• unwind ai
  • Posts
  • AI Chip with 500,000 tokens per second ⚡️

AI Chip with 500,000 tokens per second ⚡️

PLUS: OpenAI delays Voice Mode, New Figma AI features

Today’s top AI Highlights:

  1. AI chips startup Etched bets big on Transformer AI models

  2. Figma’s new AI features help designers get more creative and stay in the flow

  3. OpenAI is delaying Voice Mode’s rollout

  4. Watch AI edit your videos with Caption’s AI Edit feature

  5. Convert text to human-like speech with Eleven Lab’s iOS app

& so much more!

Read time: 3 mins

Latest Developments 🌍

The race for AI dominance is heating up, but a major bottleneck is hardware. GPUs are becoming increasingly expensive and inefficient. Meta used 50,000x more compute to train Llama 400B than OpenAI used on GPT-2 in 2019. Scaling the next 1,000x will be very expensive.

This is where Etched steps in, with a radical approach to AI hardware – a specialized chip designed specifically for transformers, the architecture behind most cutting-edge AI models.

Etched has developed Sohu, the world’s first transformer ASIC (Application-Specific Integrated Circuit). Sohu is a bold bet that transformers will continue to dominate, and that focusing on this single architecture can unlock unprecedented performance and efficiency.

Key Highlights:

  1. Transformer Performance: Sohu crushes existing hardware, achieving a staggering 500,000 tokens per second throughput for Llama 70B, making it an order of magnitude faster than even the latest GPUs.

  2. Cost-Effectiveness at Scale: Sohu delivers incredible efficiency. A single 8x Sohu server can replace a massive 160 H100 GPUs, dramatically lowering the cost of running AI models.

  3. Specialized for Transformers: By dedicating itself solely to transformers, Sohu achieves an exceptional 90% FLOPS utilization, compared to GPUs’ typical ~30%, making it far more efficient at processing AI tasks.

  4. Open Software Ecosystem: Etched is committed to open-source software, offering developers the flexibility to customize and optimize transformer models for Sohu.

The popular design platform Figma has introduced a suite of AI features to streamline design workflows and boost creativity. These new tools address common pain points of designers, from finding the right assets to generating realistic mockups. The features are currently in beta and will be free for all users until 2024, but the usage limits and pricing will be announced later.

Key Highlights:

  1. Visual Search: Consider this as Google Lens for Figma. It lets you find visually similar designs from your team’s files by simply uploading an image or even selecting a section of your current design, making it easier to reuse existing work and get inspiration from past projects.

  2. AI-powered Text Tools: These tools let you instantly translate text, adjust its length, or rewrite it to find the perfect tone for your design. No more struggling with copywriting for mockups – AI can do it for you.

  3. Realistic Mockup Generation: Forget about using lorem ipsum or placeholder images. Figma’s AI can now generate realistic copy and images for your mockups, making them more engaging and convincing.

  4. Quick-Click Prototyping: Transform static mockups into interactive prototypes with just a click. This makes it easier to bring your ideas to life and get feedback from stakeholders quickly.

  5. Automatic Layer Renaming: Keep your design files organized and developer-ready with this feature. It automatically assigns meaningful names to your layers.

  6. Generate Designs from Text: Have an idea for a design but struggling to get started? Figma’s new AI feature lets you generate UI layouts and component options from a simple text prompt. Just describe what you want, and the AI will create a first draft for you, helping you explore different design directions.

Quick Updates from OpenAI 🤌

  1. OpenAI has made the ChatGPT desktop app available for all Mac users. You can take ChatGPT’s help with just a shortcut (Option + Space), be it any website or app. This lets you ask about anything you’re reading or doing without breaking your flow. Here’s how you can use it to maximize your productivity!

  2. OpenAI has shared an update on the advanced Voice Mode in ChatGPT. The rollout has been delayed to reach “our bar to launch” in terms of safety and reliability. It will be rolled out in alpha with a small group of users to gather feedback, and all Plus users might have access in the fall. The new video and screen-sharing capabilities will be rolled out separately.

We all can feel it 🥲

😍 Enjoying so far, share it with your friends!

Tools of the Trade ⚒️

  1. AI Edit in Captions app: Upload an unedited video, choose an editing style, and watch AI fully edit your video in seconds. It automatically adds transitions, B-roll, graphics, and captions, eliminating the need for complicated editing software.

  1. Notion Sites: This new feature in Notion lets you quickly create and publish websites using over 10,000 templates. Its AI feature helps you write and complete content for blog posts, about pages, and more. Personalize your site with custom domains, themes, and favicons, and easily manage content with drag-and-drop building blocks and databases.

  2. ElevenLabs Reader: Convert text into natural, expressive speech with human-like voices, making it perfect for reading articles, PDFs, and ePubs aloud. You can control the reading speed and use it with various text formats.

  3. Awesome LLM Apps: Build awesome LLM apps using RAG for interacting with data sources like GitHub, Gmail, PDFs, and YouTube videos through simple texts. These apps will let you retrieve information, engage in chat, and extract insights directly from content on these platforms.

Hot Takes 🔥

  1. Are you wondering why Sora or the GPT-4 voice feature hasn't been released yet?
    It is the same reason that Udio and Suno got sued today!
    OAI will be exposed to lawsuits of all kinds unless they can prove categorically they didn't use any copyrighted data to build these models
    The problem with using YT data to train models is that some of the data belongs to large and powerful Hollywood studies that love to sue people and pay lawyers! ~
    Bindu Reddy

  2. Generative video is sucking up all available compute again, expect it to worsen ~
    Emad Mostaque

Meme of the Day 🤡

That’s all for today! See you tomorrow with more such AI-filled content.

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!

PS: We curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!

Reply

or to participate.