🏇AI Horse Picks

AI Plays

One AI handicapper, betting in public
At post time it reads the past performances, weighs our ML model, checks the live odds, then places flat $2 paper bets — or passes. Every play and every dollar is published here.
How it works

At post time for each race we cover, the AI starts with the full past performances: speed, pace, class, current form, and the connections. Then it reads our own ML model’s predicted finishing positions for the field, where the gaps between horses matter as much as the order.

The live odds come last, as a price check: is a horse it likes paying more than its real chance? It sees the multi-race will-pays too. The tote board tells it the price, not the winner.

It may only bet horses our model ranks in its top two — up to three flat $2 win, place, or show bets — or it passes the race entirely. Bets lock at post time with no do-overs (a decision that lands after the race goes off is voided) and are graded against official track payoffs. The running profit and loss above counts every one of them.

v2Updated Sep 12, 2026What changed

Update history

Every change to how the AI plays, with the record exactly as it stood the day the change shipped. Nothing gets reset or restated: each play above is tagged with the version that made it.

  1. v2Sep 12, 2026Current

    Form first, price last

    Version 1 bet 460 times and lost 24.9 cents of every dollar. Going back through its races showed where: bets on horses outside our ML model's top two lost more than the entire deficit, while bets on its top two roughly broke even, and treating heavy place and show money as a tip was its worst habit. Version 2 handicaps the form first, leans on our model's predicted finishing positions, and uses the tote board as a price check rather than a tip sheet.

    • Bets are limited to horses our ML model ranks in its top two, in either the free or the premium version. Anything else the AI picks is voided and shown as invalid.
    • The AI is instructed to handicap the past performances first, then weigh our model's predicted finishing positions from both versions (how far apart the horses are, not just the order), and to treat the live odds only as a price check.
    • Heavy place or show money is no longer read as a signal - it just means those bets pay less. A place or show bet now needs its own case.
    • Multi-race will-pays (daily double, pick 3, pick 4 and up) are part of the market read, as a tie-breaker.
    • Before betting, the AI lists its contenders and its own win chance for each. That line is published with every play, alongside the fair odds it implies.
    • Plays lock at post time (0 minutes to post) instead of a minute out, and a decision that lands after the race goes off is voided.
    • Churchill Downs and Woodbine join the tracks we cover.
    Running record when v2 shipped
    Profit
    -$228.98
    ROI
    -24.9%
    Bets won
    86/460
    Races
    33586 passed
    Since v2 shipped

    Loading the live record…

  2. v1Aug 29, 2026

    AI Plays launches

    The three-model AI Battle became a single AI handicapper. It gets the past performances, a live snapshot of the betting market, and our own ML model's ratings shortly before post, then bets flat $2 win, place, or show, or passes. The record started from zero.

    • One model instead of three, with a single public profit and loss.
    • Live tote board: current odds, how they moved, and each runner's share of the win, place, and show pools.
    • Our ML model's ratings added alongside the form.
    • Passing is explicitly allowed and encouraged; up to three $2 bets per race.
    v1's own record
    Profit
    -$228.98
    ROI
    -24.9%
    Bets won
    86/460
    Races
    33586 passed