Europe's biggest AI development

Mistral’s New Model Makes Waves Beyond Tech

On Monday, French AI startup Mistral quietly dropped their new LLM on X. But behind this casual drop is seriously powerful tech—and the potential to shape the future of European AI.

The details: Mistral’s new LLM, known as Mixtral 8x7B, is an open-weight LLM with a permissive license (i.e. it’s mostly open-source) that reportedly matches or even outperforms OpenAI’s closed-source GPT-3.5. It also outperforms Meta’s LLaMa 2 70B while running 6x faster—making it the strongest open-weight model to date. 

How it works: Mixtral 8x7B uses an advanced technique called “mixture of experts” (MoE) to match rivals’ performance at a far lower cost. MoE divides the LLM into several parts, each specialized in different aspects of the training data. The model dynamically chooses which part to use based on the input it receives, optimizing performance for different prompts. Other highlights:

• Mixtral 8x7B’s small size allows it to run locally on machines without dedicated GPUs (like Apple’s computers with M2 Ultra CPU).

• Because it’s (basically) open-source, developers around the world have free access to Mixtral 8x7B’s powerful capabilities. 

Mistral’s larger impact: Last week, the European Union finally passed the EU AI Act, a milestone deal on AI regulations. The Act requires general purpose AI models classified as ‘high risk’ to adhere to extensive safety requirements, while largely exempting open-source LLMs from strict regulation.

Unsurprisingly, Mistral appears to have been leading the charge against the regulation of open-source foundation models—and it looks like it succeeded.

Big picture: As the most well-seeded startup in European history—and now boasting one of the most powerful LLMs—Mistral is the EU’s shot at competing against Silicon Valley giants.

The company’s evident involvement in the EU AI Act makes it clear: the region wants AI sovereignty and regulation, but first needs Mistral to succeed.  

The Ultimate Growth Tool for YouTube Channels

It’s never been easier to optimize your YouTube channel—and seriously increase views and subscribers.

Loved by over 10 million creators and businesses, TubeBuddy is a browser extension that uses advanced AI to create, manage, and grow your channel.

With TubeBuddy, you can layer tools directly on top of YouTube’s website:

• Boost click-through rates with the Thumbnail Analyzer tool.

• Get to the top of YouTube’s search results with the Keyword Explorer tool.

• Easily compare different titles and thumbnails with the A/B testing tool.

Start growing your channel like the top creators today.

Tesla’s New AI Robot

Yeah, the Cybertruck is cool and all, but have you seen the new Optimus Gen 2 humanoid? Let’s explore Tesla’s new AI-powered robot, the latest advancement in AI hardware.

Some context: In a demo video released on Tuesday, Tesla showed off the latest version of its humanoid robot. This comes a year after the company presented the first Optimus prototype in a shaky demonstration that left the tech crowd…disappointed (I’m being diplomatic here).

The details: Compared to last year’s version, the Optimus Gen 2 is 10 kg (22 lb) lighter, 30% faster at moving around, and a lot more human-like overall. Features include:

• Improved balance and full-body control (thanks to features like human foot geometry)

• Tactile sensing on all fingers, allowing for the delicate handling of objects (like eggs, as shown in the video demo)

• More range of motion in the hands and neck

Zoom out: AI hardware notoriously innovation lags behind software. But this year saw some notable improvements in the space (remember when Nvidia’s model taught a robot how to spin a pen?)—and Tesla’s new updates to Optimus continue to pave the path forward.

Microsoft’s Phi-2 Packs a Punch

Microsoft Research just released a tiny new AI model that achieves state-of-the-art performance. Enter: Phi-2.

Size matters: As a small language model (SLM) with just 2.7 billion parameters, Phi-2 was trained on a much smaller dataset than LLMs like GPT-4 (which is rumored to have been trained on more than 1 trillion parameters).  

What makes Phi-2 so special? It rivals LLMs up to 25x its size.

• According to Microsoft:Phi-2 outperforms the 7 billion parameter models Mistral 7B and LLaMa-2 70

• Because it can run directly on-device, Phi-2 could open the door for greater AI integration into everyday applications.

A vertically-integrated future: Microsoft recently expanded its tech stack with the introduction of custom chips last month. Coupled with Phi-2's launch, this signals the company's shift towards complete vertical integration in AI and cloud computing.

Bonus Story:

Google just released MusicFX, a new AI tool that can generate music with just a few words. This latest innovation, part of Google’s AI Test Kitchen, uses Google’s MusicLM and DeepMind’s watermarking technology, SynthID. The catch: to protect artists, the tool will not generate music for queries that mention specific artists or include vocals.

Interested in learning more? Read about it here.

• Researchers at the ETH Zurich developed a new technique that could make LLMs up to 300x faster. 

• Read more about the EU’s groundbreaking AI Act negotiations here. 

• Bye-bye Discord—Midjourney is finally bringing image generation to its own website. 

• Meta unveiled Audiobox, a new model for generating voice and sound effects.

• Snapchat shipped more generative AI features for its Plus subscribers.

More important AI news: Dive deeper into this week’s hottest AI news stories (because yes, there are even more) in my latest YouTube video:

Next-level AI video: Watch me explore and breakdown everything AI’s hottest new video tool, Pika 1.0, can do (spoiler: it’s pretty mind-blowing):

Free tools you need to try: You don’t have to spend a single cent on some of the best AI tools out there. Here are a few of my favorites: