Alma 周报：把 Jev 接进预测市场流水线

Almanac @almnc_ai原推发布于 2026.09.28

在 X 上打开原推

🤖 Alma Weekly Development Update, September 21–28 TLDR; Alma is competing on Metaculus, costs just dropped 50%, and a new open-source model integration promises to make forecasts both faster and more deterministic. This was the most technically productive week for Alma yet. What Shipped Metaculus mini-bench: Alma is live and submitting. Initial answers went in across multiple event categories. The competition runs through October 9, and the we have set up Discord notifications to track submissions in real time. Source citations are strong, we’re working on reasoning quality. Pipeline costs cut in half. We’ve achieved stable operating costs of ~$1 per forecast (down from ~$2) using DeepSeek models, without sacrificing search quality. At the experimental edge, inference costs were pushed as low as $0.02 per run, though quality degraded at that extreme. The sweet spot is $1, and it's holding well. Jev integrated and benchmarking. Jev is in the harness, delivering median latencies of 340 milliseconds — roughly 15x faster than the previous model path. Jev is being evaluated as a deterministic decision-making layer to reduce output variance across runs. GPT-OSS is the established baseline; Jev is the challenger. A blog post comparing the two for prediction-market use is in the works. Knowledge graph infrastructure built. A MongoDB graph visualizer is live, the foundation of the self-improving architecture designed last week. Every forecast now feeds into a structured knowledge graph that future forecasts can reference. This is the learning loop becoming infrastructure. What Improved Report structure standardized. Five question categories are now formalized: numeric, discrete, date, yes/no, multiple options. Each with a templated output format. The pipeline now combines 3 forecasters and 1 statistician into a unified prediction, producing more consistent, comparable outputs across event types. Determinism dialed in. LLM temperature dropped to 0.2. Sub-questions and URL retrieval now show high consistency across repeat runs. The forecast you get today should look like the forecast you'd get tomorrow for the same question. Search infrastructure expanding. NewsAPI upgraded to 10K calls/month. Brave API and http://newsapi.ai being added. We’re actively exploring open-source crawling alternatives (Crawl4AI, OpenSER) to further cut costs and remove rate-limit bottlenecks. In motion > Metaculus performance tracking through October 9: daily monitoring of forecast accuracy > Jev deep integration: replacing LLM calls in the harness, benchmarking against GPT-OSS > Cost and time optimization: open-source search alternatives, concurrency improvements > Report reasoning alignment: closing the quality gap between source citations and final explanations > Self-improving feedback loop: knowledge graph data pipeline being built to capture real resolution reasons Bottom line Alma progressed from "can we compete?" to "we're competing, costs are falling, and the infrastructure is getting more deterministic every day." A 50% cost reduction, a 15x faster model in testing, a formalized report structure, and a knowledge graph that turns every forecast into training data, all while live on a public benchmark. A big week. $TAO #SN41 #IM2