impact on AI safety.

GEN AI PROTECTION
Around the table 

• OpenAI has inked a deal with News Corp (owner of WSJ, Barron’s, MarketWatch, & more) to integrate its news content into ChatGPT. The agreement could be worth up to $250M, according to the WSJ.

• Humane, the overhyped Ai Pin that wanted to replace the iPhone, is reportedly seeking a buyer.

• Nvidia’s data center arm grew 427% since last year, driven by demand for its AI chips.

• Amazon is planning to overhaul Alexa with an AI makeover later this year and charge a subscription for it.

• DeepL, which competes with Google Translate, secured $300M to improve its text translation and writing tools.

GEN AI STARTUP
The first step into GenAI 

It is tough to understand what you need to focus on if you want to learn about the field, considering the amount of work that has been done on the subject.

Those articles are great starting points to get a strong overview of what you need to know if you want to work with LLMs:

- Large Language Models: A Survey: https://lnkd.in/guWeHZHz
- A Survey of Large Language Models: https://lnkd.in/gFtc3ptC
- A Comprehensive Overview of Large Language Models: https://lnkd.in/gb2jesjX

GEN AI AT WORK
 Job Listings this week

Remember that we have launched our beta job watcher

• Robust Intelligence is seeking a Senior Product Manager focused on Security. Link

• OctoAI is looking for a Staff Technical Product Manager to join their team. Link

• Cohere is seeking a Member of Technical Staff, Evaluation to join their team. Link

• Mistral AI is looking for a Revenue Development Representative to join their team. Link

Thank you, see you next week!