Smaller LLM - 15x Faster than Llama 2 🤯

PLUS: Replit's ModelFarm, Stable Audio by Stability AI

Today’s top AI Highlights:

1. DeciLM 6B with Incredible Speed and Quality

2. Text-to-Audio Model by Stability AI

3. Replit ModelFarm for Building Generative AI Applications

4. Text-to-Image Model by Hugging Face

& so much more!

Read time: 3 mins

Latest Developments 🌍

Small Model, Big Impact 🤏

DeciAI has released and open-sourced DeciLM, a 5.7B parameters model that is incredibly fast, and among the most accurate open-source LLMs in the 7 billion parameter class.

Key Highlights:

• The model is based on Deci's Neural Architecture Search engine (AutoNAC), coupled with Deci's inference SDK for throughput enhancement.

• DeciLM's decoder-only transformer architecture employs Variable Grouped-Query Attention (GQA), which varies the number of attention groups, keys, and values across transformer layers.

• DeciLM 6B's throughput is 4.8x that of Llama 2-7B when using an optimal batch size, and 15x when combined with Deci’s Infery-LLM SDK.

Replit’s Platform to Farm Generative AI Apps🧑‍🌾

Replit has introduced Replit ModelFarm, a revolutionary platform for building Generative AI applications, providing the shortest path from idea to production-ready software.

Key Highlights:

• Replit ModelFarm streamlines Gen AI app development by eliminating the need for complex API key management, enabling developers to focus on building.

• Replit ModelFarm supports chat models, code models, and text embeddings, enhancing its versatility for developers.

• Until October 15th, Hacker and Pro subscribers can access select Gen AI models from Google Cloud Vertex AI via Replit's libraries in Python or JavaScript/TypeScript at no cost.

Quick Pass to Custom Audio Creations 🎺

Stability AI introduced Stable Audio, a text-to-audio latent diffusion model, conditioned on text metadata, audio file duration, and start time. This conditioning allows control over the content and length of the generated audio.

Key Highlights:

• Stable Audio features a state-of-the-art architecture, with a variational autoencoder, text conditioning, and a U-Net-based diffusion model for high-fidelity audio generation.

• The model was trained on an extensive dataset of over 800k audio files, spanning music, sound effects, and single-instrument stems, along with text metadata, amounting to more than 19,500 hours of audio content.

• It demonstrates exceptional performance by rendering 95 seconds of stereo audio at a 44.1 kHz sample rate in less than one second on an NVIDIA A100 GPUs.

Swift and Efficient AI Magic for Images 🔮

Hugging Face has released Würstchen, a text-to-image model, a highly efficient diffusion model that's changing the game with extreme spatial compression, reduced training costs, and lightning-fast image generation.

Key Highlights:

• Würstchen's architecture features a two-stage compression process, utilizing VQGAN (Stage A) and Diffusion Autoencoder (Stage B) components, while Stage C operates in the highly compressed latent space, achieving an extraordinary 42x spatial compression, a level of data reduction previously unseen.

• Trained on image resolutions spanning from 1024x1024 to 1536x1536, Würstchen demonstrates adaptability even at resolutions like 1024x2048.

• Along with computational efficiency, it generates images swiftly, outperforming even Stable Diffusion XL, while remarkably reducing training costs by a factor of up to 16x.

Adobe Firefly Goes Mainstream 🎨

Adobe's Firefly generative AI tools, including features like Generative Fill in Photoshop, are now officially available after beta testing. Adobe is also launching a Firefly web app for exploring generative capabilities.

They've introduced a credit-based system for faster access to Firefly-powered workflows, along with a bonus scheme for Adobe Stock contributors whose content helps train AI models.

---

Tools of the Trade ⚒️

• Glide AI: Integrate AI into your apps as simply as adding columns, without managing prompts, choosing models, API complexity, or cache results to optimize cost and performance.

• EY.ai: A new platform by EY to evaluate a business’s current level of AI adoption and uncover gaps, identify opportunities for value creation, and equip teams with AI tools for enhanced productivity.

• LaunchFlow: Quick building and deployment of real-time IoT applications with minimal code, cloud integration, and serverless architecture.

• Noah: AI work assistant that integrates with various tools like Google Drive, Notion, Zendesk and more, with enterprise-level security and compliance.

• CodeWiz: Your AI sidekick for instant web framework assistance, multilingual support, and and saved conversations for an enhanced coding experience.

😍 Enjoying so far, TWEET NOW to share with your friends!

---

Hot Takes 🔥

• “Like the rest of America but even more so, California is not a society. It is an economy.” ~ roon

• I hear a lot of folks in our AI community complain about openAI -- they don't publish, don't release models, maximum-for-profit, etc., so they are more like closedAI rather than openAI.
  This is true, but you have to give it to those guys -- they showed the true potential of LLMs no one thought was possible. ~ Russ Salakhutdinov

---

Meme of the Day 🤡

Source

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

Real-time AI Updates 🚨

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

PS: I 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!