23 Sep 2025

AI News Digest

🤖 AI-curated 7 stories

Today's Summary

Microsoft’s got a new playground for AI enthusiasts with Windows AI Labs, where they’re rolling out experimental features in apps like Paint and Notepad. It’s a clear move to make AI a regular part of our desktop lives, not just something floating in the cloud. Meanwhile, Replit is making waves with a hefty $250 million boost and its autonomous Agent 3 tool, pushing the envelope on automated app development and giving folks without a coding background a real shot at building software.

Stories

Microsoft launches Windows AI Labs to test experimental AI features inside core apps

Microsoft this morning announced Windows AI Labs, a pilot program that gives early testers access to experimental AI features across Windows apps — with Paint the first app to get the capability. The Labs program is designed to validate new AI ideas quickly by gathering user feedback on usability and market fit. Why it matters: Windows AI Labs signals Microsoft’s continued push to embed generative-AI features at the OS level (not just in cloud services), accelerating how AI tools reach everyday users and creating another battleground with other platform players integrating AI across devices and productivity apps.

Surat‑based Rocket raises $15M seed to scale its AI app‑building platform

Rocket, an AI-powered platform that converts natural‑language prompts into production‑ready apps and websites, raised $15 million in a seed round led by Salesforce Ventures and Accel (with Together Fund also participating). Rocket says it already has ~400,000 users across 180 countries and plans a North American HQ in Palo Alto to accelerate product development and global expansion. Why it matters: the round highlights continued investor interest in tools that democratize app development (vibe‑coding / no‑code + generative AI), and Rocket’s growth suggests strong demand for faster, AI‑assisted workflows for building software.

Read more → The Economic Times (ETtech)

REFRAG: new decoding technique promises 16× longer context and ~30× faster RAG decoding

Researchers (authors list on the arXiv paper) introduced REFRAG, a retrieval‑augmented‑generation (RAG) decoding framework that compresses retrieved passages into chunk embeddings and selectively expands only the most informative chunks via an RL policy. Reports summarizing the arXiv preprint show REFRAG can drastically reduce decoder input length (e.g., ~16×), speed time‑to‑first‑token by up to ~30×, and extend effective context windows — all while maintaining perplexity and downstream accuracy on long‑document RAG tasks. Impact: this could make large‑context, knowledge‑intensive LLM applications (enterprise RAG, long‑document summarization, multi‑turn assistants) far more practical by cutting latency and memory costs without changing existing decoders.

ParaThinker: Tsinghua team proposes ‘native parallel thinking’ to bypass chain‑of‑thought bottlenecks

A new arXiv paper (ParaThinker) proposes training LLMs to generate multiple diverse reasoning paths in parallel and then synthesize them — a native parallel‑thinking paradigm that aims to avoid the ‘tunnel vision’ failure mode of long sequential chain‑of‑thought scaling. Summaries of the preprint report substantial accuracy gains on hard reasoning benchmarks (e.g., double‑digit % gains for smaller models) with only modest latency overhead, suggesting that allocating compute across parallel trajectories (width) can be a far more efficient route to stronger reasoning than simply lengthening single chains. Impact: if robust in wider settings, parallel inference architectures could shift how models are scaled and deployed for reasoning tasks, enabling smaller models to match larger ones for many applications.

Read more → Emergent Mind (paper summary / aggregator)

AppZen raises $180M Series D to scale agentic AI for enterprise finance

AppZen announced a $180 million growth round led by Riverwood Capital (reported Sept. 22, 2025). The company, which builds agentic AI that automates finance workflows (expense audits, AP, corporate card programs) will use the funding to scale its Mastermind AI Agents and expand go-to-market globally. The raise signals continued investor appetite for vertical, domain‑trained AI that replaces manual back‑office work and helps CFOs cut costs — a trend prompting incumbents and enterprise software buyers to accelerate AI automation projects. For the industry, the deal highlights investor focus on productionized, revenue‑generating AI applications rather than purely research model plays.

Microsoft launches Windows AI Labs to pilot experimental AI features in Paint, Notepad and File Explorer

Microsoft has started a Windows AI Labs pilot to give testers early access to experimental AI features across Windows — with Paint being the first app surfaced. The program is designed to accelerate feature validation and collect user feedback on usability and market fit for AI tools (image generation and Photoshop‑style editing, AI actions in File Explorer, smart features in Notepad, etc.). Why it matters: embedding experimental AI directly into core Windows apps speeds mainstream exposure to generative and assistant features, lets Microsoft iterate quickly on real‑world UX, and signals a broader push to make AI capabilities part of everyday desktop workflows.

Replit raises $250M (valued at $3B) and unveils Agent 3 — an autonomous tool that builds, tests and fixes apps

Replit announced a $250 million funding round that pushes its valuation to about $3 billion and said it has launched Agent 3, a more autonomous AI assistant for app building that can test apps, fix code, and compose custom agents and workflows. Why it matters: Agent 3 advances the 'vibe coding' / agentic app‑generation trend by automating end‑to‑end development tasks, lowering the barrier for non‑technical teams to create production apps and intensifying competition among code‑generation platforms. The round — led by Prysm Capital with participation from Google’s AI Futures Fund and Amex Ventures — will fund further R&D and go‑to‑market expansion.