. . AI: Tiny 2B Local Models for Android, and Engineers Get Two New Shiny Tools! (4.3.24)

Friends, in today's AI news, researchers and developers make a breakthrough in getting us local models on IoT devices. We also get a new text to music generation tool and two software engineering tools to help fix bugs and develop open-source and decentralized AI projects.

As always, these are the AI stories the AI community is talking about the most. We hope you find them useful and we appreciate you sharing these summaries with friends and colleagues.

-Marshall Kirkpatrick, Editor

Octopus-V2-2B Model Unlocks Highly Trainable, IoT and Edge Capable Models

First impacted: Android API developers, Smart device (like Alexa) users
Time to impact: Medium

Stanford researchers have introduced Octopus-V2-2B, a Gemma-based model tailored for Android APIs. Gemma, a set of open language models stemming from Google's Gemini models, serves as the foundation. The team says it operates 35 times faster than Llama-7B, boasting a 98% accuracy rate. The model achieves its efficiency through innovative techniques like single symbol function names and establishing a meta-language. This approach enhances throughput and simplifies training, requiring as few as 100 samples. Despite its intricacy, Octopus-V2-2B can function on devices like PCs and smartphones; it processes typical queries in 1.1 to 1.7 seconds, highlighting its suitability for real-time applications while addressing concerns related to latency, privacy, and cost. [Paper page - Octopus v2: On-device language model for super agent] Explore more of our coverage of: Android API, AI Efficiency, Smart Devices. Share this story by email

Stable Audio 2.0 Launches AI Music Generation Tool

First impacted: Music producers, Advertisers
Time to impact: Short

Stable Audio has launched version 2.0 of their offering and is capable of custom-length music generation for commercial use. The tool uses the latest audio diffusion models and supports text-to-audio and audio-to-audio conversions based on user-uploaded samples. We signed up for the free account and had some fun playing around with the model. The example on the page was also representative of the model's capability and relatively high quality. [Stable Audio - Generative AI for music & sound fx] Explore more of our coverage of: AI Music Generation, Audio Diffusion Models, Commercial AI Tools. Share this story by email

Replit Launches AI Model to Repair Code Errors

First impacted: Software Developers, Programmers
Time to impact: Short

Replit, a developer platform, has launched an AI model that it says can correct code errors, according to a company blog post. The model, trained on hundreds of millions of Language Server Protocol (LSP) diagnostics, is designed to offer solutions more frequently than the LSP, which only provides fixes in 10% of cases. It is supposedly capable of handling complex multi-line changes, single-line addition/removal, and single-line edits, offering a flexible output space that can be expanded or constrained based on the use case. [Replit — Building LLMs for Code Repair] Explore more of our coverage of: Replit AI Model, Code Error Correction, Language Server Protocol. Share this story by email

Shog Launches Platform for Decentralized AI Access

First impacted: AI researchers, Open-source developers
Time to impact: Short

ShogAI has launched Shog Hub, a platform they say offers open-source and decentralized AI, including access to AI models, source code, datasets, and research papers from the Shoggoth Network. In addition to the launch, ShogAI has updated their node to Shoggoth v0.4.5 and plans to announce several projects in the near future, one potentially including collectively funded decentralized AI computing platforms based on some of the conversations on social media platforms. [Explore Open Source & Decentralized AI] Explore more of our coverage of: "Decentralized AI", "Open Source Platforms", "AI Models". Share this story by email

That’s it! More AI news tomorrow.