Writing — Varjosoft

Thesis

Under the Shared Sky

A founding thesis. Intelligence is becoming infrastructure.

The danger is not only that some companies will win and others will lose. It is that power itself becomes less dependent on social consent. What is missing is a civic layer built for machine speed...

Read the thesis →

Technical

Fitting Gemma 4 (~52 GB) into 12 GB 3-bit compression for a 26B MoE model. No calibration. 4.79/5 quality at 3.7x lower GPU cost than the previous best.
Apr 2026
Self-hosting LLMs in Europe: real costs, 5 models Five open models on European serverless GPU cloud. Cold start times, per-session costs, and when always-warm breaks even.
Apr 2026
Compressing both sides of LLM memory with one algorithm Compressing both weights and KV cache with one algorithm. Same math, different failure modes, and a community that fixed what the paper got wrong.
Apr 2026 · Substack
A 30B model in 17 GB of GPU memory Zero-calibration weight compression. Load any BF16 model from HuggingFace, compress in 4 seconds. Qwen3-30B fits on a consumer GPU.
Apr 2026
4x more concurrent users on the same GPU TurboQuant+ KV cache compression for vLLM. Fused CUDA kernels, benchmark results across 15 model configs, and what breaks.
Mar 2026
Brute-force vector search can beat indexed search When quantized brute-force scan outperforms pgvector's ANN indexes. TurboQuant, compression tradeoffs, and where each approach wins.
Mar 2026

Essays

Built with

Spegling The research behind these articles was done with Spegling, a personal authority system that maintains structured knowledge across sessions. Interested in early access?

spegl.ing

Find me elsewhere

Substack Medium LinkedIn Bluesky