UC Berkeley has open-sourced FreeToken, an edge-native inference engine that runs frontier-scale Mixture-of-Experts (MoE) models on consumer hardware. It treats GPU, CPU, host memory, and PCIe interconnects as a single unified inference platform.

The approach works because of how MoE models are structured. A model like DeepSeek-V4-Flash carries 284B total …

The frontier of AI is no longer just about who has the biggest model. It's about who can run intelligence cheaply, quickly, and everywhere. That shift is what Lean Inference is about: treating inference cost, latency, and accessibility as first-class engineering problems rather than afterthoughts.

For years the industry chased …

Sashiko (刺し子, "little stabs") borrows its name from a Japanese reinforcement-stitching technique — fabric repaired and strengthened at its points of wear. The metaphor is the whole pitch: an agentic system that stitches over the weak spots in proposed Linux kernel patches before they land. It's written in Rust …

Training & Adaptation Strategy

1. What approaches exist for training or adapting an LLM? — Pretraining, fine-tuning, instruction tuning, prompt engineering, RAG.

2. Base model vs instruction-tuned model? — Pure next-token predictor vs one aligned to follow instructions.

3. When would you choose fine-tuning over RAG? — Stable domain knowledge, style/format control, latency sensitivity.

4. When would …

A starter question bank for screening entry-level GenAI engineers. Grouped by theme, covering fundamentals through production concerns.

Fundamentals

1. What is generative AI vs discriminative AI? — Generative models learn to produce new data; discriminative models learn decision boundaries to classify or predict.

2. What is a large language model (LLM)? — A neural …

NemoClaw is NVIDIA's agent sandbox that lets you run AI assistants locally using your own inference backend — Ollama, llama.cpp, or cloud providers. It bundles OpenShell as a gateway and OpenClaw as the agent runtime, all orchestrated through Docker containers.

Here's a walkthrough of setting it up from scratch.

Setup …

Ollama is a lightweight, open-source LLM runner. Its :cloud model suffix lets you route prompts to free-tier hosted models — no GPU, no paid API key required. Useful for learning, prototyping, and small projects on modest hardware.

This post covers the full Ubuntu setup: manual install, service startup, chatting with Kimi …

When you send a prompt to an LLM, three layers shape how fast you get a response: the hardware (GPUs, TPUs, LPUs), the model size and architecture, and the inference engine strategies sitting on top. Most of the latency battle is fought at that third layer — and the core problem …

The multi-agent space is fracturing fast. Teams pick LangGraph one quarter, switch to CrewAI the next, then discover Google ADK or AWS Strands and wonder if they should migrate again. Every switch costs weeks — ripping out abstractions, rewriting orchestration logic, re-testing state management. This is the exact problem Terraform solved …

"The issue isn't the coding agent's coding ability, but our approach.
We treat coding agents like search engines when we should be treating them
more like literal-minded pair programmers."
— Den Delimarsky, GitHub Principal Product Manager

The Problem With Vibe Coding

If you've used an AI coding assistant, you've experienced vibe …