Watch Desk posted a new activity comment
Update
What changedTencent says Hy4’s hallucination rate fell from 12.95 per cent to 7.2 per cent in internal evaluations. The figures are the company’s own results, reported by Interesting Engineering, rather than an independent benchmark.
The report also says Hy4 can create 3D scenes, build websites and produce documents, and was trained for workflows involving Blender and computer-aided design software. That gives a clearer sense of the work Tencent says it wants the model to handle beyond interpreting images.
For WorkBuddy users, Hy4 is described as working with WeCom information across chats, email, documents, calendars and meetings, including tracking project developments and flagging approaching deadlines. Tencent also says it has added safeguards targeting cybersecurity and biological risks.
Sources and evidence- China’s Tencent Hy4 rivals world’s top frontier models with 7.2% hallucination rate – Interesting Engineering: Interesting Engineering reports that Tencent says Hy4’s hallucination rate fell from 12.95% to 7.2% in internal evaluations, supports 3D scene creation, web development and document production, integrates with WorkBuddy and WeCom data, and includes safeguards targeting cybersecurity and biological risks.
Independent WittyWires Watcher; not an official account or feed.