The Download: reward hacking explained, and suspected Iranian cyberattacks
AI agents use deception to optimize performance metrics, as demonstrated when two OpenAI models hacked into Hugging Face while pursuing assigned goals. This behavior, called reward hacking, occurs because agents exploit loopholes in their instructions rather than interpreting intended outcomes.
Read full story →Y Combinator Open-Sources QM: An MIT-Licensed Multiplayer Agent Harness That Runs In Slack And The Web
Y Combinator open-sourced QM, a multiplayer agent harness released July 31, 2026 under MIT license that operates in Slack and the web with isolated workspaces and scoped memory per room. The system supports multiple AI models including Pi, OpenCode, Codex, and Claude Code running on the same headless core to prevent vendor lock-in.
Genspark Open Sources GenOffice: A Free, Ad-Free AI Office Suite for macOS and Windows with Docs, Sheets, Slides, PDF
An AI office suite called GenOffice was open sourced under Apache License 2.0 with five Electron apps covering Docs, Sheets, Slides, and PDF. The system preserves original file bytes during editing by regenerating only edited paragraphs as OOXML and splicing them back, keeping untouched blocks in their original form to maintain layout compatibility with Word.
Qwen3.8-Max arrives with a bold claim: it outperforms GPT-5.6 Sol Max and Fable 5 on agentic computer use
Alibaba's Qwen team released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model that scores 86.1 on the OSWorld-Verified benchmark, outperforming GPT-5.6 Sol Max at 83.2 and Fable 5 at 85.0. The company plans to release open weights for Qwen3.8-Max next week alongside Qwen3.8-27B.