[{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/agent-editing/","section":"Tags","summary":"","title":"Agent-Editing"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/claude-code/","section":"Tags","summary":"","title":"Claude-Code"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/cuda/","section":"Tags","summary":"","title":"Cuda"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/gguf/","section":"Tags","summary":"","title":"Gguf"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/inference/","section":"Tags","summary":"","title":"Inference"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/llama.cpp/","section":"Tags","summary":"","title":"Llama.cpp"},{"content":"Local inference experiments running Qwen 3.6 models on llama.cpp. Tweaking parameters to avoid paying a token subscription!\nSetup #AMD Ryzen 5 5600 + NVIDIA RTX 3090 (24 GB VRAM). Tried both vLLM and llama.cpp, with the latter being the easiest to setup.\nllama.cpp #GGUF-based server with BLAS (BLIS) acceleration and full CUDA offload. Runs the MoE model Qwen3.6-35B-A3B in Q6_K quantization, fitting entirely in 24 GB VRAM at ~23.4 GB with 256k context support.\nKey config:\nKV cache: f16, mlock enabled Batch size: 8192, ubatch: 2048 (optimized for MoE memory access patterns) CUDA priority: --prio 3 --poll 100 (reduces MoE kernel launch latency) Having some sysadmin chops helps here; I experimented with different setups to get to this point:\ninitially my AI machine was a TalosOS K8s VM node on proxmox with PCIe passthrough. ollama ran as a container via the nvidia node labeler operator ollama is consistently behind in support for various vendors; if I waited, it would\u0026rsquo;ve taken upwards of 5 months for me to test out the qwen3.6 models when they were first quantized by the unsloth team! Ubuntu and debian support was kind of spotty; I frequently wanted to try new CUDA drivers that were simply not in the apt repos yet Finally settled on CachyOS (I kinda use Arch btw) for the rolling updates. Tweaked my ufw settings, made sure nvidia CUDA libs and drivers were installed, BOOM - ready for the races.\nPerformance: ~43 tok/s, no GPU OOM at full context.\nNotes # Running with a display manager (SDDM + Hyprland) consumes ~600 MB VRAM; headless mode frees that for inference. The current temp., top-k, min-p and top-p have been perfect for smaller models and coding. Smaller models (sub 70B parameters) seem to be prone to looping due to \u0026ldquo;lower\u0026rdquo; attention/focus capabilities; the 1.5 presence penalty was the sweet spot for preventing looping, especially around tool invocations. Does it properly one-shot things? No. However, making your harness specify /effort level (e.g.: claude-code) is enough to make the LLM with these settings \u0026ldquo;explore\u0026rdquo; different ways of solving issues. HARNESSES and LOOPS- probably THE MOST IMPORTANT thing in any agentic-coding setup. This is ESPECIALLY so in local setups. I\u0026rsquo;ve found that my obsession with doing everything in the terminal + tmux has helped immensely in doing this; main LLM agent is only responsible with task ingestion, delegation to sub-agent, and running tests against those results. Leveraging integration tests, playwright suites, etc keeps any-and-all LLMs on track, no matter how short your context or how small they are. ","date":"26 June 2026","permalink":"https://chr0n1x.github.io/portfolio/projects/llm-experiments/","section":"Projects","summary":"\u003cp\u003eLocal inference experiments running Qwen 3.6 models on llama.cpp. Tweaking\nparameters to avoid paying a token subscription!\u003c/p\u003e","title":"LLM Experiments"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/local-llm/","section":"Tags","summary":"","title":"Local-Llm"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/lua/","section":"Tags","summary":"","title":"Lua"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/minuet-ai/","section":"Tags","summary":"","title":"Minuet-Ai"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/neovim/","section":"Tags","summary":"","title":"Neovim"},{"content":"My Neovim setup centered around Claude Code for agent-driven development. Treats local LLMs hosted via OpenAI-compatible APIs as first-class citizens!\nArchitecture #Typical neovim plugin setup.\ninit.lua -\u0026gt; flag-consts, consts, base-settings, key-bindings, lazy.nvim lua/plugins/ -\u0026gt; modular plugin configs loaded by lazy.nvim ai-claude.lua -\u0026gt; Claude Code via 99 + claudecode.nvim ai-minuet.lua -\u0026gt; OpenAI-compatible FIM completion via minuet-ai.nvim claude-code sits on TOP of my neovim buffers and I can quickly navigate to/from it.\nIn this world of AI - there\u0026rsquo;s no point in fighting the agents. Next best thing - integrate them!\nI settled on claude code for the following reasons:\nenterprise - my employer is 100% in with Anthropic (or so it seems) compiled bin - I can install claude-code on ANY machine I own, regardless of architecture claude-code supports open-api backends! self-hosters REJOICE! AI Integration #Claude Code — two interfaces, one workflow one LLM #I run Claude Code through both ThePrimeagen/99 and claudecode.nvim, each serving a different interaction pattern:\n99 handles agent-mode tasks. \u0026lt;leader\u0026gt;C opens a floating terminal prompt that dynamically builds prompts scoped to specific segments/lines that my cursor is on. That prompt is then fed to whatever LLM backend I\u0026rsquo;m using. Buffer refreshes happen immediately even while I\u0026rsquo;m typing in some other section of the file. The goal of using this is to give me a more hands-on approach to agentic coding; the LLM may be faster than spewing out code, but if the code that I\u0026rsquo;m dealing with requires very strict navigation via specific libs or styles I still would like to \u0026ldquo;have my hands on the wheel\u0026rdquo; to some degree.\nIn most cases though, I just use claudecode.nvim (backed by snacks.nvim).\n\u0026lt;leader\u0026gt;c — summon claude-code in a floating neovim terminal above my code. \u0026lt;leader\u0026gt;ca — add current buffer to context (either the entire file or specific lines) When coupled with treesitter and telescope, this overall UX gives me the ability to fly through files the same way that I always have! Only difference now is that I leave all of the typing and cross lib/file implementation to the LLM. Most of my workflows these days are:\nopen up neovim, find the file that I think I want to work on (\u0026lt;leader\u0026gt;p, so \u0026lt;spacebar\u0026gt;p -\u0026gt; fuzzy search) if the changes are smaller and I\u0026rsquo;m familiar with the codebase/language/framework - I can usually type faster than the LLM inference time if the change would span multiple files, or if I\u0026rsquo;m unsure - I put claude -\u0026gt; sub-agent -\u0026gt; go into plan mode and have it come up with a solution while claude is doing that (with whatever LLM backend, haiku, opus, even my qwen3.6 instance) - I\u0026rsquo;ll open up the tests and try to figure out what the tests should concretely cover, what regression tests or new use-cases must function go back to claude, have it write the tests based on how I think things should work, have it come up with other test cases too put claude into a loop. For the last bit - I run everything in tmux and have a command called /watch-tests that makes claude watch any tests that it spins up for the repo by investigating tmux panes. So while changes are happening, it knows what\u0026rsquo;s a valid \u0026ldquo;end-state\u0026rdquo; because I then have tests to help steer it the right way.\nI have some secret sauce in various personal and proprietary repos that helps claude \u0026ldquo;remember\u0026rdquo; what tests to run per repo, some other tools to save on token costs\u0026hellip;all of which is a bit of its own systems engineering ;)\nAnd all of this is happening while I\u0026rsquo;m in my editor, still jumping around!\nLocal LLM — OpenAI-compatible FIM autocomplete #Minuet-AI connects to my local inference server (llm-experiments or ollama) via its OpenAI-compatible API for Fill-In-the-Middle completion. This is effectively in-line auto-completion via LLMs on top of LSP suggestions! The idea is to have something like copilot or cursor super-tab functionality. The model and URL set through environment variables — and can be a different model, powered by a completely different provider!\nConfiguration #Both Claude Code and the local FIM backend are driven entirely by environment variables. The environment variables are read at config load time, so switching between local and cloud backends is a matter of setting the right env vars before launching Neovim.\nComing from a SRE background, I prefer that all configurations are shell env-vars for configuration injection. And because my work environment has a claude subscription whereas my personal setup is powered by a local/private LLM, I need this setup to be able to be configured on the fly.\nFinal Thoughts #I\u0026rsquo;ve been using vim and neovim for over 14 years now. The first time I saw vi in use was in college, by my professor in a graduate course. I was absolutely enthralled and dumbstruck with how magical he made editing code look. A few keystrokes in less than a second and he finished editing blocks of code.\nOver the years I\u0026rsquo;ve built up my vim-motion muscle memory. While I don\u0026rsquo;t think I can min-max vim golf the same way the terminal wizards do, nor would I ever achieve the grace that my professor displayed all those years ago, I do my darndest to achieve something smooth, something that\u0026rsquo;s a joy for me to \u0026ldquo;drive\u0026rdquo;. And over time it\u0026rsquo;s proven to make me faster (trust me bro). Countless times I\u0026rsquo;ve sat in pair-programming or brainstorming sessions and while folks are futzing around with cursors or awkwardly trying to navigate between their editor and terminal, I\u0026rsquo;ve already opened all files required, a terminal, and jumping between them.\nAI is here to stay. And I have to say - it\u0026rsquo;s been a joy integrating it into what already feels like an amazing setup. A bit hectic because of its speed. But a joy nonetheless.\n","date":"26 June 2026","permalink":"https://chr0n1x.github.io/portfolio/projects/neovim-ai/","section":"Projects","summary":"\u003cp\u003eMy Neovim setup centered around Claude Code for agent-driven development.\nTreats local LLMs hosted via OpenAI-compatible APIs as first-class citizens!\u003c/p\u003e","title":"Neovim + AI"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/projects/","section":"Projects","summary":"","title":"Projects"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/qwen/","section":"Tags","summary":"","title":"Qwen"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/","section":"Tags","summary":"","title":"Tags"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/vllm/","section":"Tags","summary":"","title":"Vllm"},{"content":"Intro #HI! Welcome to my blog/portfolio\u0026hellip;thing.\nNew York-based SRE and infrastructure engineer. Previously fullstack and ecommerce, now focused on Linux systems, Kubernetes, and building tooling that makes complex systems manageable.\nProfessional Experience # Ecommerce: inventory systems (SKU management and tracking), created APIs that helped shape user UX to increase sales (e.g.: quick-add items w/ only 6 left in stock to cart right before checkout), financial reports, UX adjustments to help sales productivity. SRE: CI/CD pipelines for application deployments (emphasis on sub-10 minute recovery times for outages). Testing libraries and harnesses (this is BEFORE LLMs) Security audits (routed requests for PEN testers). Scaled deployments for applications (e.g.: goin from deploy once a month to multiple times a DAY), eventually scaling and standardizing practices across multiple teams and BUs. Ops: Infrastructure A/B automations for 0-downtime cluster cutovers. AWS (cloudformation), Chef, Ansible, terraform; now ArgoCD and K8s. Implemented OTeL monitoring and alerting (Prometheus, Grafana, Alertmanager). Outage triage, direction, RCA composition/writeups. SWE: full-stack - front-end is admittedly my weakpoint (latest experience in TS React, Spotify backstage). Have written various applications to sensibly wrap all systems that I\u0026rsquo;ve worked on above, in a myriad of flavors (MVC monoliths, to microservice-based workflow systems). Current focus: shared LLM memory/context architectures and GitOps pipelines for AI workloads.\nThe above is a very abbreviated splurge of my experience using and developing proprietary software! For more details, you\u0026rsquo;ll have to speak to me face-to-face ;)\nPersonal/FOSS Work #Homelab \u0026amp; Infrastructure: I run a TalosOS Linux Kubernetes cluster on Raspberry Pi 5s, managing ~30 application services via ArgoCD GitOps. See my project entry on my homelab for more details!.\nLLM Ops: Running local inference experiments with Qwen 3.6 models (llama.cpp vs vLLM), speculative decoding, and quantized MoE models on consumer hardware. This entire site was coded with qwen3.6-35b-a3b :D\nAutomation Tooling: Built nox-bot — a Go CLI for Telegram messaging, LLM-powered service reports, and homelab automation. Also maintain pinchscrape (OPML feed scraper) and various utility scripts (luks-utils).\nInfrastructure as Code: Documented my Talos Linux + K8s homelab setup for others to replicate. Contributing to the Talos ecosystem through configuration examples and operational runbooks.\nOpen Source \u0026amp; Community: Arctic Code Vault Contributor - code archived at Svalbard. Pull Shark x3 (100+ PRs opened). Active contributor across Adobe, Behance, and Adobe Platform orgs on GitHub. I maintain helm-charts for any services that I self-host in my homelab!\nSkills #Infrastructure: Linux (RHEL, Debian, Alpine, Arch/CachyOS), Kubernetes, Talos Linux, Docker, Ansible, Terraform, Vault, Traefik, Longhorn, CNPG, NUT UPS, AWS\nCI/CD \u0026amp; GitOps: GitHub Actions, Flux CD, Argo CD, Helm, Kustomize, Gitea\nAI/ML Ops: llama.cpp, OpenAI-compatible APIs, speculative decoding, quantized MoE models, LLM-assisted development workflows\nLanguages: Go, Ruby, Bash, Python, TypeScript\n","date":null,"permalink":"https://chr0n1x.github.io/portfolio/","section":"任","summary":"","title":"任"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/argocd/","section":"Tags","summary":"","title":"Argocd"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/arm/","section":"Tags","summary":"","title":"Arm"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/grafana/","section":"Tags","summary":"","title":"Grafana"},{"content":"Self-hosted TalosOS Kubernetes platform running ~30 services across ARM and x86 nodes via GitOps (ArgoCD) with full observability, secrets management, storage, and the ability to incorporate \u0026amp; run GPU inference.\nIntroduction #I started with a handful of Raspberry Pis and an old GPU card, a hoard of spare computer parts\u0026hellip;no grand plan - just a question about what \u0026ldquo;production-grade\u0026rdquo; infrastructure looks like when it lives in the corner of my room. At work we were moving towards k8s as the basis of all our workloads so I wanted my homelab to kind of be a \u0026ldquo;safe\u0026rdquo; place for me to experiment with various technologies and processes; an environment close to my work environment but where I was free to not only experiment but also to mess up(!), and then to find solutions that would make me feel comfort and safety.\nThat might seem kind of strange, but in my mind I don\u0026rsquo;t want to interact with technologies that make me feel stressed and anxious every day. I know that I would perform better at any task if I felt like the environment that I\u0026rsquo;m in is resilient, self-healing, secure\u0026hellip;safe. Cozy, even!\nK8s conceptually is a nice basis for all of this. I\u0026rsquo;m already an avid fan of containerization, have years of experience with docker/CRI-O/podman. To me it\u0026rsquo;s this wonderful technology that let\u0026rsquo;s me package up my \u0026ldquo;things\u0026rdquo; and go. With it I can treat a collection of computers as fungible resources. The only friction to worry about would be configuration, setup and runtime - all of which boil down to a single question: can this machine run a container?\nEventually I ended up gravitating towards TalosOS Linux over vanilla Linux because the attack surface is minimal and node configuration is a well documented, machine manageable process. A read-only rootfs means you can\u0026rsquo;t SSH in and accidentally break something, the system either matches git or it doesn\u0026rsquo;t. Flashing the OS, wiping it clean, adding nodes - all trivial and automatic.\nAt first I used helmfile to kinda automate deployments of services that I wanted to run. But eventually I settled on ArgoCD to self-apply and heal everything. Including argocd itself. And with that\u0026hellip;my home started to fall into place!\nArchitecture #Two-tier deployment pipeline:\nhelmfile boots the platform layer — cert-manager, Longhorn storage, Traefik ingress, SMB CSI, ArgoCD itself, and internal registries (ChartMuseum, Distribution). These are seeded imperatively because some secrets aren\u0026rsquo;t tracked in git (initially!).\nArgoCD eventually takes over ongoing management of everything including itself. It creates ~30 nested Application CRDs from k8s/helm/cluster-apps/, each pointing to a Helm chart. ArgoCD continuously reconciles against git HEAD, self-healing drift and pruning removed resources.\nGitOps flow: local changes -\u0026gt; make sync -\u0026gt; helmfile applies platform tier -\u0026gt; ArgoCD reconciles all service Applications → cluster state matches git.\nServices # Category Services Identity Authentik (SSO), Vault (secrets), Twingate (remote access) Infra Traefik ingress, cert-manager (DNS-01 on DuckDNS), Longhorn, SMB CSI, Trivy Media Immich, Jellyfin, Pinchflat Productivity Paperless-ngx, Tandoor, BentoPDF AI/ML Ollama + Open WebUI (cluster GPU), CachyOS llama.cpp (edge proxy) DevOps Gitea, ChartMuseum, Distribution (registry proxy) Monitoring Prometheus, Grafana, Headlamp dashboard Other SearXNG, Kiwix, KEDA autoscaling, Descheduler Hardware # RPi5 control plane nodes Worker nodes: bare-metal RPi, ZimaBlade with NVMe, NVIDIA 3090Ti Proxmox host (PCIe passthrough to AI worker VM) Longhorn block storage on worker disks NUT UPS with PiSugar integration, etcd defragmentation automation Monitoring \u0026amp; Observability #Prometheus collects metrics from all cluster nodes and services via node_exporter, cAdvisor, and service-level exporters. Grafana dashboards cover system resource utilization, Kubernetes cluster health, namespace-level CPU/memory/network trends, and SLO tracking. Alertmanager routes notifications to Telegram via nox-bot for real-time incident awareness.\nPrometheus queries feed into shell services that extract per-namespace resource trends (CPU delta, memory growth, network bandwidth changes) and report the top three anomalies on each poll cycle — enabling proactive capacity planning without expensive commercial monitoring tools.\nInfrastructure Details # Storage: Longhorn for block storage with SMB CSI for NFS/SMB mount integration; etcd weekly snapshots backed up and managed via cron-driven defragmentation. Networking: Traefik ingress controller with cert-manager DNS-01 challenges on DuckDNS, split between internal (rannet.duckdns.org) and edge-accessible (rannet-edge.duckdns.org) domains. Secrets management: Vault with Vault Secrets Operator for Kubernetes-native secret injection; Authentik for identity and SSO across services. Power resilience: NUT UPS monitoring with custom systemd integration — on-battery scripts trigger graceful service shutdowns and notifications via Telegram. Architecture Decisions # Two-tier GitOps (helmfile -\u0026gt; ArgoCD) . helmfile to start and seed the cluster, but ArgoCD takes over ongoing reconciliation and self-healing for all service deployments. Pure GitOps top to bottom, including references to secrets (VSO).\nTalosOS Linux for all nodes. Chosen over vanilla Ubuntu/K3s because the attack surface is minimal (read-only rootfs, no container runtime exposed), node configuration is declarative and versioned, and upgrades are atomic across all nodes. The tradeoff is reduced interactivity — debugging requires SSH into a read-only system and using talosctl instead of familiar kubectl/debug containers. But to me this is and \u0026ldquo;a blessing in disguise\u0026rdquo; - I don\u0026rsquo;t have to care about what kind of machine I want in the cluster, I just need to answer one question - can this run TalosOS?. Upgrading the entire fleet is then trivial too!\nHardware tiering \u0026amp; segregation via workload \u0026amp; labels. Control plane on RPi5s + nvme (adequate for etcd and API server), media/AI workloads on the RTX 3090Ti host, bare-metal RPi for lightweight services. This avoids over-provisioning a single node type and lets each workload use hardware it was designed for. K8s lets me do this with taints/labels. It\u0026rsquo;s all GitOps. Code.\n","date":"22 June 2026","permalink":"https://chr0n1x.github.io/portfolio/projects/homelab/","section":"Projects","summary":"\u003cp\u003eSelf-hosted TalosOS Kubernetes platform running ~30 services across ARM and x86\nnodes via GitOps (ArgoCD) with full observability, secrets management, storage,\nand the ability to incorporate \u0026amp; run GPU inference.\u003c/p\u003e","title":"Homelab"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/kubernetes/","section":"Tags","summary":"","title":"Kubernetes"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/longhorn/","section":"Tags","summary":"","title":"Longhorn"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/prometheus/","section":"Tags","summary":"","title":"Prometheus"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/talosos/","section":"Tags","summary":"","title":"Talosos"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/bash/","section":"Tags","summary":"","title":"Bash"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/disk-encryption/","section":"Tags","summary":"","title":"Disk-Encryption"},{"content":"I don\u0026rsquo;t trust big tech anymore. Neither should you! This project documents how I set up \u0026amp; automate rsyncs to my LUKS drives.\nThe setup in the blog is simple - zero out your drives, set a passphrase, rsync things over, close drive.\nThe reason why I set up this repository and its scripts though is because I found the semantics around initializing and then mounting the devices to be cumbersome.\nWith this repo, assuming that the scripts are in $PATH - all I have to do now is something like:\nluks-mount sda cold-store full-sync luks-close sda cold-store The idea being that I as a human know that I have a drive sda that\u0026rsquo;s listed in lsblk and I call it \u0026ldquo;cold-store\u0026rdquo; - I don\u0026rsquo;t want to have to keep in mind that luksUtils will place it in /dev/mapper or whatnot.\nI use this particular setup to back up the most crucial portions of my homelab with storage on my NAS. I use the SMB CSI operator for applications running on my K8s cluster that do not need crazy high throughput, so this setup also lets me safely dump application data into my LUKS drives too. For example, I run an instance of Immich because I don\u0026rsquo;t want Google to train its automations/AI on the faces of my family. With this setup applications running on k8s are covered; I have backups and they\u0026rsquo;re encrypted! Yay for data sovereignty!\n","date":"14 April 2026","permalink":"https://chr0n1x.github.io/portfolio/projects/luks-automations/","section":"Projects","summary":"\u003cp\u003eI don\u0026rsquo;t trust big tech anymore. Neither should you! This project documents how\nI set up \u0026amp; automate rsyncs to my LUKS drives.\u003c/p\u003e","title":"LUKS Automations"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/openssl/","section":"Tags","summary":"","title":"Openssl"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/redundancy/","section":"Tags","summary":"","title":"Redundancy"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/tags/security/","section":"Tags","summary":"","title":"Security"},{"content":"","date":null,"permalink":"https://chr0n1x.github.io/portfolio/categories/","section":"Categories","summary":"","title":"Categories"},{"content":"By default a reductionist #Probably to a fault. Or maybe I\u0026rsquo;m just lazy and prefer to strip problems down to the bare-bones. If I were to put marketing jargon to it - I seek \u0026ldquo;Zen\u0026rdquo; when it comes to solving problems.\nWhen something comes my way, be it a painting, a short story, or a problem - if I\u0026rsquo;m genuinely interested in it my mind will start to churn in all manners. \u0026ldquo;What happened in the past that led to this result? What would happen if xyz?\u0026rdquo;. For paintings I\u0026rsquo;d look at the proverbial canvas, colors, medium and try to picture where and how the artist approached the process of creation. For works of writing I would try to see where the story/plot is \u0026ldquo;flowing\u0026rdquo; given the prevalent motifs in the work up to that point.\nFor solving problems, I\u0026rsquo;d try to picture how I would want to solve it, what an \u0026ldquo;ideal\u0026rdquo; solution would be and then try to work out the various moving parts that would need to be orchestrated to carry things out.\nThis attitude of mine extends to my work. All problems or states \u0026ldquo;flow\u0026rdquo; from a source, and they can all have an \u0026ldquo;ideal\u0026rdquo; solution or equilibrium state.\nMy attention is usually between these spaces:\n\u0026lt;source\u0026gt; -\u0026gt; \u0026lt;problem\u0026gt; -\u0026gt; \u0026lt;solution\u0026gt; I\u0026rsquo;m always looking to remove a layer of unnecessary complexity or decision surface. With infrastructure, that\u0026rsquo;s how I move forward - trying to achieve this \u0026ldquo;ideal\u0026rdquo; state where all \u0026ldquo;primitives\u0026rdquo; or \u0026ldquo;moving parts\u0026rdquo; of the system are idempotent and easy to understand, easy to combine.\nIn that way, you could say that I\u0026rsquo;m a very big fan of POSIX standards - do one thing, and do it very well. I try to achieve that in all my work by the time it\u0026rsquo;s in production.\nFor samples of my thought processes and notes around my work, see the projects section.\nAge of LLMs #Like everyone in tech, I\u0026rsquo;m trying to find my niche in all of this. I\u0026rsquo;m going through the same explosion of generated code, comments, messages, blogs, images, etc etc (ETC!) that we see. It seems these days that in order to survive in big tech you either need to be swimming in the kool-aid, praying that with AGI we\u0026rsquo;re on the precipice of the second coming \u0026hellip;or something in-between. Either way we collectively seem to be having an existential crisis of sorts if not feeling completely dejected and despondent.\nPersonally, I\u0026rsquo;ve settled on these key principles/opinions for myself, to steer my anxiety-addled mind:\nAI is here. Agents are here. No point in fighting it (regardless of how the datacenter situation turns out.) Unless proven otherwise (huge emphasis on otherwise) - it\u0026rsquo;s a tool. We should all learn how to leverage it where practical. I will never communicate with other humans using AI. Role of the Human in an Age of LLMs #Like any industry going through some upheaval, it\u0026rsquo;s better to accept that any technology causing disruption is here to stay in one shape or another. As a Millennial that viscerally experienced the world transition from no-internet to one day having cable/DSL and access to online games\u0026hellip;I know first-hand what this feels like, and I\u0026rsquo;m not going to fight it.\nHowever - it\u0026rsquo;s a tool. It\u0026rsquo;s incredibly fast at typing, debugging, finding all the nooks and crannies in which your code is broken/insecure.\nIt should NEVER replace the interpersonal relationships you have, though.\nMaybe I\u0026rsquo;m an extremist in this way, and in the future I may have to eat my words, but if you work with me you will never have to ask did this guy copy-paste a message from ChatGPT and send it to me? or does he have an LLM responding to his emails?\nqueue higher-level SVPs that get thousands of emails saying that I\u0026rsquo;m probably not an important employee\nI genuinely believe that it\u0026rsquo;s important for all of us, as humans, to strictly adhere to this kind of stance though. Who wants to personally interact with an LLM? When you\u0026rsquo;re down and have run out of tokens, with no LLM to interface with bots on your behalf - where will you go?\nIn this day and age where AI slop dominates our collective reality I believe that it is vital for us to maintain a shred of humanity somewhere.\nThe only way to do that is by leaving an avenue of true human interaction open to those that seek it. And this is something that needs to be consciously and intentionally preserved. Protected, even.\n","date":"1 January 0001","permalink":"https://chr0n1x.github.io/portfolio/engineering-philosophy/","section":"任","summary":"","title":"Engineering Philosophy"}]