40 posts
The real heat in tech is no longer just another model release. AI coding agents are entering real development workflows: taking tasks, editing repositories, running tests, opening pull requests, and forcing old questions about context, safety, and review back onto the table.
NVIDIA DGX Spark is easy to misunderstand. It is not a toy for ordinary users to run every chatbot locally, nor a shrunken data-center training cluster. What it really sells is 128GB unified memory, Grace Blackwell, the NVIDIA software stack, and a desktop form factor.
In a WSL environment with Ubuntu 26.04 LTS and an RTX 4060 Laptop GPU, this note uses micromamba to prepare CUDA 12.4, GCC 13, and cuBLAS without modifying system directories, builds the llama.cpp CUDA backend, and compares CPU and CUDA speed with Qwen3 4B.
On a local development machine with an RTX 4060 Laptop GPU, 8GB VRAM, and 32GB memory assigned to WSL, I tested Qwen3 4B, Qwen3 8B, Gemma 4 E4B, and Gemma 4 12B for load time, generation speed, VRAM usage, multimodal capability, and practical trade-offs.
The real value of local open models is not a benchmark score, but whether they can enter tasks that happen every day. This article puts Qwen3 4B, Qwen3 8B, Gemma 4 E4B, and Gemma 4 12B into three workflows: development assistance, image understanding, and writing organization.
A continuously updated local open-model installation note. Starting from a Windows + Ubuntu 26.04 LTS development machine with 64GB host memory, 32GB assigned to WSL, and an RTX 4060 Laptop GPU with 8GB VRAM, this article first decides whether Ollama should run on Windows or WSL, compares llama.cpp, LM Studio, vLLM, and other options, and then lists models worth keeping locally.
Starting from Andrew Ng's comments on small teams and high-context engineers, this article explains why AI pushes senior developers toward a broader product full-stack: problem framing, prototype, code, copy, compliance, launch, operations, and iteration.
A practical test of gemma4:12b audio input on a local test machine: the model metadata declares audio support, but Ollama's native /api/chat path could not reliably accept audio; the working path was the OpenAI-compatible input_audio block.
A practical test from official docs to local verification: Gemma 4 12B can understand video as frame sequences, but Ollama did not accept MP4 directly in this test. The stable path is frame extraction plus multi-image input.
World Cup prediction should not ask a model to guess a champion. A more useful approach combines historical results, team strength, goal distributions, tournament simulation, and live information into one probability framework.
Once browser agents move from reading pages to clicking buttons, filling forms, and submitting actions, their product value rises together with the risk of mistakes and overreach. Guardrails must mature before automation is granted broad authority.
Agent systems are no longer just about writing a prompt. What determines production quality is how we select, trim, order, validate, and replay context so the model gets just enough information at each step.
Edge AI is not a smaller copy of cloud LLMs. It is about latency, power, privacy, caching, product boundaries, and engineering constraints that live much closer to the device.
After a RAG system goes live, the question is not whether it can answer demos. It is when it answers wrong, why it answers wrong, and whether those failures have clear boundaries.
A practical look at the protocol boundary of Codex App Server, the product and engineering scenarios it fits, and how it compares with Claude Code and Cursor in automation, cloud agents, and custom integrations.
A practical test of Gemma 4 12B through Ollama: generating Canvas code, connecting it to OpenClaw, calling QVeris tools, and judging where a 12B local model is useful and where it still needs guardrails.
A token and cost breakdown from one real long Codex session: why context is expensive, which conversations cost the most, how prompt caching helps, and practical ways to split sessions, compress logs, use files, and preserve stable rules.
A practical Q&A about Codex collaboration modes, GPT-5.3-Codex-Spark, statusline configuration, skills, slash commands, codex exec, and the Codex SDK.
A sanitized field note on using Codex, Claude Code, Cursor, gh, and lark-cli for real engineering tasks: what work is worth delegating to AI, what boundaries must stay firm, and how to make AI-assisted work verifiable and shippable.
A practical record of turning idle RTX 4090 machines into local model servers with NVIDIA drivers, Ollama, OpenAI-compatible APIs, model downloads, benchmark scripts, and first-round model evaluation.
Cursor Composer 2 can trigger image generation in real workflows, and Cursor's changelog points to Google Nano Banana Pro. This article records the workflow, cost intuition, and a same-aspect comparison with Codex-generated images.
A practical guide to combining Codex and Cursor: product boundaries, Codex IDE, Codex CLI, the Codex extension inside Cursor, Skills, Cursor Rules, session context, and a workable coexistence workflow.
Cursor's TypeScript SDK exposes the agent capability behind IDE, CLI, and Cloud Agents as a programmable interface. This article explains runtimes, hooks, MCP, examples, use cases, and a CI failure repair workflow.
A practical explanation of Hermes skills as procedural memory, how skill_manage edits SKILL.md files, and why progressive loading keeps skills usable without filling the model context.
A practical overview of Hermes Agent: installation notes, model setup, differences from OpenClaw, terminal versus dashboard workflows, channel support, migration, and common commands.
A workflow-oriented look at using Cursor for long-form web fiction: repository structure, single source of truth, rules, daily writing prompts, consistency checks, context-window tradeoffs, and model-cost choices.
A grouped reading of OpenClaw v2026.3.22-beta.1 and v2026.3.22, covering ClawHub-first plugin installs, browser relay removal, image generation, plugin SDK migration, pluggable sandbox backends, model/search plugins, and gateway hardening.
A practical walkthrough for installing OpenClaw, running onboard, checking the gateway, opening the dashboard, and connecting WeChat through the ClawBot plugin.
A release-note reading of OpenClaw v2026.3.13-beta.1 and v2026.3.13-1, covering compaction, memory, browser automation, gateway RPC, control UI, channels, and security fixes.
A practical look at Cursor's position in 2026: it is not dead, but its main battlefield has changed. The useful question is not replacement, but whether Claude Code should become part of your workflow.
Instead of turning tool discussion into a succession story, this article separates what Cursor and Claude Code are each good at: IDE context, browser interaction, Auto mode, terminal workflows, and underestimated Cursor capabilities.
A reflection on using Cursor while working on www.zicode.com: writing posts, developing a WeChat draft sync script, reviewing agent history, polishing documentation, and extracting an open-source tool.
A development retrospective on the official QVeris Cursor/VS Code extension: browser OAuth callback, SecretStorage, MCP configuration, workspace rule injection, host differences, and the last mile of making tools actually usable.
A short Gradio note: set share=True in launch() to create a temporary public link for a local machine-learning demo, and use Hugging Face Spaces for permanent hosting.
A practical Windows 11 setup note for running DeepSeek R1 locally through Ollama, then connecting it to Chatbox, MaxKB, Open WebUI, and VS Code-based AI tools.
A practical early experiment with VS Code, Cline, and DeepSeek V3: configuration, cost impression, a Gomoku coding test, what worked, and where the AI-generated code failed.
An early hands-on note on Cursor as a VS Code-based AI editor, using a small C++ Gomoku program to test code understanding, alpha-beta AI logic, object-oriented refactoring, and project-level assistance.
A short introduction to Google's lightweight Gemma LLM family, including Kaggle resources, common model formats, 2B and 7B variants, and basic hardware expectations.
A simple local Gemma setup note: download the PyTorch 2B checkpoint from Kaggle, clone gemma_pytorch, set PYTHONPATH, install dependencies, and run the official script directly.
Study notes from Jeremy Howard's Kaggle notebook on how neural networks work, using a quadratic function, loss, gradients, learning rate, ReLU, and simple function composition to build intuition.