Rost Glukhov | AI Systems & Infrastructure

GitHub’s monthly trending page works for a language, because a language filter already defines the set.

A self-hosted AI agent needs more than a long context window.

Go 1.27 shipped in August 2026, and for most modules the upgrade is a single go.mod line followed by a go fix pass — but a handful of changes bite silently.

Unbounded fan-out is the default failure mode of Go concurrency: every go statement is one more task in flight, with no ceiling and no owner.

Virt-Manager (the libvirt GUI) copies a KVM disk and assigns the clone a UUID and MAC.

Amazon S3 has become the de facto API for object storage.

Go ships a production-capable HTTP server in its standard library, and since Go 1.22 the built-in router handles method matching and path parameters on its own.

Newsletter platforms split into thin email senders and full publishing systems with archives and memberships.

WordPress SEO plugins add the metadata, sitemaps, schema, and AI layer above WordPress core, and in 2026 that layer includes plugins that run inference on your own hardware.

Self-hosted SEO now covers crawlers, Search Console dashboards, rank trackers, MCP servers, and local AI.

AI coding agents can already write functions, modify repositories, run tests, and open pull requests.

Deep Research has become its own category of software, not just a model pointed at a search box.

In September 2026 the open-model efficient frontier sits between 25B and 34B parameters: near-frontier agentic work on one 24 GB GPU, at a fraction of 70B-class hardware.

Persistent memory turns an agent from a re-explained tool into one that carries context forward — but it opens a failure mode stateless chat avoids: an interpretation can become memory, retrieved as fact, and justify a stronger version of itself.

Mnemosyne is a local-first memory provider for Hermes Agent, storing working memory, structured facts, temporal data, and episodic history in local SQLite — no hosted service, no mandatory network calls, and unusually granular write control.

An agent that proposed “move persistence into a shared library” six months ago will happily propose it again unless something durable says it was already rejected – and OpenSpec has no built-in state for that today.

OpenSpec is a free, open-source CLI from Fission AI that gets you and your coding agent to agree on a change in plain Markdown before any code gets written, without the phase-gated ceremony of heavier spec-driven frameworks.

Migrating an AI assistant is not the same as copying an application config.

Ollama and llama.cpp are often compared as if they were rival inference engines.

ROCm and Vulkan both accelerate AMD GPUs for local LLM hosting, but they are not interchangeable.

A model can advertise a 128K context window and still fail at 40K tokens on a 16 GB GPU.

AI hype follows a rhythm: roughly every three years, the architecture everyone is betting on gets replaced by something newer.

Agent Skills and MCP servers are often presented as competing ways to extend an AI agent.

OpenCode’s command-line interface is built for scripting, CI pipelines, and unattended agent runs.

Superpowers packages a full spec-driven methodology into installable Claude Skills, enforcing brainstorming, planning, subagent-driven implementation, and strict TDD instead of leaving that structure up to you.

Pi Coding Agent is a minimal, open-source terminal coding harness that ships with four default tools and leaves most of the behaviour to extensions, skills and your own workflow.

Markdown looks like one language until the same file renders differently on GitHub, Hugo, Obsidian, or Pandoc.

Every API call feels like a simple transaction - until enough of them accumulate that your fine-tuning data, evaluation harnesses, and tool schemas are all shaped around one vendor, and switching stops being a routing change.

Most Claude Code sessions get slow and cluttered for the same reason: every exploratory grep, every log dump, and every “let me check one more file” stays in the main conversation forever.

Ollama is one of the easiest ways to run a local language model, but convenience can conceal the moment when a local experiment becomes a shared inference service that needs better scheduling and observability.

AI coding agents ship features fast, but specs, tests, and code quietly drift apart.

A dead-letter queue is the safety net that catches messages your consumers cannot process, so one broken payload does not block or silently drop everything behind it in the queue.

An LLM Wiki fails when old facts remain plausible, contradictions become polished, and generated summaries drift from their sources.

Syncthing keeps files synchronized across devices you control, making it one of the most practical tools for a self-hosted knowledge infrastructure that avoids cloud lock-in.

A circuit breaker stops your Go service from hammering a failing dependency, preventing cascading failures that consume goroutines, sockets, and memory until the entire system collapses.

Docker Compose and Podman Quadlet solve overlapping problems but come from different design centers, and choosing between them depends on whether you think in application stacks or Linux services.

The AI hardware landscape has shifted significantly in 2026, with NVIDIA, AMD, and Intel all competing for developers who need GPUs capable of running local large language models and AI inference workloads.

Running Hermes Agent on a headless server while connecting from a desktop client on another machine requires two server processes and a single client connection.

Developers comparing Spec-Driven Development setups in 2026 are usually not asking which model is smartest.

Spec-Driven Development works when the specification is a workflow, not a document you file away after kickoff.

Prompt injection gets most of the security attention in LLM systems, and it deserves attention, but it is not the whole problem once agents start calling tools and delegating work to other agents.

Most AI agent demos still behave like chat completions with extra steps: you send a prompt, wait a few seconds, and get an answer back in one response.

Docker Compose on a Linux server should start on boot, stop cleanly on shutdown, and survive reboots without manual intervention.

Installing Docker on Ubuntu should be simple, but in practice several Docker-shaped options compete for the same command name, each with different packaging, upgrade behavior, and security implications.

APT failures are common on long-lived Ubuntu machines, and they usually appear after a release upgrade, a third-party repository change, a removed PPA, a manually installed .deb, or an interrupted package installation.

A 70B model generates one token per forward pass, and each pass reloads weights from VRAM, computes attention across the context, and synchronizes memory.

Single-agent AI systems peaked in 2025 — you gave one LLM a prompt, some tools, and a goal, and it did reasonably well on bounded tasks.

Two writes that should succeed together will eventually fail separately.

Go’s context.Context is simple enough to use badly — and that is the problem.

Spec-Driven Development is one of those ideas that software engineers have reached for before and then set aside when the effort stopped paying.