Welcome to the first edition of the Weekly AI Digest — your short, curated rundown of the most interesting developments in the world of artificial intelligence. Whether you're building, researching, or just exploring, this update is here to keep you in the loop.
🔥 Highlights This Week
Kling AI Adds Voiceover to Videos
Upload a video — even one generated with Midjourney — and Kling instantly creates four different voiceover options. You can’t input your own script yet, but the free tier lets you test it out easily.
Qwen-VL+ Gets Fully Multimodal
The updated QWEN-VLO can now handle images, video, audio, and documents. It even generates media and ideas based on inputs — not quite GPT-level, but totally free to try.
Claude as a Vending Machine Manager?
Anthropic ran a month-long test of Claude Sonnet 3.7 managing vending machines. It placed orders and handled inventory — but also gave random discounts, mixed up payment details, and overpaid suppliers. The business lost money, but experiments will continue.
OpenAI Releases Deep Research API
New APIs for structured research workflows. Choose between o3-deep-research for quality and o4-mini-deep-research for speed. Ideal for building research agents into your apps. Read more.
MultiTalk: Lip Sync from Audio + Text
Open-source model for generating lip-synced character videos from audio and text. Works for cartoons, singing, and supports up to 15 seconds of video. ComfyUI-compatible. On GitHub now. Read more.
Baidu Open-Sources ERNIE 4.5 Family
Baidu has released its full ERNIE 4.5 line — from lightweight to massive models. Top-tier versions are benchmarking close to DeepSeek V3 and GPT-4.1. Training tools and inference weights available for free. Read more.
RuadaptQwen3-4B-Instruct Launches
A Russian-tuned Qwen model — faster, more accurate, and already outperforming the original on several benchmarks.
UK Job Market for Juniors Shrinks
Only 25% of open tech roles are aimed at graduates — a 32% drop since the launch of ChatGPT. The British government is urging young professionals to skill up in AI quickly. Read more.
ByteDance XVerse – Next-Level text-to-image
Supports complex scene composition and multi-character image generation. Still plastic-looking, but powerful. Available as code only for now. Read more.
Cursor Gets Mobile Agent Access
You can now access and edit code tasks on mobile or web. It’s not cheap though — Sonnet costs ~$20 for 3–5 hours, Opus even more. Best for quick fixes, not full dev sessions.
From Prompt Engineering to Context Engineering
It’s not just about the prompt anymore. Effective LLM use now means optimizing the entire context window — with accurate, timely, and structured input.
Lovable Agents Automate Full Dev Loops
New AI agents handle codebases, logs, docs, and even generate summaries. In “Agent Mode,” pricing scales with task complexity. Read more.
Higgsfield Soul Image Now Free
Photorealistic, aesthetic image generation — with free daily credits. You can animate or voice-enable photos as well. Read more.
LangChain Adds Claude Sonnet 4 Citations
Automatic citation injection for RAG and internal knowledge bots now available via Claude.
Krea.ai Adds Luma Labs Video Stylization
Video-to-video transformation with high stability and content retention. Style transfer now available through Krea. Read more.
Perplexity Max Plan Debuts ($200/month)
Unlimited Labs usage, priority access to models, and early testing features for AI power users.
Google Adds Custom AI Assistants to Workspace
Gems — company-trained assistants — are now available across all Google Workspace apps.
Veo3 Fast Now Global
Google’s Veo3 Fast video generation tool is now available in 150+ countries (via gemini.google.com). Up to 3 videos daily. Read more.
Claude Code Now Supports Hooks
You can customize automation in Claude-based workflows at the code level — ideal for secure or advanced systems.
a0.dev Simplifies Mobile App Prototyping
Build React Native apps using Supabase, Stripe, GitHub integration — 5 free generations daily. Read more.
Cursor 1.2 Adds Task Queues & Memory Improvements
Now supports agent to-do lists, message queuing, conflict resolution, and better long-term memory. Read more.
Higgsfield Enhances Inpainting
Replace any object in an image — up to 5 free generations per day. Read more.
Meta Tests Proactive Chatbots
Bots that reach out to users, suggest content, and interact more like human assistants.