Welcome back!

Apple has unveiled technical details behind their on-device model powering Apple Intelligence. Using fine-tuned adapters and other techniques, Apple has struck a balance between performance, size, and energy efficiency, while still surpassing much larger models. Let’s explore...

Today’s Summary:

• Apple on-device AI outperforms larger models

• Enterprises scramble for AI talent

• Mistral raises $640M

• Apple will also collaborate with Google Gemini

• Apple stock hits record high after AI announcements

• Elon Musk drops OpenAI suit

• Microsoft ends Copilot GPT Builder feature

• 2 new tools

TOP STORY

Apple On-Device AI Will Run 3B Models Outperforming Larger Models

The Summary: Apple shared technical details about its on-device 3B parameter model powering Apple Intelligence. It outperforms larger models like Mistral-7B on the required use-cases. 

The model uses a series of LoRA adapters, specializing the AI on precise capabilities, which are dynamically loaded and swapped depending on the current task. Rigorous evaluations demonstrate the model superiority on metrics like instruction-following and writing quality.

Key details:

• 3B on-device LLM runs with 0.6ms latency, 30 tokens/sec

• Uses grouped-query-attention, shared embeddings, quantization

• Dynamically loads LoRA adapters for task specialization

• Outperforms Mistral-7B, Microsoft Phi-3-mini, Google Gemma on IFEval benchmark

• For more complex tasks, Apple uses an additional model running on Apple Private Cloud that matches the performance of GPT-4-Turbo.

Why it matters: Apple's new 3B on-device AI model shows that small, optimized models can outperform larger ones on specific tasks. This technology promises to make AI faster, more efficient, and accessible for everyday use on our personal devices, without the need for massive cloud servers.

GUIDE

Enterprises Scramble for AI Talent Amid Skills Crunch

The Summary: As AI becomes a priority, IT leaders face a severe shortage of AI talent. The vast majority of employers expect to leverage AI in the next few years, but 75% struggle to find the required skills. CIOs are leading the charge, combining hiring, upskilling, contracting, and retention tactics to build teams.

Hiring experienced AI professionals remains intensely competitive, so many enterprises are building AI skills internally through training programs. The needed skills span data, modeling, governance, and identifying use cases.

Key details:

• 92% of employers expect to integrate AI solutions by 2028

• 93% plan to leverage generative AI within 5 years

• 73% prioritizing hiring talent with AI skills/experience

• 75% struggling to find required talent

• Enterprises upskilling current staff for AI roles

Why it matters: AI can transform businesses, but it depends on skilled professionals. The current shortage of AI talent is a big challenge, slowing down adoption. Companies will need to be creative in hiring and training to keep up.

MISTRAL AI

Mistral AI Raises $640M, Unveils Model Customization

The Summary: Mistral AI closed a $640 million Series B, valuing the startup at $6 billion. Alongside the funding, Mistral unveiled tools for customizing and fine-tuning their AI models to provide better performance, speed, and control for specific use cases. The offering includes online serverless fine-tuning services on their La Plateforme, and a free SDK for developers.

Key details:

• Availability of Mistral fine-tuning services using LoRA adapters

• Serverless or free SDK for developers to use on their infrastructure

• Initially supports Mistral 7B (open-source) and Mistral Small

• $640 million Series B funding round at $6 billion valuation

Why it matters: Fine-tuning existing AI models allows for better responses, flexibility and efficiency for specific business applications. Mistral fine-tuning tools makes this process easier, allowing enterprises to customize their own generative AIs at low cost. The $640M war chest provides firepower for Mistral to maintain its AI roadmap, which promises to further advance open source AI.

QUICK NEWS

Quick news

• Apple confirms plans to also work with Google Gemini

• Apple stock surges to record high after AI announcements

• Elon Musk drops suit against OpenAI and Sam Altman

• Microsoft kills off Copilot GPT Builder after 3 months

TOOLS

🥇 New tools

• Invisibility - Mac shortcut app to ask multiple AI models

• Elai - Create avatar-based training videos with quizzes