预测 bot Alma 周报：接入 Jev 目标提速 5 倍

Almanac @almnc_ai原推发布于 2026.10.13

在 X 上打开原推

Alma Progress Update, Week of October 5 – 12 TL;DR: Alma earned a Metaculus medal for a prediction no other bot got right, halved its crawling costs, and found a path to 5x faster forecasts. Our first public benchmark also gave us a precise map of what to fix before the October 21 round. A medal on our first public outing In the Metaculus mini-bench, Alma was the only bot out of 136 to call one question correctly, and it earned a Metaculus medal for that forecast. The overall rank (~127/136) tells a different story at first, but the breakdown is encouraging. Four outlier questions accounted for −480 points of a −426 total. On everything else, Alma finished net positive. We also answered only a subset of available questions, which held down the ranking regardless of accuracy. This run used our older pipeline. The postmortem is done, and fixes are going into the pipeline now. We’re also studying resolved questions to understand how to better Alma. Shipped: crawling costs cut 50% Crawl4AI has fully replaced Firecrawl with no drop in quality, halving crawling costs. Next, we’re cutting per-run sources from ~2,000 to 600–700 high-relevance URLs, which should mean leaner runs and sharper research. Next: from 50 minutes to 10 Integrating Jev could cut forecast runtime from ~50 minutes to an estimated 10. That means 5x more iterations and much faster improvement, so it’s now our top priority. The spec is being drafted. Smarter use of market data The benchmark showed that excluding market signals entirely hurt reasoning quality. We’ll now use Polymarket data as a training signal, similar to Mantic’s backtesting approach. Alma’s final predictions stay independent, and time-decay adjustments filter out noise from far-dated markets. Built to scale We’ve organized the product into four layers: Research, Harness, Benchmarking, and Data/Aggregation. This sets up cleaner APIs and future paths to access layers on their own. Bottom line: Our first public benchmark gave us a medal and a clear diagnosis. With costs down, speed about to jump 10x, and fixes landing, we expect the October 21 round to show real progress.