Wonder News
A concise morning read on AI-generated games, model releases, creator tools, research, and education signals worth tracking. Page 4.
Newsletter Agogic cut Fréchet Music Distance from 272–286 to 159 with performance-timed tokens, while a 34-fold model-size increase barely moved the metric. Also: video editing, layered design, agents, robots, tutoring, and UPCEA 2027.
Read article →
Newsletter Qwen-CUA operates software through screenshots, keyboard, and mouse events. Also: world reactivity, shared coding, agent repair, video efficiency, robot data, programming practice, and the National ESEA Conference 2027.
Read article →
Newsletter DungeonBench tests AI policies across 561 combatants, 229 spells, and 25 tactical scenarios. Also: hidden Minecraft rules, video compositing, code agents, robot world models, and Game On Expo 2027.
Read article →
Newsletter Microsoft opened Project Perception with MAI-Cyber-1-Flash and says its MDASH configuration scored 96% on CyberGym. Also: secure code generation, prompt audits, agent training, replay fidelity, visual models, and SecurED 2027.
Read article →
Newsletter DeepSeek says V4-Flash-0731 beat V4-Pro Preview in all nine agent and coding benchmarks it reported, though the tests used an unreleased harness and included two internal evaluations.
Read article →
Newsletter PAC-MAN reports a 95% dodge rate on a real Unitree G1 using onboard depth vision. Also: embodied data, physical world models, longer video generation, data-efficient 3D creation, coding-agent tests, and AASA NCE 2027.
Read article →
Newsletter In two AI-research trials, frontier agents had six days and thousands of dollars of compute but failed to answer the central questions. Also: long video, 3D cameras, embodied benchmarks, robot policies, coding agents, memory security, and ASU+GSV 2027.
Read article →
Newsletter Wonder reports minute-scale, camera-controlled video worlds at 16 FPS. Also: search-free world-model control, image-based driving, faster video generation, GUI and code benchmarks, agent memory, robot policies, 3D streaming, and a new tabletop expo.
Read article →
Newsletter DreamStyle3D reports native stylized 3D asset generation within 10 seconds. Also: camera-controlled video, film benchmarks, avatars, agent memory, code repair, tactile world models, robot evaluation, and PAX East 2027.
Read article →
Newsletter In nearly 6,000 office-automation runs, some procedural skills made LLM agents fail tasks they had solved without the skill. Also: 3D-scene action, persistent world-model sessions, robot rendering, tactile rollouts, 4D reconstruction, generated video, interfaces, tutoring robots, and SXSW 2027.
Read article →