Researchers built a finance-specific prediction harness to test whether frontier AI can beat human analysts at forecasting quarterly earnings. The evaluation ran seven high-capability models (GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5, GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.8 Flash, Kimi K3) on 456 earnings instances for companies with >$5B market cap and ≥8 broker estimates. Models made one-week-ahead forecasts for four headline metrics (revenue, gross margin, operating income, adjusted EPS) using a point-in-time gated data stack, time-aware retrieval, substituted web access, and expert-guided prompts. Performance was measured by normalized prediction error (adjusted for each company’s surprise volatility), surprise correlation (Spearman ρ versus actual − consensus), and hit rate; a bias-corrected consensus baseline (adding each company’s median historical surprise) provided a stronger expert comparison.
Results show a capability inflection: all tested models beat raw analyst consensus, and the latest frontier models (Fable 5.1, Opus 5.5, GPT-6 Astra) outperform the bias-corrected consensus, with GPT-6 Astra leading on overall error and hit rate while Fable and Opus excel on surprise correlation. Traces reveal models gather evidence, quantify event impacts, and revise away from consensus, driving gains especially on below-consensus calls. An ablation study attributes most of the accuracy improvement to access to up-to-date data, demonstrating that frontier models plus a finance-tailored harness achieve superhuman performance on earnings prediction.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Four coding agents were given a $100 budget to build a functional PDF editor, but their implementations had noticeable bugs and shortcomings. The experiment highlights that current AI agents struggle with goal pursuit and user interaction, leading to subpar software quality. (blog.nielstron.de)
Lithos-ai open-sourced lithos-metal, enabling Qwen3.8-27B to run at over 200 tokens per second on an Apple M5 Max laptop. This allows ultra-fast inference for coding agents with a single command. (twitter.com)
Google Cloud introduces the Gemini agent, a universal AI tool that handles questions, content creation, coding, and knowledge work within familiar applications like Gmail and Docs. It features industry-specific skills, security, cost controls, and integrates seamlessly into existing workflows. (cloud.google.com)
Temporal developed pi-temporal to enable the Pi coding agent to recover from machine failures during tasks by leveraging a fleet of workers and durable execution. This setup allows seamless task continuation across machine failures without repeating interrupted commands. (temporal.io)
![OpenAI cannot make AI safe on its own [pdf]](https://hntoday.pixelshot.io/screenshots/20261008/0217d5d1410143c5_tn.jpeg)
OpenAI has acknowledged that it cannot ensure the safety of artificial intelligence systems on its own. The document discusses the challenges and limitations faced by OpenAI in making AI safe without external collaboration or oversight. (mikitabalesni.com)
Yoshua Bengio urges AI researchers at frontier companies to prioritize safety and consider leaving if safety measures are insufficient. He warns that current development races risk pushing humanity toward catastrophic outcomes and calls for slowing AI progress to prevent loss of control. (transformernews.ai)
Today's best Hacker News stories, summarized and screenshotted, one email a day.