📰 Dev.to · Multigrid

A decision table by task, the break-even volume derived from labelled assumptions, and the three considerations that are not about cost at all.

What the classic n-gram metrics compute, the reporting traps that make published scores incomparable, and the cases where they still beat a model judge.

How one vector space holds many languages, the language-bias failure that makes cross-lingual results look worse than they are, and how to build the language-pa

The Nature result, the follow-ups that reversed its conclusion, and the follow-ups that reversed it back — with the experimental regime each one holds under.

What the llms.txt format actually specifies, why publishing it is not the same as anything reading it, and the one-line log query that tells you which is happen

A closed schema, a prompt that demands evidence spans, and a validation pass that rejects every entity and relation not grounded in the source text.

What each status code and error type actually means underneath, which ones are safe to retry, and the failures that arrive with a 200.

Five charts, each shipped as SQL against a stated schema, with what each one is for and what it cannot tell you.

Judge prompt design, the published human-agreement figures, and the validation step that has to happen before a judge's score is allowed to mean anything.

The ingestion pipeline as LlamaIndex's real product — readers, nodes, metadata — and the four points where its query abstractions stop helping.

How to attach a source to every fact — named graphs, RDF-star, edge properties or an assertion table — and the retraction query that makes it worth the storage.

A Jupyter setup for LLM experiments: a spend guard in the first cell, a disk cache so re-running is free, and a path for getting the code out into a module.

The published quality-over-quantity results — LIMA, AlpaGasus, the textbook-quality work — what each actually compared, and a filter stack you can run over an e

The three-stage method that turned a text continuation engine into something that follows instructions, and the cost the paper reports for it.

The pipeline between a question and a citation list, stage by stage, marking which stages you can observe and which nobody outside the lab can.

There is no such thing as a compliant model. There are business associate agreements, de-identification, and a set of questions to put to a vendor.

Five classes of question that similarity search structurally cannot answer — multi-hop, aggregation, negation, conjunctive constraints and completeness — each w

The families of evaluation tooling, the four axes on which each one locks you in, and the exit test to run before you commit a year of results to any of them.

What each extra dimension costs in bytes and in query work, derived in full, set against what model vendors have published about the quality it buys.

What has to be in the trace for a complaint about yesterday's answer to become a reproducible request today.

The finding, the tasks it was demonstrated on, the scale condition attached to it, and the much larger claim it turned into.

What coreference is, the datasets that define its hard cases, and why unresolved pronouns are a concrete cause of retrieval failure in multi-turn assistants.

What each common VRAM tier can hold, derived from the memory formula rather than reported, plus the speed that tier's bandwidth implies.

Crawl, chunk, embed and serve semantic search over a site — with a re-index job that only re-embeds the pages that actually changed.