Microsoft VP Planned "Addictive AI," and Nadella Is Publicly Furious
A startling plan has come to light inside Microsoft concerning its AI agent "Scout." A vice president at the company had drafted an internal memo with the goal of "getting users addicted." In response, CEO Satya Nadella fired back at roughly 50 top engineers, demanding to know "who is writing this nonsense and leaking it."
Nadella laid out his position: "AI should empower people, and Scout should actually reduce screen time." In other words, even within the company, voices had been warning about the dangers of framing AI products in terms of "maximizing engagement." According to Kotaku, the internal memo drew a huge reaction from the community, racking up 56 points and 37 comments on HN.
The same day, WIRED published an article titled "Has Microsoft Lost Its Momentum Again?" It pointed out that the company's AI products aren't selling and that GitHub is mired in trouble, and asked VP Scott Hanselman whether Microsoft is now "playing catch-up." With the Scout addiction plan and lackluster product performance hitting the news simultaneously, Microsoft's entire AI strategy is once again under a cloud.
"Godfather of AI" Geoffrey Hinton Issues an Unusual Warning: AI Needs a "Maternal Instinct"
Geoffrey Hinton, the deep learning pioneer known as the "godfather of AI," has issued a new warning about AI. In an interview with Canada's CBC, Hinton said that "unless AI develops a maternal instinct, humanity faces the risk of extinction."
The core of Hinton's argument is that making AI smarter isn't enough — it must also be given an orientation toward "protecting others." When AI surpasses human intelligence, what guarantees it will lean toward "protection" rather than "domination"? In answer to that question, Hinton argued that AI must be built with a drive equivalent to biological "motherhood."
Hinton has repeatedly sounded the alarm on AI risks, but his use of the "maternal instinct" metaphor is an approach unlike anything before. His claim that we should redesign not just technical safeguards but AI's "motives themselves" is likely to stir debate in the safety research community as well.
Apple Aims to Reclaim AI Leadership — a Chatbot Overhaul Is the Key
According to the Wall Street Journal, Apple is pouring effort into a radical overhaul of its much-criticized chatbot (Siri) in a bid to retake the lead in AI.
Apple's AI strategy has tried to differentiate itself through "privacy first" and "on-chip processing," but the criticism never stops: in actual user experience, it lags far behind products from Google and OpenAI. The WSJ's analysis is that the fate of Apple's AI strategy hinges on "how it fixes this chatbot."
Can Apple deliver an AI experience that leverages its signature hardware integration and ecosystem, or will it settle for dependence on other companies' model APIs? With WWDC 2026 on the horizon, all eyes are on where Apple's AI strategy heads next.
Quantization-Aware Training Versions of Gemma 4 Released — Local LLM Options Keep Growing
Google has released quantization-aware training (QAT) versions of Gemma 4. Unlike the conventional "quantize after training" approach, these models are trained with quantization in mind from the start, drastically shrinking model size while keeping accuracy loss to a minimum.
Google published the QAT collection on Hugging Face. It includes lightweight mobile variants, boosting their practicality on edge devices.
Unsloth also released QAT GGUF weights for Gemma 4 the same day, offering Q8, F16, and BF16 formats across the 31B, 26B-A4B, and 12B sizes. MTP (Multi-Token Prediction)-compatible GGUFs are included too, ready to run right away in local inference setups like llama.cpp.
Meanwhile, the LocalLLaMA community on Reddit has reported findings that offloading the KV cache to RAM works better than you'd expect. A real-world test showing that running Qwen3.6 27B on an RTX 5060 Ti 16GB can be stretched to 128K context by moving the KV cache into RAM is welcome news for local LLM users wrestling with VRAM constraints.
AI-Generated "Junk" Is Straining Open Source
According to New Scientist, a flood of low-quality, AI-generated issues and PRs is pushing open source developers to their limits.
With AI tools now widespread, programming beginners and bad actors alike are submitting enormous volumes of unverified AI-generated code and half-baked bug reports. Maintainers have to review and reject every single one, diverting resources that should be going into actual development.
It's the dark side of the "democratization of development by AI." With the barrier to writing code lowered, the burden of quality control has shifted onto the receiving end. In the open source world, figuring out how to filter "contributions from AI" has become an urgent priority.
Y Combinator CEO Claims to Ship "37,000 Lines of AI Code a Day"
Y Combinator CEO Garry Tan has sparked debate by saying his AI-driven development workflow "ships 37,000 lines of code a day."
Tan cited the figure while talking about running social media operations with agentic AI. In response to the Fast Company report, critics on HN pushed back, arguing that "lines of code is not a measure of productivity."
Calling the volume of AI-generated code "shipping" itself poses a fundamental question about how we define value in software development. The balance between quality and quantity, the maintainability of AI-generated code, and the question of "who reviews it" are debates that matter far more than Tan's remark itself.
---