Chokri Boubakri
Chokri Boubakri is the editor of The Signal. An AI automation strategist and e-commerce growth specialist based in Sfax, Tunisia, he builds AI-powered SaaS tools and runs performance campaigns for clients across the Maghreb, the Gulf and Europe. The Signal is where he files what he finds about open models, local AI setups and the tools built around them.
Jaaz: An Open-Source, Local-First Canvas for AI Images and Video
Jaaz is an open-source Canva AI alternative and local-first creative canvas for AI images and video, with an AI agent that can use local ComfyUI or cloud models. License unverified.
ComfyUI-LTXVideo: Lightricks’ Extra Nodes for LTX-2 Video
ComfyUI-LTXVideo adds LTX-2 custom nodes and ComfyUI video workflows from Lightricks, including example workflows, control LoRAs, HDR output and dubbing tools. Requires a 32GB+ GPU. License unverified.
Wan 3.0: Alibaba’s New AI Video Model
TL;DR: AI video is moving fast, and Alibaba clearly does not intend to watch from the sidelines. With Wan 3.0, […]
The Mystery Model That Turned Out to Be Z.ai’s GLM-5.3-Flash
For six days, an anonymous model called 'ox-alpha' ran free on OpenRouter and OpenCode with no declared owner. Then Z.ai confirmed it: the mystery model was GLM-5.3-Flash all along.
Study: Every Model-Based Prompt-Injection Defense Eventually Breaks
A 20,000-attack evaluation across GPT-5.4, Gemini 2.5 Pro, and Claude Sonnet 4.6 found only one defense held: filtering the model's output in separate application code.
Microsoft and Google Both Roll Out New AI Pricing Tiers on September 1
Microsoft Foundry adds regional pricing premiums and Google's Gemini Enterprise Agent Platform starts billing Sessions and Memory Bank usage — both effective the same day.
What AP’s Updated AI Newsroom Standards Actually Permit
AP revised its AI guidelines in July 2026, adding rules for coding assistants and a disclosure threshold. Here's exactly what changed, sourced from AP's own announcement.
GitHub’s HydraFusion Picks a Different AI Model for Every Coding Task, Automatically
GitHub's new Copilot CLI research preview builds a custom workflow per task instead of running everything through one model, claiming a 67% cost cut on one benchmark with quality close to Claude Opus 5.
DeepSeek Ships an Experimental Vision Model, Claims It Rivals Opus 4.8 on Agent Tasks
DeepSeek quietly released DeepSeek-V4-Flash-Vision-Exp, a multimodal variant the company says closes the gap with Anthropic's Opus 4.8 on agent benchmarks that require visual understanding.