mrsaynothing.dev
AI crawlers vs search crawlers: the 7-night census
Seven nights, 8,654 bot knocks, 247 human visits. AI crawlers vs search crawlers, measured: Perplexity beat OpenAI 9-to-1 and 65% knocked unnamed.
The Thinking Bill
A thinking model spent 866 tokens and 53 seconds to say “Bug”. The three switches that turn thinking off in Ollama — and the measured bill that says when to flip them.
The 7% Ghost
An 8GB card holds 8.59 decimal GB but only ~7.3 usable GiB. Sizes measured today, KV math shown, one honest fit table — the 9B the tables love fails at 8k.
GGUF VRAM calculator: the three numbers that decide
Weights + KV cache + overhead: the three numbers that decide whether a GGUF quant fits — printed as formulas, checked against four public GGUF files today.
Run LFM3 on Linux with Ollama (voice included)
ollama v0.40.13 brings Liquid AI's LFM3 — text, audio, omni — to Linux and Windows. The 1.7B rung benched on a CPU VM: 17 tok/s, 1.2 GB, one honest catch.
llama.cpp v0.8.0: the same LFM3 week, one layer down
Ollama wrapped LFM3 for pull-and-run the same day; llama.cpp v0.8.0 is where that week actually landed — realtime audio endpoints, six new architectures, sparse attention on a fused mask, MoE experts cached on host RAM.
ollama d1 vs d3 — which decision model do you need?
Two decision models shipped into ollama five days apart: d3 (2b-27b, runs on Linux CPU today) and d1 (2b, Apple-only this week). Which one to call.
Can vLLM run GGUF? The GPU-border rule
Can vLLM run GGUF? Yes — on GPUs only. The honest vLLM–GGUF matrix: six green hardware rows, every CPU row red, and the border rule the guides skip.
Why is my Ollama so slow? Five usual suspects, ranked
Nine times in ten it's CPU offload you didn't notice — ollama ps shows the split in one line. Then: cold reloads after the five-minute keep-alive, an oversized context, a heavier quant than you need, and swap.
What happens when Ollama runs out of VRAM? The offload cliff
Ollama never refuses a model that misses VRAM — it splits it. Fitted layers stay on the GPU, the rest fall back to CPU, and tok/s falls with them. How to see the split in one line, and the three honest fixes, ranked.
ollama d3 refuses to chat: the CPU bench ledger
Ollama's d3 scores tickets instead of talking. The 2b and 4b rungs on an 8-core CPU box: yes 98% in 0.8 s warm, a 64 KiB door fuse, and a verdict that wobbles.
What is Ollama's d3? Decision models that score, not chat
Ollama v0.40.9 ships d3: four decision models from vLLM Semantic Router that return scored choices instead of chat. Sizes, memory, index scores, fine print.
How to run pplx-decider locally with Ollama
Perplexity's 27B decision model sits in Ollama's library. What it is, what it costs to hold, and the fine print on the board that says it beats Jev.
The rate limiter that ate the test suite
The e2e suite expected 401 and got 429 — the daily IP cap had spent the day's budget on earlier runs. The wall, built as a machine: twelve window slots, a tape that remembers, an arm that drops at three.
llama.cpp vs ollama performance is a category error
Both shipped this week. One is the engine room, one pins it with a warranty — and the speed gap you are hunting lives in your config, not the wrapper.
llama.cpp connection refused after update? the port moved
llama-server v0.7.0 moved its default port from 8080 to 9931. Connection refused after an update? That is the whole bug — here is the flag that puts 8080 back.
Ollama connection refused on 127.0.0.1:11434 — every fix
Connection refused means nothing is listening on 11434. The three causes — server down, OLLAMA_HOST moved, wrong machine — and the checks that settle each one.
The rsync generator that refuses a bare --delete
Paths, excludes and bandwidth in, a paste-ready command out — dry-run by default, and every --delete ships with a dated backup net that makes the mirror reversible.
Three Boxes — the tale of one model in three containers
A field guide in five spreads: the same Qwen3-0.6B weighs in three ways, opens to show its first bytes, forgets on purpose at 4-bit, and picks its shelf — carry one crate the whole way and the page remembers yours in the verdict.
jj v0.46, second read: the on-ramp widened, the undo grew guardrails
Two weeks, +216 stars and one release later: worktree colocation makes the trial cheaper, cross-workspace undo is refused by default — and the bench is still owed.
vLLM on Windows: three doors, one supported
The docs say OS: Linux — WSL2 is the supported road, the native fork costs the fork tax, and Ollama was probably your real answer anyway.
Git undo keep changes: five verbs, one fear
The FAQ for every spelling of 'undo my commit but keep the work' — one command answers the family, and the only flag that deletes work is never required.
Can vLLM run GGUF? Ten questions, answered first
The FAQ hub: plugin yes, GPU only, repo:quant_type plus the tokenizer that is never optional — every answer before the explanation.
What decision models can't do: six limits, stated by the record
No reasons, no verification by consensus, licence walls, silent printed limits, unproven calibration — the wave scoped, not killed.
Jev decision models explained: a label, a number, no essay
Tiny models answer yes-or-no with the odds attached — the six local checkpoints, the licence trap, and why 0.51 and 0.99 are different decisions.
