Daily AI Newsletter: A new AI chip from Meta

In today’s email:

🔥 OpenAI and Meta ready new AI models capable of ‘reasoning’
👨🏻‍🏫 Texas is replacing thousands of human exam graders with AI
👨🏻‍💻 Karpathy is Back with llm.c: A Pure C Implementation of GPT-2 in <1000 Lines
🧰 7 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

Meta unveils its newest custom AI chip as it races to catch up

Meta has launched its latest custom AI chip, the "next-gen" Meta Training and Inference Accelerator (MTIA), aiming to boost its generative AI capabilities and reduce dependency on third-party GPUs. This new chip, an advancement from the previous 7nm MTIA v1 to a 5nm process, offers more processing cores, higher power consumption, and enhanced memory and clock speed, promising up to three times the performance of its predecessor. Meta's move comes amidst a backdrop of intense competition in the AI chip market, with rivals like Google, Amazon, and Microsoft unveiling their custom solutions.

Despite the technical upgrades, Meta acknowledges that the next-gen MTIA is not yet deployed for generative AI training workloads, suggesting a gradual integration strategy alongside existing GPUs. The company's rapid development pace—from design to production in under nine months—highlights its urgency in catching up with competitors, although it still faces significant challenges in achieving hardware autonomy and cost efficiency in the high-stakes AI sector.

Meta's investment in its own AI chip development underscores a broader industry trend where tech giants are increasingly focusing on custom hardware to power their AI ambitions. As Meta strives to enhance its AI capabilities and reduce reliance on costly external hardware, the competitive landscape continues to evolve, with key players pushing the boundaries of AI chip innovation and performance.

Your Personalized Photography Studio, On Demand!

Artflow.ai - The first-ever tool that lets you create images with two consistent identities. Design your dream scenario, from facial expressions and outfits to exotic locations, all without stepping outside your door. Perfect for group photos, transform your visions into shared memories with just a few clicks.

Try now

Google brings AI-powered editing tools, like Magic Editor, to all Google Photos users for free

Google has announced that its AI-powered editing tools in Google Photos, including Magic Editor, will now be accessible to all users for free. These features, which were previously exclusive to Pixel devices and Google One subscribers, include the Magic Eraser for removing unwanted objects, Photo Unblur, and Portrait Light for post-photo lighting adjustments. The rollout will begin on May 15, but there are specific hardware requirements for different devices to use these tools.

Magic Editor, a standout feature, allows users to perform advanced edits like changing backgrounds and repositioning subjects using AI. Initially available on Pixel 8 devices, it will now be extended to all Pixel devices, and users on iOS and Android can enjoy 10 Magic Editor saves per month, with additional access available through a Premium Google One plan.

In addition to Magic Editor, other tools such as Sky suggestions, HDR effects, and Portrait Blur will be available to all Google Photos users without a Google One subscription. However, some advanced features like Best Take, which merges similar photos to create an optimal image, will remain exclusive to Pixel 8 users.

Mistral AI drops new ‘mixture of experts’ model with a torrent link

Mistral AI has launched its new sparse mixture of experts (SMoE) model, Mixtral 8x22B, offering an innovative approach to AI with a download available via a torrent link. This release follows other major AI model launches, positioning Mistral as a noteworthy competitor in the AI landscape. Unlike traditional rollouts, Mistral opted for a torrent distribution, drawing attention to its substantial size and potential challenges for local use, indicating a significant advancement in AI technology.

The Mixtral 8x22B model utilizes a sparse MoE architecture, enhancing efficiency by selectively engaging different model experts for various tasks. This technique allows for a more dynamic allocation of resources, optimizing performance while managing costs. Mistral's approach aims to combine the strengths of various models, achieving superior results and setting a new standard in AI capabilities.

Despite the buzz, the full capabilities and performance benchmarks of Mixtral 8x22B are yet to be fully disclosed. The AI community is keenly observing how this model stacks up against other major players like Meta's upcoming Llama 3 and OpenAI's GPT-4 Turbo. With its innovative distribution and promising architecture, Mistral's latest release is a notable development in the evolving landscape of artificial intelligence.

Other stuff

OpenAI and Meta ready new AI models capable of ‘reasoning’
Three major LLM releases in 24 hours
A new bill wants to reveal what’s really inside AI training data
A list of top AI agents
Texas is replacing thousands of human exam graders with AI
OpenAI’s Altman Pitches Global AI Coalition on Trip to Middle East
Karpathy is Back with llm.c: A Pure C Implementation of GPT-2 in <1000 Lines
Human influencers are being replaced by artificial intelligence — and maybe that's a good thing.
Majorly improved GPT-4 Turbo model available now in the API and rolling out in ChatGPT.