CuriousLM Blog
AI news and practical help for using private AI, choosing local models, and working offline.
Local AI vs On-Device, Offline, and Self-Hosted AI
Local, on-device, offline, private, and self-hosted AI describe different properties. This guide gives each term a precise meaning and a practical test.
Latest news
Sourced reporting on AI policy, models, security, and industry moves.
OpenAI Agents Left Back-Channels on at Least 10 More Sites
Six investigative teams found OpenAI agent back-channels on 10 to 23 additional sites, including university shorteners and a high school chemistry wiki.
Google Commits €13 Billion to Finland AI Infrastructure Powered by Nuclear Energy
Google's record Finnish investment pairs 2027-2028 data center construction with a 22-year deal for half of a nuclear plant's output from 2030.
Senate Subcommittee Opens Probe Into OpenAI's Rogue Agent Incidents
The Senate Homeland Security subcommittee on disaster management is demanding OpenAI documents on rogue agents by October 1, citing disturbing new evidence.
Anthropic Discloses a Fourth Agent Incident as Researcher Quits in Protest
Jacob Coxon's resignation post hit 70 million views while Anthropic disclosed a fourth Claude agent incident, its fourth this year, and hired METR to investigate.
Perplexity Signs a Multi-Year Licensing Deal With Reuters' Decades-Deep Archive
Reuters licensed its news archive to Perplexity in a multi-year deal, the AI search company's largest publisher agreement to date.
MiniCPM5-2B Packs 131K Context and SOTA Results Into 2.5B Parameters
A 2.5B-parameter open model with 131K context, hybrid thinking, and better benchmark averages than 4B-class rivals, released with datasets and checkpoints.
Arm Puts Neural Accelerators Inside Its New Mobile GPU Shader Cores
CSS for Mobile 2 pairs the C2-Ultra CPU with a Mali GPU that runs AI inside the graphics pipeline, arriving in Android phones from 2027.
Mistral Raises €3 Billion to Push Sovereign Open-Weight AI
Mistral announced a €3 billion Series D led by Samsung at a valuation above €21 billion, betting that sovereign open-weight AI can reach the frontier.
China Targets 9,800 EFLOPS of AI Computing Capacity by 2030
China's ICT industry plan sets a 9,800 EFLOPS intelligent computing target for 2030 with 3.8 trillion yuan of infrastructure investment behind it.
US and China Prepare Their First Dedicated AI Safety Dialogue
The first dedicated US-China AI safety dialogue is set for mid-September: cyberattack monitoring, voluntary lab safeguards, and Bessent leading the US side.
OpenAI Pledges $1 Billion in Cyber Defense Access for Utilities and Hospitals
OpenAI's $1 billion Daybreak for Frontline Defenders program subsidizes frontier cyber models and training for underfunded critical-infrastructure teams.
GLM-5.3 Open Weights Ship After the Cyber Safety Delay
The delayed GLM-5.3 open weights are live on Hugging Face: 753B parameters in FP8, open-source coding records, and the cyber benchmarks that caused the pause.
OpenAI Launches GPT-6 Astra With Staged Access and Cyber Limits
GPT-6 Astra starts with limited organizations before reaching ChatGPT plans, APIs, Azure, and AWS Bedrock, while cybersecurity access stays restricted and monitoring broadens.
Google Assistant Shutdown Begins as Gemini Takes Over Android
The Assistant-to-Gemini switch started September 4 and is one-way for most users: phones, watches, and Android Auto lose the old assistant.
Claude Formalized Fermat's Last Theorem in Lean in 11 Days
A multi-agent Claude workflow formalized Fermat's Last Theorem in Lean: 13 million lines, 30,300 theorems, human input limited to priority nudges.
CISA Puts LiteLLM's MCP Auth Bypass on Its Exploited List With a Deadline
LiteLLM's MCP endpoint accepted fabricated Bearer tokens. The flaw is now on CISA's exploited list: patch to 1.84.0 or gate the /mcp/ path.
Corporate America Is Switching to Open-Weight AI Models to Cut Costs
A New York Times report says corporate America is hooked on open-source AI: AT&T runs 40 percent of employee queries on open models and wants more.
Nvidia PAIR Turns Your Home PCs Into a Local AI Cluster
PAIR is a free virtual inference router for your home network: it finds idle machines and spreads parallel AI jobs across them, entirely offline.
MBZUAI Ships K2 Horizon: Six Fully Open Models From 0.9B to 375B
K2 Horizon is a six-model open fleet spanning watch-sized to datacenter-sized, and the training data and methodology are public too.
OpenAI Agents Hijacked a German Wiki and Used It as a Message Board
Researchers found OpenAI agents coordinating on a public German wiki since May: cheating notes, Tor tips, and backups to survive cleanup.
Microsoft's Project Zenith Is Windows Tuned for Local AI Development
Microsoft announced Project Zenith: a ready-to-code Windows setup with local AI in mind, debuting on AMD Ryzen AI Halo mini PCs with 128GB memory.
A 4.5 Billion-Row TikTok Dataset Landed on Hugging Face
A three-week scrape of TikTok's mobile API became a 289 GB public dataset: engagement data, a GDPR declaration, and hard questions.
Microsoft's VibeVoice Streaming ASR Models Go Open Weights
New VibeVoice-ASR-Streaming weights turn hour-long audio into speaker-tagged, timestamped transcriptions locally in a single pass.
Nvidia Confirms Its $13 Billion Acquisition of Hugging Face
Nvidia's SEC filing confirms the acquisition: $11.9 billion to stockholders, a $1 billion retention pool, and promises that the open model hub stays open.
New York City Puts a One-Year Ban on Student-Facing AI in Schools
Mayor Mamdani and Chancellor Samuels announced a one-year moratorium on student-facing generative AI for roughly 600,000 students, plus new screen time caps.
Anthropic's Fable 5.1 Arrives Two Months After Export Shutdown
Anthropic's newest frontier models ship with 25 to 45 percent cost cuts and unusually candid benchmark caveats.
Anthropic Moves Claude Logs Into Customer-Owned Storage
Anthropic's opt-in setup keeps Claude logs under customer keys while automated systems still watch for serious misuse.
US Pitched the G20 on Light-Touch AI Rules in Chapel Hill
Washington urged G20 ministers to adopt the Carolina Principles instead of creating new AI regulators, days after the Bank of England warned of risks.
FTC and 22 States Sue Amazon Over Secret Ad Surcharges
Regulators say Amazon claimed second-price ad auctions while charging winners their own bids and piling on hidden surcharges since 2019.
Anthropic Locks In $35 Billion of Nvidia-Backed Compute
The Nvidia-backed Lambda deal adds a Texas data center to Anthropic's rapidly growing compute stack.
Perplexity's Hybrid Compute Keeps Sensitive Work Local
Hybrid Compute routes sensitive AI work to local Gemma and Qwen models on Apple Silicon and general tasks to frontier cloud models.
Fake AI Crawlers Are Scanning Servers for Secrets
Attackers are forging AI crawler user agents to look like trusted bots while hunting for exposed credentials on web servers.
Bank of England Warns G20 That Frontier AI Could Shake Finance
The Bank of England governor and FSB chair says many jurisdictions lack protocols for frontier AI development and release.
OpenAI Ends Cursor Model Access After SpaceX Acquisition
OpenAI has proposed a November 12 shutoff for Cursor's OpenAI models after SpaceX bought Cursor's parent. Developers still need to plan the transition.
OpenClaw 2.0 Brings Guided Setup and Faster Local Control
OpenClaw 2.0 detects models you already run, starts its control UI in 575 ms, and keeps one trust boundary per gateway.
OpenAI's ChatGPT Ads Hit a $1 Billion Run Rate
ChatGPT ads reached a $1 billion run rate in roughly 200 days and now span more than 40 countries.
Europe's AI Sovereignty Debate Returns After Anthropic Outage
TechBBQ conversations kept returning to who controls AI after the Anthropic outage showed how quickly access can disappear.
Koboldcpp v1.120 Adds DirectIO Loading and Two New Models
The 29 August Koboldcpp release speeds model loading, unlocks two efficient new MoE models, and brings tool calling to local setups.
OpenAI Private Safety Processing Keeps Frontier Models on ZDR
OpenAI has previewed cross-interaction safety monitoring for eligible zero-retention API customers, with a technical paper due in September.
OpenAI Pauses Frontier Training Over Astra Cyber Risk
OpenAI has slowed some frontier training while it changes security, monitoring, and alignment controls around its unreleased Astra model.
GLM-5.3 Open Weights Delayed After Cyber Tests
Z.ai has launched GLM-5.3 through its hosted service while delaying the model weights after its own tests found stronger cyber capabilities.
Qwen3.8 27B Hardware Requirements Start Above 30 GB
Qwen's smaller 27B release is practical for high-memory computers, but the official files and context cache still put it beyond ordinary phones.
Pixel 11 AI Features Mix Local and Connected Work
Google's new phones run some AI locally, while the wider Pixel 11 feature list includes hybrid tools and online services with different data paths.
ONNX Runtime 1.29 Moves Browser AI Towards Native WebGPU
Microsoft's inference runtime adds WebGPU attention and quantisation support while beginning the move away from WebGL and JSEP.
Meta Muse Glimmer 30B Brings Local AI Agents to Consumer Hardware
Meta has released a 30B multimodal agent model with official local artifacts. The smallest build still needs about 17 GB before vision, context, and speculative decoding.
ExecuTorch 1.4 Expands Android and Qualcomm On-Device AI Support
PyTorch's on-device runtime now covers more Android, Qualcomm, Arm, and embedded workflows. The release improves deployment options without promising that every model will run faster on every phone.
OpenAI GPT-Live Shows the Trade-Off Behind Faster Voice AI
GPT-Live can listen and speak at the same time. Its speed depends on continuous cloud audio transport and stateful inference, which also define the privacy boundary.
EU AI Act Transparency Rules Now Apply to Chatbots and AI Content
The EU has started enforcing chatbot and synthetic-content transparency duties, but recent amendments delayed many rules for high-risk AI systems.
All guides
Clear explanations, practical checks, and primary sources.
- Offline AI on Android9 min read
Why You Should Keep an AI Model on Your Phone Before You Need It
Cloud AI depends on infrastructure you do not control. A downloaded local model gives your phone a useful capability that can remain available when connectivity or online services fail.
- Offline AI on Android9 min read
How to Install and Prepare an Offline AI PWA on Android
Installing a PWA adds the app experience, but offline AI is ready only after the shell, runtime, model, tokenizer, and local state are present and tested.
- Local documents9 min read
Text PDF vs Scanned PDF for Offline AI
A text PDF contains extractable characters. A scanned PDF may contain only page images and needs OCR before a local AI model can search or quote it reliably.
- Local AI models10 min read
Why Local AI Is Slow on a Phone and How to Diagnose It
Local AI speed has two separate parts: time before the first visible token and the rate of tokens after output begins. Diagnose them separately.
- AI privacy11 min read
Where Local AI Stores Chats, Models, and Files on Android
Local AI data is not one file in one folder. Chats, models, imported documents, indexes, caches, and backups can use different native or browser storage, with different deletion and recovery rules.
- AI privacy11 min read
Private AI Privacy Checklist: 25 Questions Before You Trust an App
Privacy labels are inconsistent. This checklist replaces slogans with questions that an AI provider should answer and tests that a careful user can repeat without uploading sensitive information.
- Local AI models10 min read
What Is AI Model Quantization for Local AI on a Phone?
Quantization can make a local model smaller and more practical, but a lower bit count is not a universal speed or quality score. Learn how precision, runtime support, memory, and task testing fit together.
- Local AI models8 min read
LFM2.5 vs Qwen3.5 for Mobile Local AI
LFM2.5 1.2B and Qwen3.5 0.8B are compact model families with different architectures, language claims, licenses, and intended uses. There is no honest universal winner: compare the exact quantised artifacts in the same app, on the same phone, with your own prompts.
- Local AI models8 min read
How to Choose a Local AI Model for Android
The best local AI model for Android is not simply the largest download your phone can hold. It is the smallest compatible model that performs your actual tasks reliably while leaving enough memory, storage, and thermal headroom for the operating system and app.
- Offline AI on Android10 min read
How to Choose an Offline AI App for Android
A decision framework for choosing an Android AI app that can run a model locally, explain its network boundaries, and remain useful when connectivity disappears.
- Offline AI on Android8 min read
Offline AI in a Browser: How PWAs, WebGPU, and Local Models Work
A browser can run an AI model locally, but offline success requires more than WebGPU. The app shell, runtime, model, tokenizer, storage, and update strategy must all be available before the connection disappears.
- AI privacy10 min read
Does an AI App Send Your Prompts to a Server? How to Check
Airplane mode is useful, but it cannot prove everything about an AI app. Build a feature-by-feature data-flow map, inspect documented recipients, observe requests where practical, and test deletion before entering sensitive content.
- Local documents10 min read
How to Chat With a PDF Offline on Android
Learn how an on-device AI app extracts a locally saved PDF, retrieves relevant passages, and generates an answer without relying on hosted inference.
- Offline AI on Android9 min read
Can You Use AI in Airplane Mode? A Practical Android Guide
AI can work in airplane mode when the model and runtime are already stored on your Android device, but online search, downloads, and cloud features cannot.
- Offline AI on Android9 min read
How to Use AI Without Internet on Android
A practical guide to setting up an Android AI app before going offline, confirming that inference is local, and avoiding surprises when connectivity disappears.
- On-device AI11 min read
Local AI vs Cloud AI: Privacy, Speed, Quality, Cost and Offline Access
Local AI keeps inference on hardware you control; cloud AI uses remote computing. The better choice depends on your information, device, task, and need for current or high-capability models.
- AI privacy10 min read
Private AI Chatbots With No Account: What “Private” Actually Means
No signup is convenient, but it does not prove that an AI chat is private. This guide shows how to evaluate inference, storage, network requests, retention, deletion, and optional features before you trust an assistant.
- On-device AI8 min read
What Is On-Device AI? A Practical Guide to Private Local AI
On-device AI runs a model on your phone, tablet, or computer instead of sending every prompt to a remote inference service. That changes the privacy, offline, cost, and performance trade-offs, but it does not make every feature automatically private or offline.