Beneath the Claim

Beneath the Claim

By Alexander WitteveenBusinessManagement
Download on the App Store

Beneath the Claim episodes

  • One Layer Below the Question
    CIONET put five questions to me on camera about scaling AI responsibly. The format allowed four and a half minutes. Here are the elaborated answers those questions deserve.
    20 min
  • Stage 1: Only the Seller Got Paid
    The first stage of a buildout is supposed to be the one where the early projects pay. This one paid exactly one party, and it was the party selling the equipment. What the filings of 2022 and 2023 showed, and when they showed it.
    19 min
  • The Model: Bubbles Break on Price, Not Demand
    Every infrastructure boom is right about the traffic and wrong about who gets paid for it. The last two times we built this much on borrowed money the tracks stayed and the money did not, and both ran the same five stages.
    27 min
  • One Dollar, Round and Round
    What you pay for AI is not what it costs to serve you. Follow the difference down the stack, one payer at a time, and it comes back to your own receipt. The loop from the customer’s seat, in ten minutes.
    16 min
  • The Source Won’t Save You
    Asking for the source is the instinct that feels like diligence. In the two best documented cases available it fails: the source published several different values, and the caveat was dropped by one of the people who wrote it.
    26 min
  • Nobody Called It Sabotage
    A company that sells AI adoption software surveyed people already using AI, folded seven different behaviours into one word, and published the cure. The word travelled. The reaction it is producing costs more than the behaviour ever did.
    34 min
  • The Scoreboard Is the Product
    The same model scored 0.637 and 0.488 on the same benchmark because two teams spelled the answer differently. Every AI number you have been shown is a scoring decision somebody made, and the one rule that encodes what you want is the part no vendor ships.
    25 min
  • The Half-Life of an AI Claim
    Almost no AI forecast can be graded, because almost none carries a date, a number and a company staked on it. One did. Scoring it two years on shows which layer of a vendor pitch is worth your money, and which one is engineered to expire.
    27 min
  • You Don't Need a Monorepo. You Need Coherence.
    The monorepo-versus-polyrepo war is back, and both camps are arguing about the wrong variable. What an AI agent, and a new engineer, can actually use is contextual coherence, and for a brownfield enterprise the road there almost never runs through a big-bang migration.
    12 min
  • The Default Nobody Chose
    Ask the wiki and your organisation has a branching standard. Ask the repositories and the truth comes back: a dominant flow nobody decided, a long tail with no model at all, and branch lifetimes nobody governs, just as agents start authoring at volume.
    13 min

About Beneath the Claim

From the publisher's feed

Long-form essays on software, organisations and the claims made about them.