Notes tagged
#ai
31 posts
- 2 min read
29 Hours of Silent Maximum-Depth Reasoning
Set the model to think at a moderate level. It quietly ran at full depth for over a day instead — no error, just silence.
- 3 min read
The Host Froze the First Night. Two Days Later, a Private AI That Saves ₽75,500 a Month
Bought a used GPU — the host froze the first night. Two days later: five real fixes, from power draw to the finance model.
- 3 min read
0.56 + 0.38 + 0.04 + 0.03 = 1.01
The VRAM budget for two models on one card never added up — a routine vLLM upgrade finally proved it.
- 3 min read
GPT-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.
- 4 min read
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.
- 3 min read
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.
- 3 min read
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.
- 5 min read
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.
- 2 min read
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.
- 3 min read
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.
- 2 min read
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.
- 2 min read
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.
- 3 min read
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.
- 4 min read
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.
- 3 min read
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.
- 3 min read
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.
- 1 min read
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.
- 4 min read
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.
- 5 min read
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.
- 2 min read
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.
- 2 min read
Next specialization: neonatology. Started with the sources
Expanding into medicine, starting with neonatology. The model wasn't the bottleneck — the sources were.
- 2 min read
The 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.
- 2 min read
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.
- 2 min read
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.
- 2 min read
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.
- 4 min read
Recorded 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.
- 4 min read
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.
- 3 min read
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.
- 3 min read
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.
- 5 min read
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.
- 2 min read
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.