Tag: llm

Train a 5.8M-parameter GPT on Sherlock Holmes: 59 seconds on an RTX 3070 vs 48 minutes on its CPU, and the overfitting that starts near step 1,250 of training.

Build causal self-attention and RoPE in plain PyTorch, then measure naive vs fused attention on an 8GB GPU: 6400 MiB vs 8.3 MiB of VRAM at 16k tokens.

Build a byte-level BPE tokenizer in about 100 lines of Python, then compare it to tiktoken's gpt2, cl100k_base and o200k_base on prose, code, YAML and German.

Cloudflare's Clef decision model tied Jev on accuracy in a private benchmark, abstained on junk better, and cost 4.2x more per decision. Real numbers inside.

Prompt-level defenses against injection are probabilistic and lose to a patient attacker. Here's the capability-level design that holds, with working configs.

Loading model weights, not math, is your GPU bottleneck. Continuous batching with llama-server serves many concurrent requests for close to the cost of one.

A 2026-09-22 patch left a self-hosted model's weights untouched but made it more confident on nonsense input. Six systems compared, real numbers only.

Jev, Von, and SemIf scored head to head on real commit history: a frozen 4B general model beat a purpose-built 395M decision model on every task tested.

I ran TypeSafe's Jev model against 840 blog posts to fix tag drift. Real cost numbers, benchmarked accuracy, and the failure mode that almost cost pageviews.

Komodo's community MCP server beats mounting the Docker socket or handing an AI agent SSH: scoped auth, per-tool RBAC, and confirmation before deletes.

Nano Banana's blurry images come from two causes: permanent generational loss from iterative editing, and a June 2026 outage Google fixed within 15 hours.

Coding agents write flawless code and leave hero.png blank. Compare self-hosted GPU image generation against subpixel, a ChatGPT-subscription CLI tool.