AI just made three bold moves reshaping privacy, transparency, and open search. From browsers giving users a hard AI off-switch to tools that literally open up model brains and private Perplexity-style search you can self-host, here’s what’s new:

🌐 Firefox AI Kill Switch — Control Stays with the User
After privacy backlash, Mozilla confirms a real, permanent AI kill switch coming to Firefox—letting users fully disable all AI features, no prompts, no grey areas. AI stays optional, not forced.

🧠 Gemma Scope 2 — X-Ray Vision for AI Models
DeepMind drops the largest open-source interpretability toolkit ever, decoding how Gemma models reason, hallucinate, or jailbreak, giving researchers unprecedented visibility into AI internals.

🔍 Perplexica — Private, Self-Hosted AI Search
An open-source Perplexity alternative that runs on your own machine, delivers cited answers, supports multiple LLMs, and keeps your data private—no black-box search, no data leaks.

AI isn’t just getting smarter, it’s getting more transparent, more private, and more user-controlled.

Epic Firefox AI Kill Switch Reveal

Firefox faces backlash after Mozilla CEO Anthony Enzor-DeMeo's December 16 blog post hinted at heavy AI integration, including an "AI window," sparking fears among privacy-focused users who favor the browser as a Chrome alternative. Enzor-DeMeo responded on Reddit (Dec 17) promising a "real kill switch" in Q1 2026 for full user control, emphasizing AI as a choice—though critics argue features should be strictly opt-in (off by default) rather than opt-out. Firefox for Web Developers (@firefoxwebdevs) clarified on Mastodon (Dec 18) that all AI is opt-in with an internal "AI kill switch" to permanently disable everything, no future prompts, addressing "grey areas" like toolbar buttons. Jake (@jaffathecake) urged the technical community not to prejudge, seeking trust amid critiques of Mozilla's trend-chasing (e.g., FirefoxOS, brief 2022 crypto donations). A Vice article highlights Firefox's Gecko engine and customization legacy, but users demand explicit opt-in.

Gemma Scope 2: DeepMind's Massive AI Interpretability Toolkit

Google DeepMind released Gemma Scope 2 on December 18, 2025, an expansive open-source toolkit for interpreting the Gemma 3 family of language models, spanning 270M to 27B parameters—the largest such release by an AI lab, involving 110 petabytes of data and over 1 trillion trained parameters. Building on the original Gemma Scope, it uses sparse autoencoders (SAEs) with JumpReLU and Matryoshka training, plus skip- and cross-layer transcoders, to decode internal "features" across all layers, revealing multi-step reasoning, hallucinations, jailbreaks, sycophancy, and chain-of-thought faithfulness in chat-tuned models. Available on Hugging Face with notebooks, APIs for JAX/PyTorch, and a Neuronpedia demo, it enables safety researchers to audit biases, debug agents, and trace emergent behaviors like those in scaled models discovering cancer pathways, fostering robust interventions for reliable AI.

Perplexica: Open-Source Perplexity AI Alternative

Perplexica is a privacy-focused, self-hosted AI answering engine mimicking Perplexity AI, running on your hardware with local LLMs via Ollama or cloud providers like OpenAI, Claude, Gemini, and Groq—delivering cited answers from SearxNG web searches while keeping data private. Key features include smart modes (Balanced, Fast, Quality soon), six focus modes (Academic, YouTube, Reddit, Wolfram, Writing, Web), image/video/file search, domain-specific queries, smart suggestions, Discover feed, and local search history; upcoming: Tavily/Exa support. Docker deployment is simplest: docker run -d -p 3000:3000 -v perplexica-data:/home/perplexica/data -v perplexica-uploads:/home/perplexica/uploads --name perplexica itzcrazykns1337/perplexica:latest, accessing http://localhost:3000 for setup—no extra SearxNG needed. Boasts 27.6k GitHub stars, MIT license, API access, browser search integration, sponsors like Warp/Exa; troubleshooting covers Ollama/Lemonade connections.

Hand Picked Video

BG Blur transformed from a basic blur tool into a powerful AI video enhancer with a minimal design, smarter background blur, privacy upgrades, and live previews for creators.​

Top AI Products from this week

• ConnectMachine - ConnectMachine is a privacy-first AI Agent that elevates how you connect. Craft bespoke digital business cards, share only what you choose, and explore your network with intuitive voice or text queries.

• Flowdrop V1.1 - Flowdrop turns chat into automation. Just tell us what you want. Our AI builds it for you. Connect Gmail, Slack, Sheets, Telegram, X, LinkedIn and more. Zero code. NEW THIS WEEK: Human in the loop forms so you can approve steps before they run. See changes live. Keep what works. YouTube video to text.

• Whisper Snapper for Mac - Transcribe videos, podcasts, meetings, and voice memos with the fastest, most accurate AI models—running locally on your Mac or in the cloud with your own API keys. Speaker labels, timestamps, and export to SRT, Markdown, PDF, and more. Try free or one-time lifetime $9.99 Pro.

• wafer - Wafer is the GPU development stack that lives inside your editor: profiling, compiler explorer, and GPU docs—all in one place. Writing GPU kernels means jumping between fragmented tools: editing here, profiling there, docs in a browser, compiler explorer in another tab. Wafer pulls everything back into your IDE.

• ​OverseerOS - Paste any YouTube channel link and OverseerOS turns it into a growth blueprint you can actually follow. It decodes what works across the channel’s top videos, X rays viral uploads (pacing, title patterns, thumbnails), and helps you spot rising channels in any niche.

• Kirkify - Kirkify is an AI meme studio that turns photos into bold satirical edits in seconds. Exports come with a built-in satire/AI disclosure label + watermark by default, plus ready-to-post sizes for social.

This week in AI

• Cursor Acquires Graphite - Cursor acquires Graphite (Dec 19, 2025), a code review platform for 100k+ engineers, to merge AI coding with PR workflows. Graphite runs independently; future integrations promise smarter reviews and dev tools.

• Solaris Unveiled - MASK Architects debuted Solaris, the world's first fully solar-powered autonomous motorcycle. Its 360° curved photovoltaics, foldable wings, and hybrid energy system enable self-charging without fuel or grids. Lightweight aerospace materials boost efficiency.

• Unitree's Robot App Store - China's Unitree unveils world's first humanoid robot "app store" platform. Users control bots via smartphones, share datasets, demos like Bruce Lee moves & ballet. Aiming embodied AI in daily life.

• LORE Laws of Reasoning - LORE formalizes LRM behaviors: compute scales linearly with complexity (monotonicity, compositionality). LRMs satisfy monotonicity but fail compositionality.

• MMRAG-RFT Framework - Two-stage RL fine-tuning for explainable multi-modal RAG: Stage1 rule-based point-wise doc ranking filters irrelevants; Stage2 reasoning-based list-wise ranking + answering. SOTA on WebQA/MultimodalQA.

Paper of The Day

UniRel-R1 introduces relation-centric KGQA, where answers are informative subgraphs linking seed entities, unlike entity-centric methods. The framework combines k-hop subgraph selection, multi-stage pruning (using hub penalties to remove generic nodes), and RL-tuned LLMs (via GRPO with rewards for format, connectivity, entity/relation informativeness). Experiments on 7 KGs show 35%+ connectivity gains and 245% reward boosts over baselines; Llama models generalize better sans semantics.

To read the whole paper 👉️ here