Apple’s Free Local LLM Is Hiding in macOS Terminal
There's a working local AI model in your Mac's terminal that most people never notice — free, private, and requiring no account or API key.
llmlocal-llmon-device-ai
AI
LLM security, production ML, and AI risk — the technical side of artificial intelligence.
5 Essential Tools to Supercharge Your Claude Workflows
Five open-source tools—Graphify, Ponytail, Task-Observer, Claude-Mem and Headroom—that save tokens, keep context and make Claude Code more reliable.
llmai-toolsclaude-code
Harness Engineering: The Harness Around the Brain
The loop makes an LLM behave like an agent, but the harness is what makes it survivable in production — tool registry, context manager, state store, guardrails, and evaluators around the model.
llmagentsagent-harness
Loop Engineering: The Brain Behind the Agent
An LLM is a stateless next-token predictor. The 'brain' of an AI agent isn't the model — it's the loop you build around it. What actually happens in one iteration, and where the loop breaks in production.
llmagentsagent-loop
Why Quantized Small Models Will Dominate AI's Future — An Honest Take
The shift from giant generalists to tiny specialists. Why INT8/INT4 quantized small language models — not ever-larger frontier LLMs — are where most real business value will actually be captured.
quantizationsmall-language-modelsslm
Build LLM Vocab: Tokens, Embeddings, Vocabulary Size, and Context Windows Explained
The core plumbing behind every large language model — tokens, tokenisation, embeddings, vocabulary, and context windows — explained in plain English with examples you can actually try.
llmtokenstokenization
LLM Context Windows and Memory: How Models Handle Extended Dialogues
Why LLMs forget, how much they can actually remember, and the four techniques — bigger windows, RAG, persistent memory, and efficient attention — that push past the limit.
llmcontext-windowtokens
Fine-Tuning vs Retrieval-Augmented Generation (RAG): Choosing the Right Approach for Custom LLMs
Fine-tuning bakes knowledge into the model's weights; RAG lets it look things up. A clear-eyed comparison across cost, latency, freshness, and accuracy — plus when to use both.
fine-tuningragretrieval-augmented-generation
Fine-Tune an LLM on Your MacBook with LoRA: A Hands-On Guide
Fine-tune an open-source LLM on a MacBook in under ten minutes. A working LoRA + Phi-2 pipeline on Apple Silicon, trained on the GDPR, with no cloud bill attached.
lorafine-tuningphi-2
The Hidden Risks of Exposing LLM APIs
Shipping an LLM behind a REST endpoint feels routine — until you realise the model itself is an attack surface. Prompt injection, exfiltration through retrieval, and timing side-channels, with fixes that actually hold up.
llm-securityprompt-injectionrag-security
Building RAG Systems That Don't Fall Apart in Production
A ten-document RAG demo is an afternoon project. A half-million-document RAG system with real users is an engineering discipline. Here's what actually breaks, and the architecture that holds up.
ragllmvector-search
11 posts