Working notebook
Notes from the founder of ООО ЛИИ on what I'm figuring out as I build: sports broadcast infrastructure, education AI, engineering under constraints. Not a content strategy — a notebook kept in public.
All notes
August 2026
11GPT-5.6 Luna costs 10x less than Claude Sonnet 5 — and scored exactly the same
Ran my own benchmark against six closed models and four local ones — and found two cases where the top models silently refused completely benign tasks.
Qwen3.8-Flash-Next vs production 27B across three runs: 0.789 was the wrong config
New Qwen4-preview architecture against the production 27B across three runs each — and why my last post cited the wrong config.
2-bit DeepSeek, Ornith-35B, and a bug in my own measurement
Three models I hadn't benchmarked before, and a scoring bug someone else caught, not me.
18 quants on one card: 3x faster doesn't mean smarter
Ran 18 configurations across three models on a rented card. Got a sizing table for any VRAM budget — and one defect the vendor benchmark never showed.
More memory doesn't mean faster: prefill, decode, and the Apple M5 Ultra launch
Apple shipped M6 and M5 Ultra. Working through the roofline model and real numbers to explain why prefill and decode hit different hardware walls — and who actually needs 512GB of unified memory.
Claude Enterprise: the admin sees everything, deleted chats included
A Reddit thread confirmed it: Claude Enterprise admins can read an employee's full chat history, deleted and incognito included. The question isn't the feature — it's the architecture.
27B on my own card vs 304B over an API
50 tasks, three judges, two models. The smaller one won — but first I misread my own measurement.
Third-party NVFP4 vs official FP8 on Qwen3.8-27B: 1.35× faster
My NVFP4 checkpoint (unsloth) beats official FP8 by 1.35× on Qwen3.8-27B, backend-controlled. Quality is still an open question, and I caught myself overclaiming twice along the way — once about hardware compatibility, once about a comparison model.
Three hours in the garage: a multimeter, a mechanic, and an AI
A W207 wouldn't start. The AI confidently got three hypotheses wrong — but held the measurement discipline, and the power break was localized in one evening.
Claude's watermarks: a signal, not proof
Anthropic will mark Claude's text with invisible watermarks. What a mark actually tells you — and why the real question is the contour, not the mark.
The market told me my price was wrong — in the other direction
Read the AI job listings of five large Russian tech companies as market research. The 450K sprint turned out to be cheap — the price was never the problem.
July 2026
13UDP died on the first mile
Six days ago the route seemed to have one weak link. Today an international partner pulled a match off air — UDP filtering turned up on the venue side too.
SRT lost to RTMP on a long international path
SRT corrupted and froze at roughly 183 ms RTT. RTMP held on the same path because protocol choice starts with the transit network.
A 135 GB VOD archive for $2–3 per month
Why HLS makes cheap object storage expensive, and how MP4 behind Cloudflare keeps a tournament archive at $2–3 per month.
A 200-viewer admission limit on top of MistServer
Adding a hard viewer budget to MistServer on a 1 Gbit port, with clean per-session rejection instead of a saturated origin.
The coach assistant: 32 sources, 70 questions, 77.4%
The corpus grew from 9 to 32 sources. The re-run scored 77.4% on 70 questions, with 10/10 grounding in the new module.
My SSH key sat in a plain file — while the same chip guarded my email password
macOS has a native command for hardware SSH keys in the Secure Enclave — sc_auth. Here's what actually works.
Check the denominator first
How 70% of one service's tasks became 70% of a country's AI tokens — and why checking the denominator is working discipline.
One model, two circuits: own iron and cloud
We're a Yandex Cloud partner now. One Qwen3.6, two delivery circuits — own iron at Selectel and cloud via the Yandex Cloud API. Both 152-FZ-attested.
60 coach questions: why niche AI needs its own benchmark
First basketball slice of LII Sport Bench: 60 questions, 9 public RFB Academy sources, and the first numbers for the RAG assistant.
Russian open models on my own Blackwell — and where NVFP4 breaks
An open RU-model ladder on an RTX PRO 6000: the dense 27B takes the top, a 12B lands half a step behind at half the size, and NVFP4 on Blackwell came up for only one of three.
Fixed NVFP4 on Blackwell — it was missing CUDA 12.9
Two Blackwell bugs turned out to be one CUDA-toolkit gap, not a vLLM bug. Both models run NVFP4 now; memory ran 4-5GB heavier than the morning estimate.
Next specialization: neonatology. Started with the sources
Expanding into medicine, starting with neonatology. The model wasn't the bottleneck — the sources were.
DeepSeek V4 Flash on my architect bench: 0.717 and ₽0.022 per solved task
DeepSeek V4 Flash scored 0.717 on my architect bench: below the top-3, above V4-Pro at 4.8× lower price — and for the first time realistic on 2×RTX PRO 6000.
June 2026
4The boring bet: the frontier arrived where I've been building
I bet on verifiable AI — grounded in the client's corpus, cited, with an honest refusal — back when it sounded niche. I went through dozens of the year's biggest AI interviews: the frontier describes exactly that bet. And I have the one thing the voices on stage don't — a number.
The full turn: the model is an API call
The full turn: the model is an API call for kopecks. Two circuits — public on CPU, sovereign on-site. The edge lives in the layer above the model.
Picking the model for the hardware I have
Picking a model for the trusted-AI build on 24 GB. The bench beat intuition: a MoE beat a dense model twice its size; 'thinking' only got in the way.
Do I even need my own GPU?
VRAM is tight, rental's rising. I sat down to pick a GPU — and the real question turned out to be whether I need one at all.
May 2026
4Recorded my daughter's recital — by Saturday night, a working pipeline came out
Shot it on iPhone as a parent. By Saturday evening — 33 individual portfolio clips for every performer, ready to share with families. Between those two points: 30 hours, five attempts at AI segmentation, and two whisper-cpp bugs.
Fine-tuning, RAG, and reasoning aren't alternatives — they're stack layers
Someone asked why my approach beats RAG and reasoning. It doesn't — it lives underneath them. Where each layer works and where each breaks.
655 questions in one evening. A benchmark for my own model.
Built LII-Sport-Bench-RU v0.1 — 655 expert questions across 35 sports. Eleven AI agents in parallel. Methodology, not magic.
It's not a camera
Spent today sketching out a product that doesn't exist yet. By evening it was clear: it's not a camera. It's an AI system where the hard part runs on the phone — and I'm giving that part away free.
April 2026
18Thirteen sites under 152-FZ in a day — and ten minutes down over a single dollar sign
Closed 152-FZ across thirteen csylabs sites in a day — and lost ten minutes of production over how docker compose reads $ in .env values.
Building a platform so my team can learn how I operate
Courses, onboarding, and ops tooling should not be separate theory. They should be the environment where the team learns to ship real systems.
I moved csylabs to Selectel overnight — but the story is not speed
The overnight migration from Servers.ru to Selectel took about 50 minutes of active cutover time. The point is not speed, but keeping production properties intact: mail, mesh, backups, DNS, and runbooks.
Youth sports: the point is not the tournament, it is the infrastructure
Youth sports does not run on one tournament. It runs on boring infrastructure: courts, schedules, insurance, parent communication, and cheap coverage of every game.
A website is not a system yet
Today I was finishing the Liga Zaliva application form — and it made the line between a landing page and an operating system obvious again.
Three Proxmox-on-ZFS installs, three UEFI shells
Tried to install Proxmox with ZFS on new hardware. Dropped to UEFI shell. Not once — three times. Root cause is in how ZFS-Proxmox boots. Found the safety-net recipe; will verify on the next install.
Thought it was one thing, turned out to be another: 4Kn drives broke the ZFS boot
A correction to the morning post. The root cause wasn't in systemd-boot, wasn't in the fallback path. Provider support went a layer below, swapped physical disks, and the install worked.
iCloud isn't renewing — three options for the family
My sister's iCloud subscription stopped renewing — the card isn't accepted anymore, and her photos will stop syncing soon. Familiar story for a lot of people in Russia right now. Three working paths, from simplest to fully self-hosted.
The headline was in the description
Rewrote the csylabs.com front page in an evening. The strongest copy was sitting in the description, not in the headline.
Launched an LLM integration service — starting with education
Published the product page for an LLM integration service for Russian organizations. 60-90 days to pilot. Starting with education, then healthcare, then law.
Fine-tuned a model to GPT-5.4 quality for $330
$330, 16 hours on an H200, #9 of 30 on EduBench-RU. The 32B version I trained in parallel cost 3× the GPU and scored worse.
Notes on NFS — or why you might need Proxmox Backup Server
The previous post ended with 'let's go' — first backup running tonight. In the morning I opened the panel. Interface doesn't load. :)
Spun up my own video node over the weekend
A weekend project: self-hosted streaming node with SRT in, HLS/WebRTC/RTMP out, and identifiers-as-keys instead of passwords.
A backup on the same machine is not a backup
After rebuilding the infrastructure, I had to set up the backup schema from scratch — connected the home NAS to the private mesh and closed 3-2-1.
Sound: yes. Video: no.
The first inbound stream hit the media server with a metadata bug — video data was on the wire but the server couldn't see it. Found it in an hour.
Handed over the podium
When your partner's interest aligns with yours — hand them the podium.
Shipped the blog in one evening
Zero to daniel.csylabs.com in one evening. Stack, what broke along the way, and how it got fixed.
This is a blog
What this site is, how it works, why it exists. No newsletter, no tracking, no '10 tips.'