One embedding model, four different maximum lengths
The most downloaded sentence-embedding model ships four config files declaring 128, 256, 512 and 512 as its maximum inpu...
Guides & How-tos · Worked examples · Real numbers
Step-by-step how-tos, head-to-head comparisons and honest roundups — every guide worked through with current figures, then backed by a RECATOOLS calculator or AI-directory entry so you can act on it.
The most downloaded sentence-embedding model ships four config files declaring 128, 256, 512 and 512 as its maximum inpu...
OpenAI charges 2x input and 1.5x output on the full request once a prompt passes 272,000 tokens, so the 272,001st token...
All three major providers price a cache read at a tenth of normal input, so caching looks like the same deal everywhere....
Every hosted model has a date after which it stops answering. Almost everywhere that is an error you can see in your log...
A long session is cheap. An interrupted one is not. Every AI coding agent re-reads your whole conversation at a tenth of...
Twenty questions with known answers, run over 205 real documents. Prefixing titles to chunks raised hit@1 from 14 to 17...
Seventy-eight lines with no dependencies join a local embedding model, a vector store and MCP. The interesting part is n...
A 46 MB model, running on your laptop, turns text into positions in space. The demonstration: a query whose only content...
Every retrieval tutorial says "point it at your vector database" and none of them tells you how to run one. One containe...
MiniMax H3 generates video with the sound already in it, and stops at fifteen seconds. Five ComfyUI nodes now exist to g...
MiniMax Music 3 writes complete five-minute songs with sung vocals, and the weights are public. Here is what hardware yo...
A parameter count is not a size. Mistral 7B is 14.5 GB of weights before the KV cache exists, and most model repositorie...
Giving an agent tools is one line of configuration, and that line does not say what any of them can do. We read all 77 t...
robots.txt is the one place a publisher's position on AI training has to be written down in public. We read it for 19 ne...
We checked the 1,000 most-downloaded models on the Hugging Face Hub. 99% ship a model card, so documentation is not what...
We tried to tier all 1,370 entries in our own AI directory against the EU AI Act. A keyword scan gave 42 hits, 26 of the...
We set the ten OWASP LLM risk categories against what our directory of 1,366 AI tools actually records. Four have a fiel...
Every business owner wants a map of where AI is reliable and where it is not. We built five task types — drafting to spe...
Tell the model it is an expert. Offer it a $200 tip. Threaten it. Say please. Add "let's think step by step". We ran all...
A language model emits one symbol at a time, in order. An image model starts from noise and refines the whole canvas at...
Every guide to controlling an AI bill opens by telling you to trim your prompt. We ran the arithmetic over all 27 models...
The usual answer is five neat categories — chat assistants, copilots, command-line agents, autonomous runtimes, model pr...
The usual case against fine-tuning is that it is expensive. That case is wrong — a small training run costs about four d...
The same sentence costs seven tokens in English and fifteen in Malay. That gap is usually explained as a money problem;...