Apple's Camera Lead Discusses iOS 27's AI Photo Features

In an interview with WIRED, Jon McCormack, Apple's head of camera software, talked about the generative AI features built into the new Photos app in iOS 27. While the feature adds "fake pixels" to photos, McCormack emphasized that it is not "AI for AI's sake" — the goal is to give users genuine "superpowers."

Apple's full-fledged adoption of generative AI in photo editing signals that the company's AI strategy is steadily spreading into practical domains. Apple's AI features have been rolled out incrementally so far, but applying them to photography — an everyday use case — may be the first time many users become consciously aware of AI at work.

Uncensored Versions of Gemma 4 Released — Four Models at Once

On Reddit's LocalLLaMA community, user LLMFan46 released uncensored versions of Gemma 4 (Uncensored Heretics) — four models all at once. The lineup is as follows:

gemma-4-31B-it-qat-q4_0-uncensored-heretic** (QAT quantization variant)

gemma-4-26B-A4B-it-qat-q4_0-uncensored-heretic** (MoE QAT variant)

gemma-4-12B-it-qat-q4_0-uncensored-heretic** (smaller QAT variant)

gemma-4-12B-it-uncensored-heretic** (non-QAT variant)

Each model is available in multiple formats, including Safetensors, GGUF, NVFP4, and GPTQ-Int4, covering a range of hardware environments. Community-driven releases of uncensored models are a testament to how vibrant the open-source LLM ecosystem has become.

llama.cpp Thread Optimization Delivers Up to 80% Performance Gains

Also on LocalLLaMA, user AXYZE8 reported substantial performance improvements from tuning llama.cpp's `--threads` argument. Testing the Gemma 4 26B A4B QAT model on an Intel 250K Plus (18 cores: 6 P-cores + 12 E-cores) produced the following results:

threads (P-cores only): average 49.1 tok/s

threads: average 88.8 tok/s (+80%)**

threads (all cores): average 66.5 tok/s

The conventional wisdom for hybrid CPUs has long been "stick to the P-cores only," but on Arrow Lake-generation chips, increasing the thread count can actually pay off. That said, 18 threads (using all cores) degraded performance instead, showing that the optimal thread count varies significantly depending on the CPU architecture. Running your own benchmarks in your own environment matters.

A Method for Directly Probing LLM Hidden States

A technical article that drew attention on Hacker News, "Don't let the LLM speak, just probe it," explains a technique for extracting information by directly probing an LLM's internal representations (hidden states).

The conventional approach has been to have the LLM generate text and then analyze the output. By probing hidden states directly, however, it may be possible to determine whether the model "knows the answer" before any text is generated. This opens up new approaches to hallucination detection and model reliability assessment.

"Europe 2031" — What Happens If AI Policy Goes Wrong

"Europe 2031," another project that made waves on Hacker News, explores what happens if Europe gets its AI policy wrong. It presents concrete scenarios covering the impact of the EU's AI regulation (the AI Act), the risk of regulation stifling innovation, and the challenge of staying competitive with the United States and China.

European AI regulation is a pioneering effort on the world stage, but striking the right balance between regulation and innovation remains a major challenge. This project offers a valuable perspective for policymakers.

A Wave of New AI Agent Tools

Several open-source AI agent tools have been announced.

**Ciris** is an open-source AI agent supporting 29 languages on both iOS and Android. A full-fledged AI agent running in mobile environments is a symbolic release that demonstrates the practicality of edge AI.

**OpenDream** is a local-first memory system for AI agents, notable for its "dreaming" feature. It provides a mechanism for agents to organize and consolidate their memories.

**AgentForge** is a framework that adds "28 production-grade skills" to AI agents, aiming to improve the reliability of code generation.

Japanese Community Discussions — The Engineer's Role in the AI Era

On Qiita, debates about what engineers should be in the AI era are thriving. An article titled "The people who do the work nobody else did will become the key players of the next era" argued that the very premise that "everyone can use AI" leads to team dysfunction. The view that skill gaps emerge in how well people use AI — and that those gaps determine an organization's productivity — is a refreshingly practical one.

Another article, "What Makes a 'Capable Engineer' in the AI Era," voices anxiety about degenerating into a mere AI operator. Rather than fearing replacement by AI, the argument goes, we need to think about what the core skills for truly mastering AI really are.

---