Local AI Inference Speed: What to Expect (September 2026)
Real tokens per second from local LLMs on consumer hardware: how quant and context change speed, why memory bandwidth is the limit, and what helps.
Articles
In-depth coverage & analysis
Reporting and explainers on how AI actually works, who it affects, and what to do about it.
Border AI towers watched 1,050 deaths and never raised an alarm
MIT Tech Review's 15-month investigation finds Anduril, RVSS, and Elbit surveillance towers failed to catch people dying within range of their cameras.
How to Read an AI Privacy Policy Before You Sign Up
The five sections that actually matter in any AI privacy policy, what each major assistant's policy really says, and the red flags that mean walk away.
Best Local Coding Models by VRAM (September 2026)
Open-weight coding LLMs sized by VRAM: Qwen2.5-Coder, Qwen3-Coder, Devstral Small 2, KAT-Coder, GLM-4.7-Flash, and the trade-offs at each tier.
How to Delete Your Data From ChatGPT, Claude, and Gemini
The exact steps to delete chats and accounts from ChatGPT, Claude, and Gemini, and what each one keeps on its servers after you hit delete.
Vocci Ring Records Everything. Who Knows You're Recording?
The $249 Vocci ring is the latest always-on AI notetaker. We dug into its privacy policy, where the data flows, and the recording laws.
Local AI Power Draw and 24/7 Inference Cost (September 2026)
NVIDIA and Apple TDPs, the US residential cents-per-kWh average, and what 24/7 local inference actually adds to your electricity bill by card.
FOIA Records Show Medicare AI Prior Authorization Is Broken
EFF obtained ~1,000 pages of FOIA records on Medicare's WISeR AI pilot. Vendors shipped untested code; providers report patient harm.
AI Assistants That Don't Train on Your Conversations
Which AI assistants train on your chats by default, which keep data only briefly, and where the real opt-out toggle actually lives.
What Thinking Mode Actually Costs You (September 2026)
Reasoning tokens, latency, and KV cache cost of thinking mode in local LLMs. How to toggle it per runner and when it is the wrong choice.
Inside OpenAI's Project Lily: Who Reads Your ChatGPT Prompts?
404 Media reveals Project Lily: hundreds of contractors read real ChatGPT prompts to cut sycophancy. OpenAI admits sensitive details slip past scrubbing.
How much context can a local LLM actually hold? (September 2026)
The advertised context window is one number. What your GPU can actually serve is smaller. The KV cache math, RoPE scaling, and the practical ceiling.
How to Safely Download and Verify Open-Source LLMs
Sourced steps to safely download and verify an open-source LLM: safetensors, GPG-signed commits, pinned revisions, and HF / Ollama integrity checks.
Fable 5.1 Cracks a 370-Year-Old Cipher in 44 Minutes
Anthropic's Claude Fable 5.1 reportedly decoded a royalist cipher first published in 1653 using 176k tokens. The plaintext is verifiable. The framing is not.
Local LLM Backends Compared: Ollama, LM Studio, llama.cpp, vLLM
How Ollama, LM Studio, llama.cpp, and vLLM differ on model format, OpenAI-compatible API, hardware support, and which to pick for a home server.
How to Choose an Open-Weight Model Family (September 2026)
Qwen, Llama, Mistral, Gemma, DeepSeek, Phi - which family to commit to, what each is good at, and the licence traps that send you back to negotiate.
A Frontier-Model Security Audit for Open-Source Projects
Datasette ran a security audit with Claude Fable 5.1, GPT-5.6 Sol, and GPT-6 Astra. The workflow matters as much as the bugs.
How to Evaluate Local LLMs on Your Own Workload (September 2026)
Why general benchmarks like MMLU and GPQA don't predict your results, the index-rebase trap, and a recipe for picking a local model from your own prompts.
AI Built a Cross-Platform Zero-Click Worm in Three Days
How Calif Research used AI to find a WeChat bug and write a working zero-click worm in days, not months. What changes when the exploit loop is automated.
Serve a Local Model to Your Team (September 2026)
Auth, reverse proxy, HTTPS, concurrency and audit: how a small team shares one local model without exposing it to the public internet.
Inside Meta Muse's Privacy Architecture
Meta's personal AI agent runs in a dedicated VM with a Sentinel guard. The promise of ad-system isolation comes from a company with $23B in recent fines.
Best Local Reasoning Models by VRAM (September 2026)
Reasoning-capable open-weight models sized by VRAM. DeepSeek-R1, Qwen3 with thinking, QwQ, Phi-4 reasoning, gpt-oss, and the licence traps.
Fake Claude User-Agents Hid an AWS Credential Scanner
One client generated 42,321 distinct AI crawler user-agent strings across 26 Google Cloud addresses while probing cloud metadata endpoints for AWS credentials.
Local AI When One GPU Isn't Enough (September 2026)
Multi-GPU options for self-hosted AI when one card runs out of room: how Ollama, llama.cpp and vLLM split models, and what the interconnects actually cost.