Skip to content

Writing

How it was actually built.

Posts about decisions that cost something — the constraint we had to enforce in code, the asset we drew rather than licensed, the feature we cut. Written when there is something to say, not on a schedule.

5 posts shown

  1. RetrievalProductEngineering

    Put the knowledge in the index, and the model can stay in the building

    "It has to know our company's material" and "the material cannot leave the building" are usually two requests from the same person. As long as knowledge is something a model has to memorise, the two fight each other: the models that memorise the most are the largest, and the largest run in someone else's data centre. Put the knowledge in an index and let the model only read, and the two stop conflicting. This post makes the case from the published work: why retrieval works, why it pairs so well with a model on your own hardware, and what we built from it.

    9 min read

  2. AI for sciencePart 1ScienceStructural biologyBenchmarks

    From DENDRAL to AlphaFold: what let AI into science was the assessment

    In October 2024 the Nobel Prizes in Physics and Chemistry both went to machine learning, two days apart. Reading that as "AI finally got good enough" misses what happened. AlphaFold was accepted by structural biology because of CASP — a blind assessment, running since 1994, that hides the answer in advance. This is sixty years of AI for science in four stages, and what the real bottleneck was in each.

    7 min read

  3. AI for scienceScienceBenchmarksReproducibility

    What CASP14 actually measured

    The overview says AlphaFold2 was accepted because of CASP. This is the measurement underneath that sentence: what r.m.s.d.95 is, the two numbers the paper leads with, how far ahead of the field they were, and the one thing the confidence intervals will not tell you.

    4 min read

  4. EngineeringProduct

    Count the lines, do not trust them

    Ask a model for seven characters a line and a rhyme at the end, and it returns something that looks right. We changed it to count instead: syllables measured line by line, rhymes filed by the thirteen Mandarin rhyme groups, and only the lines that missed get rewritten. Here is the loop, and the pinyin bug we hit on the way.

    3 min read

  5. Product

    Seven tools, one account

    Websites, apps, lyrics, score, image, video, support — seven tools that each do one narrow job properly, sharing one account, one API surface, and the same rule about your data. This is what that rule actually says.

    2 min read

Have something in mind?

Tell us what you are building and we will tell you honestly whether we are the right studio for it.