Microsoft Overhauls Copilot and Enters the 'AI Super App' Race with AutoPilot Agents

Microsoft is reportedly planning a major Copilot app overhaul in August. The company will unify the consumer and enterprise Copilot apps, which had previously been split, cut low-usage features such as Copilot Podcasts, and introduce a new AI agent called 'AutoPilot.'

AutoPilot is an agent that autonomously handles tasks in the background and will be offered for an additional fee. As Anthropic and OpenAI successively build comprehensive AI platforms centered on chat, Microsoft is set to join the race by turning Copilot into an 'AI super app.' Copilot has long been criticized for its fragmented features and inconsistent user experience, so how much the consolidation actually improves the experience will be worth watching.

Meta AI Lead Claims Its Next LLM Has Caught Up to OpenAI's Flagship

The head of Meta's AI division has indicated that the company's next-generation large language model has caught up with the performance of OpenAI's flagship model (believed to be GPT-5). CEO Mark Zuckerberg had previously acknowledged falling behind in the AI agent space, but the new message is that the gap has closed — at least in terms of model performance.

That said, this is the executive's own subjective assessment, and no independent benchmark has verified it yet. It's also unclear which axis of evaluation 'caught up' refers to (reasoning, coding, multilingual support, etc.), so we'll need to wait for third-party verification.

Meituan's 'LongCat 2.0' Weights Released

The model weights of 'LongCat 2.0,' an LLM developed by Chinese tech company Meituan, have been made public. INT8 quantized and FP8 versions are both distributed on Hugging Face. The LongCat series has existed since the previous generation, but with 2.0 delivering improved performance plus officially provided quantized variants, this is good news for developers considering local inference.

While the pool of open-source LLM options keeps widening, checking the license terms and conditions of use remains a prerequisite.

SwiReasoning Draws Attention — A Technique That Boosts Accuracy While Cutting Token Consumption

A technique called 'SwiReasoning' is generating buzz on Reddit's r/LocalLLaMA. Users who tried it with Qwen 3.6 27B report that answer accuracy improved while the number of required tokens dropped significantly. Although tokens per second (t/s) itself dips somewhat, the total token count shrinks, so the perceived speed actually improves.

What's interesting is that the technique itself was unveiled nine months ago, yet it never gained wide adoption until now. It's a fascinating case of an inference-efficiency method reclaiming the spotlight after being evaluated in real-world settings. The implementation is public on GitHub, and a llama.cpp wrapper also exists.

Collabora Office 26.04: The Decision to Keep AI Features Optional, Not Forced

Collabora Office 26.04, the latest release of the LibreOffice-derived office suite, is out. What stands out is the design decision to keep AI features strictly optional rather than pushing them front and center. The release centers on improvements to Writer and Calc, with AI available only for users who choose to enable it.

As the push to integrate AI into every piece of software accelerates, this policy of respecting the workflows of users who don't use AI deserves real credit. Guaranteeing 'the option not to use AI' is an important stance for preserving diversity in productivity software.

---