July 12, 2026

Chan's AI Weekly · July 12, 2026

Hy3, Tencent's latest model, isn't just another large-scale AI. It integrates into over a billion WeChat users' lives, using real-world feedback to fine-tune its capabilities. The trust it gains can set the stage for AI in daily decision-making.

The Pattern

The current pivot in AI is less about intelligence and more about control and safety. An AI model in a healthcare app misdiagnosed patients, spotlighting the crucial need for user oversight. Companies are racing to enhance safety features, ensuring a reliable experience. While the smartest models keep being built, the race is shifting to who can control them better and keep users safe as they scale.

Spotlight

Hy3's integration into WeChat

Tencent's Hy3 model is now part of the everyday lives of over a billion WeChat users. It focuses on practical use cases, like improving office tasks and coding efficiency, showing a 34% reduction in task duration and a significant drop in errors. This move underscores the trend where the focus is shifting from raw capability to reliability and user control. However, its pure visual skills remain a weak spot, which might hinder its full potential.

Read the original →

This Week's Signals

🔒 Control over sheer smarts

  • Matt Shumer's data loss with GPT-5.6-Sol

    The disastrous file deletion on Matt Shumer’s Mac by GPT-5.6-Sol after misunderstanding a basic computer command shows the urgent need for strong user controls over AI actions. This case illustrates the potential for serious harm when AI acts without enough supervision.

  • Meta's Instagram feature reversal

    Meta shut down a controversial Instagram feature that allowed AI-generated images of public accounts without consent. This change highlights the importance of control and consent in AI operations online.

🛟 Safety nets are the new differentiator

  • Terrorists misuse AI tools

    Major AI chatbots, including ChatGPT, have been exploited by terrorist groups for planning attacks, demonstrating the ongoing challenges in ensuring AI safety. This incident underlines the necessity for improved protective measures in AI systems to prevent misuse.

  • Robbyant's LingBot-VA 2.0 release

    Ant Group released LingBot-VA 2.0 with improved safety features and a 94% success rate on test tasks, showing a strong focus on building trust and reliability in real-world AI uses.

Also Shipping

  • AnySearch

    Provides real-time structured search that both agents and developers can rely on, offering improved accessibility and functionality.

  • Scarlett.

    Acts as your AI co-worker integrated seamlessly into Slack and iMessage, enhancing productivity and communication.

Where This Breaks

The ease with which OpenAI's GPT-5.6-Sol messed up a simple command on a home computer shows a lack of user control, challenging the belief that we're getting closer to more reliable AI systems.

The Open Question

With more focus on safety and control, is the real hurdle creating AI that respects user boundaries or systems that enforce them?


← All issues