Original Summary

So, I'm posting this here because I've been building something for the past few weeks and I need people to break it before I take it further. We all ask ChatGPT or Gemini stuff like "will the RBI cut rates in December?" or "will petrol get cheaper next year?" and get a long paragraph that ends with "it depends on various factors." Nobody checks back later whether that answer was right. A chatbot can be confidently wrong every single day and you'd never know. So I had the idea of building a forecaster that has to put a number on it and then gets held to that number. It's called Kāla (Sanskrit for time). How it works: you ask a question about the future, like "Will India win the 2027 Champions Trophy?", "What will onion prices be in January?" or "When will the Mumbai–Ahmedabad bullet train actually start running?" It searches the news, reads dated sources, and gives you: - a probability (say 31%), or a range for numbers and dates - the reasoning in plain English, with every source linked - what would change the answer - a share card you can post with the number and a short "why" It also does a quick mode (under a minute) and a deep research mode (1–3 minutes), and you can ask in Hindi too. Under the hood it isn't just one AI. For yes/no questions it averages several different models (gpt-oss, Gemini, Qwen, Kimi), because on the hindsight-proof benchmark I built, the average scored better than the main model on its own. Honest disclaimer: the edge is small and not yet statistically certain. Now the part I actually care about, the accountability part: - Every forecast is saved the moment it's made, and once the question resolves it's scored against what really happened. Misses included, nothing deleted. - Every day, all the forecasts are fingerprinted and timestamped on the Bitcoin blockchain (using OpenTimestamps), so not even I can quietly edit or backdate a bad call. - Kāla also makes about 10 forecasts of its own every morning on public events, so the track record grows even when nobody is asking. - It's entered in two outside AI forecasting competitions: it's forecasting in the Metaculus AI tournament now, and joins ForecastBench from its 25 October round. The first batch of forecasts resolves in late October, so the public scoreboard is basically empty right now. That's partly why I need testers now. I'm also training my own model for this (Kāla-1). If you're into ML, you can literally watch it train live on the site: layer by layer, learning curves, pass/fail checks set before training started. The failed experiments stay in the public notebook too. My long-memory version recently failed its recall test and the result is sitting right there. What I need from you: - Ask it real questions you actually care about: markets, elections, your city, cricket, tech, prices, anything. (It won't give betting odds though. It&#39

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/10/18 13:08:03