06-23-Daily - AI Hot Daily
AI Hot Daily 2026/6/23
Daily curated AI + indie dev news
Today’s Summary
Sakana AI launches Fugu system,
Z.ai releases 1M token open-source model,
Boogu drawing model can write Chinese characters.AI Technology & Products
Sakana AI Releases Fugu ⭐ 9
Sakana AI has launched Fugu, a system that dynamically schedules multiple expert models using LLMs to complete complex tasks. This allows users to call through a single API endpoint, with a main model coordinating multiple expert models to collaborate, similar to OpenRouter’s Fusion feature. Fugu aims to improve performance through inter-model collaboration and is seen as a new signal for the direction of AI development.
GLM-5.2 Performance Leap ⭐ 8
Z.ai has released GLM-5.2, an open-source model with a 1M token context window. The model excels at long-context encoding tasks and rivals closed-source models. It is considered a significant advancement in the open-source model space and is expected to impact the existing market landscape.
Boogu Open-Source Drawing Model ⭐ 8
Boogu has released an open-source drawing model (version 0.1) that excels at accurately writing Chinese characters and possesses some reasoning capabilities. Although its aesthetic quality needs improvement, this model offers new possibilities for drawing applications requiring precise Chinese character output or reasoning abilities.
Moebius: 0.2B Image Inpainting Model ⭐ 8.5
Moebius is a 0.2B parameter image inpainting model that claims to achieve the performance of 10B models. Community discussions show successful browser demos based on ONNX, while some users remain skeptical about its performance matching 10B models, citing poor handling of novel objects and a 512x512 output limitation. The article also mentions its potential application in comic translation.
GLM 5.2 Introduces 1M Context ⭐ 8
Z.ai has released GLM 5.2, an open-source model with a 1M token context window and powerful long-context encoding capabilities. Developers can access it via API or download model weights for local deployment, quantization, and fine-tuning. Its low cost and high flexibility make it a highly sought-after option.
Anthropic Trains Stronger Model ⭐ 7.5
It is rumored that Anthropic has completed training a stronger model called Mythos, sparking discussions about new and old model iterations and releases. Previously, the Mythos model was “frozen” due to its excessive capabilities, raising concerns about safety risks.
AI Assistant PPT Generation Tutorial ⭐ 7
This is a detailed tutorial on creating PPTs with AI, covering the entire process from tool installation to final presentation. It includes content restructuring, design style determination, AI image generation, and animation effects, aiming to help users become experts in AI PPT creation.
AI Agent Safety Control Roadmap ⭐ 7
Google DeepMind has released a roadmap for AI agent safety controls, emphasizing the importance of system-level security and viewing AI agents as potential internal threats. The roadmap draws from cybersecurity practices, using monitoring, logging, and permissions to manage AI agent behavior.
Tallicom AI Browser Trick Contest ⭐ 7.5
Tallicom AI Browser is holding a “Trick Contest” encouraging users to create prompt or script tricks, with a top prize of 10,000 yuan. This event offers users a great opportunity to showcase and monetize their creativity.
AI News Daily ⭐ 7
An AI news daily that compiles information from June 22nd, with a Notion link provided for users to delve deeper.
AI Red Teaming and Security ⭐ 7
Gray Swan discusses AI security, particularly red teaming and model vulnerabilities. They argue that AI security differs from traditional cybersecurity and emphasize the importance of model control, agent safety, and preventing prompt injection. Anthropic’s Mythos model has raised concerns about export controls and security reviews due to its powerful capabilities.
AI Plugins Found to Contain Viruses ⭐ 6
Recently, AI plugins have been found to contain viruses. Attackers are pushing malicious plugins to the top of search engine results through paid ads. These plugins steal all conversations from users on AI chat platforms, and over 900,000 users may have been affected. Users are warned to be vigilant and check the safety of plugins to prevent personal information leakage.
Indie Development & SaaS
Vercel Supports WebSocket Beta ⭐ 8
Vercel Functions now support WebSocket connections, allowing bidirectional communication between clients and server-side code. This feature is suitable for real-time AI streaming, chat, and collaborative applications. Billing is based only on the time functions spend processing messages, not idle connection time.
Common Mistakes of Early-Stage Entrepreneurs ⭐ 8
The article points out that early-stage entrepreneurs often focus on features and neglect the importance of creating user stories. Intermediate entrepreneurs learn to tell stories, while advanced entrepreneurs tell stories first and then implement them with features. The article warns against scammers who only talk about stories without taking action.
Vercel Launches Platform-Native Feature Flags ⭐ 7
Vercel has launched Vercel Flags, a platform-native feature flagging service designed to simplify the development process and decouple code releases from feature rollouts. It integrates deeply with frameworks like Next.js, supports server-side evaluation for page performance, and allows management via a dashboard or CLI.
Indie Developer Creates Logic Puzzle Website ⭐ 6
A developer, tired of ads, built their own logic puzzle website, Puzzlelair. The site offers an ad-free puzzle experience and has sparked community discussions about AI-generated content and different puzzle engines.
Creating AI-Powered PPTs with Scratch ⭐ 7
A developer shares a method for creating AI-powered PPTs using Scratch, demonstrating how to generate content and implement dynamic effects with AI tools. This offers innovative PPT creation ideas for indie developers.
AI SEO Monetization and Investment System ⭐ 7
This system combines SEO traffic acquisition with Adsense or subscription monetization, and then invests the revenue in US stocks. Its core lies in using SEO to obtain virtually risk-free funds, which are then amplified through the stock market. It showcases a monetization and wealth growth model for indie developers.
Vercel Increases Project Limit Per Git Repo ⭐ 6
Vercel’s Hobby plan users can now connect up to 25 projects per repository, an increase from 10. This adjustment simplifies monorepo management and scenarios where multiple applications are deployed from a single codebase, providing convenience for indie developers managing and deploying multiple related projects.
Open Source Projects
sqlite-utils 4.0rc1 Released ⭐ 8
sqlite-utils version 4.0rc1 introduces database migration and nested transaction features. The migration functionality, inspired by the sqlite-migrate package, supports forward-only migrations for easy database change management. Nested transactions are provided via db.atomic(), simplifying SQLite transaction management.
Oak: A Git Alternative for Agents ⭐ 8
Oak is a version control system designed as a Git alternative for agents, aiming to improve their speed and contextual efficiency when handling complex projects. Through features like virtual mounting, it allows agents to process tasks in parallel, avoiding the hassle of downloading entire repositories or managing worktrees. Although still in early development, it has replaced Git in several months.
Patch the Planet Supports Open Source ⭐ 8
OpenAI has launched the Patch the Planet project to help open-source maintainers use AI and expert reviews to discover, verify, and fix security vulnerabilities, supporting the security and sustainability of the open-source community.
Expert-Aware Quantization Research ⭐ 7.5
Research proposes an “expert-aware quantization” technique that analyzes the frequency of model experts’ usage. It maintains high precision for “hot” experts and performs low-bit quantization for “cold” experts, achieving near Q4 model quality at approximately Q2 model size. This offers optimization ideas for running large models locally.
Industry News
Agentic Engineering Workflow ⭐ 9
A former Director Engineer at Meta/Microsoft/Atlassian shares an Agentic engineering workflow capable of submitting 40-50 production-level PRs per day. The core idea is to act as the captain, with agents forming the crew, collaborating through layered progression, involving terminal-centricity, agent memory and skill training, interactive reviews, and an “unattended” pipeline.
In-depth Analysis of Modern Browsers ⭐ 8
Addy Osmani’s in-depth article comprehensively guides through the internal mechanisms of modern browsers (primarily Chromium, with comparisons to Gecko and WebKit). Content covers the page loading pipeline, V8 JavaScript engine, multi-process architecture, security features, and provides practical optimization points for web developers, such as network optimization, structured HTML/CSS, and batched DOM updates.
Vercel Design Engineer Principles ⭐ 8
Vercel’s six design engineer principles emphasize delivering design through engineering capabilities, focusing on “usefulness” and “complete experience.” The principles include obsessing over usefulness, taking ownership of the complete experience, understanding constraints, designing for everyone, pursuing excellence, and making the team stronger. This represents an end-to-end ownership, low-ego, collaborative, and continuous improvement work style.
Past Experiences at Baidu US Research and AI Development ⭐ 8
An in-depth interview covering Baidu US Research, Scaling Law, OpenAI, and more. The content discusses Baidu’s early exploration of ultra-large models, its investment in Cerebras, and missed opportunities to be early angel investors in companies like OpenAI. The article also revisits former Baidu employees’ involvement in GPT-3 development at OpenAI and the contributions of Baidu US Research talent to AI startups.
AI Agent Control and Security ⭐ 7
The article discusses AI agent security, stating that AI agents need to be controlled like internal employees and emphasizing the importance of system-level security measures. The rise of AI agents shifts “chatbot security” to “workflow security,” requiring a combination of model evaluation, access control, and audit logs.
The Rise of AI Agents and Social Trends ⭐ 7
Data shows that US users spend significantly more time interacting with AI companions than dating apps, reflecting a demand for “unconditional” connection. The article suggests that AI development is changing social patterns and explores AI applications in work and life, such as DeepMind’s AI control roadmap and Anthropic’s improvements in robot task speed.
Cybersecurity for Business Travelers ⭐ 6
The article provides detailed cybersecurity advice for business travelers, including avoiding unknown USB devices, being cautious with public USB charging, and deleting unnecessary phone apps. It recommends using independent chargers, VPNs (with caution), and password managers, and emphasizes handling tempting information and using dedicated laptops.
Vercel CLI Supports Signing Blob URLs ⭐ 6
Vercel CLI can now directly generate signed URLs for Vercel Blob. Signed URLs allow setting expiration times and can be used for single operations, such as providing short-term access to files or allowing users to upload specific types and sizes of images. This feature is integrated into the CLI, making it convenient for developers to manage and control access to storage objects.
Social Media Buzz
Harry Zhang Open-Sources Agentic Book ⭐ 8.5
Harry Zhang (Hai Li) has open-sourced his third book, “Deep Agents in Action.” The book focuses on agent development, covering topics such as Agent Harness planning, context engineering, sub-agents, skills, and memory. This is a valuable learning resource for developers interested in agent development.
WeChat AI Assistant Does Not Support Markdown ⭐ 6
The author expresses dissatisfaction with WeChat’s “Xiao Wei” AI assistant not supporting Markdown format, viewing it as a disregard for open standards and criticizing WeChat’s product strategy. However, the author also notes that this may not be important for the target user group (e.g., elderly users) and does not change their impression of WeChat as “arrogant” and “old-fashioned.”
Potential Risks of Agreeing with Opinions ⭐ 6
The author believes that when an opinion receives a large number of likes, it may simply be repeating common consensus, leading to complacency and overlooking potential issues. The author implies that sometimes, being “unknown” is better.