A live reference page lists release dates and training cutoffs for 20 recent large models from eight labs, with data checked on 2026-09-10 and available as models.json. Each entry shows a model name, the lab, its public release date and the date its training data stops; examples include LLAMA 4 (released Apr 5, 2025; cutoff Aug 2024), GPT-5.6 Sol and Luna (released Jul 9, 2026; cutoff Feb 16, 2026), and GPT-6 Astra (released Sep 3, 2026; cutoff Apr 30, 2026). Many entries list “Not established” where vendors did not publish a cutoff. The page uses JavaScript counters that tick upward from those dates so the displayed age stays current.
The core argument is that two dates determine how up-to-date a model actually is: release date and training cutoff, and the gap between them shows how stale a model already is at launch. Browsing or web-search tooling does not change a model’s training cutoff; it only supplies ephemeral context for single queries. Four labs (Anthropic, Google DeepMind, Meta and OpenAI) publish cutoffs for at least one model, and overall 9 of 20 models on the list carry a published cutoff. Practical advice: ask a model directly “what is your training cutoff date?” and verify its answer against this dataset before trusting time-sensitive outputs.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Barndoor AI has acquired Diaphora, the creator of the open-source workflow engine Frags, to enhance governed AI automation for enterprise workflows. The combined platform enables organizations to build repeatable, governed AI workflows that can be securely scaled across teams. (barndoor.ai)
Friday is an open-source platform that enables self-hosted persistent memory for AI coding agents, enhancing their efficiency and scalability. It integrates with various developer workflows and security tools to support AI-driven software development. (github.com)
UBTECH has opened a factory in China's Guangxi region capable of producing over 10,000 humanoid robots annually, with one robot made every 10 minutes. The plant uses robots and driverless vehicles for automation, mainly producing the Walker S and Cruzr series. (en.sedaily.com)
A 4-billion-parameter model called Qwen was trained using reinforcement learning to generate query plans that are 81% faster than those by Postgres. The model improves query optimization by learning from performance measurements and adjusting its strategies accordingly. (rohanbansal.com)
A reverse-engineered Jev-like model is available on GitHub, with recent updates and commits. The project includes code, documentation, examples, tests, and licensing information. (github.com)
A fleet of AI agents requires a new type of operating system that manages authority and ownership across multiple agents, rather than a simple harness for individual models. The proposed agent OS must coordinate shared authority over context, tools, execution, and governance to ensure reliable and secure multi-agent operations. (pentad.ai)
Today's best Hacker News stories, summarized and screenshotted, one email a day.