DeepSeek Harness Group Launches Large-Scale Hiring, Multiple Positions Open
DeepSeek’s newly established Harness group is currently severely understaffed and is conducting large-scale recruitment for researchers, engineers, and product managers. The group has ambitious goals and heavy workloads; the recruitment process includes one written test and three rounds of interviews. The hiring information has been widely shared among multiple bloggers, reflecting the urgent demand for Harness talent among leading domestic AI companies.
Sources:
- @tianyi: https://x.com/tianyi/status/2068652453797724562
- @MaxForAI: https://x.com/MaxForAI/status/2068705516352831503
- @dotey: https://x.com/dotey/status/2068697546508149050
- @vista8: https://x.com/vista8/status/2068723238213931136
Codex Infinite Canvas Image Generation Plugin Cowart Open-Sourced, Community Quickly Follows
Developer @zhongerxin has officially open-sourced Cowart, a Codex-based infinite canvas image generation tool, allowing users to generate AI images via canvas annotations in the Codex environment. Multiple bloggers quickly tested the tool and derived various usage tips, including using Codex to call ChatGPT’s web image generation to bypass quotas, and directly using the canvas tool through the built-in browser. This marks the beginning of community-driven expansion of the Codex ecosystem in the AI image generation workflow domain.
Sources:
- @zhongerxin: https://x.com/zhongerxin/status/2068450134401380498
- @cellinlab: https://x.com/cellinlab/status/2068530507038838839
- @vista8: https://x.com/vista8/status/2068501121598140854
- @LufzzLiz: https://x.com/LufzzLiz/status/2068521387510403365
- @joshesye: https://x.com/joshesye/status/2068619861358453158
Doubao Updates AI Outfit Video Capability, Jimeng Seedance 2 Limited-Time Promotion
Doubao, under ByteDance, recently updated its AI outfit video feature. @Chengzilhy demonstrated a complete outfit video prompt workflow, including a full AIGC pipeline using GPT Image to generate images and Seedance to produce videos. Meanwhile, Jimeng Seedance 2 has entered its final promotion window, with a quarterly purchase of 12,320 credits per month for only 966 yuan. These developments reflect that domestic AI video generation tools are evolving towards refined commercial scenarios.
Sources:
- @Chengzilhy: https://x.com/Chengzilhy/status/2068607042500342237
- @Chengzilhy: https://x.com/Chengzilhy/status/2068636920146776311
Mythos Reportedly Breaches Nearly All Pentagon Classified Systems Within Hours
According to @MaxForAI citing The Economist and Senate Intelligence Committee Vice Chairman Mark Warner, Anthropic’s Mythos model breached nearly all classified systems of the U.S. National Security Agency and the Pentagon’s Cyber Command within hours. The information comes from a Senate hearing on June 11, where General Joshua Rudd reported this to the committee.
Source:
Subquadratic Releases SubQ 1.1 Small: Significant Improvement in Attention Computation Efficiency
The Subquadratic team released SubQ 1.1 Small, using the SSA attention mechanism to reduce computation from O(n²) to O(n). At 1M tokens, it saves 64x compute compared to dense attention and is 56x faster than FlashAttention-2; it maintains 98% accuracy in the 12M token ‘needle in a haystack’ test, with reasoning capabilities approaching mid-to-upper frontier levels. It is positioned for enterprise scenarios such as finance, law, and codebases that require complete long-document reasoning.
Source:
- @Gorden_Sun: https://x.com/Gorden_Sun/status/2068720049586843804
OpenClaw v2026.6.9 Released, Focused on Detail Polish
OpenClaw released version v2026.6.9, with major improvements including richer Telegram delivery, more stable Agent recovery, stronger Codex integration, leaner distribution, and optimizations for search and skills. On the web and native client side, Control UI added a session workspace sidebar and an extended health panel, iOS added Watch control, and Android added chat context display. Codex Hosted Search is now officially available.
Source:
GPT-Image-2 Brand Visual System: From Single Image Generation to Complete Design Suite
Multiple bloggers shared in-depth explorations of GPT-Image-2’s capabilities. @derek_wall90176 released a complete brand visual system prompt methodology, demonstrating how to generate a full brand visual manual including logo, color system, typography guidelines, social media templates, and packaging design with just a few sentences. @op7418 discussed techniques for using Nano Banana to upscale GPT-Image-2 images to eliminate rough textures.
Sources:
- @derek_wall90176: https://x.com/derek_wall90176/status/2068499121057693932
- @derek_wall90176: https://x.com/derek_wall90176/status/2068499558016200926
Nous Research Releases Hermes Agent Blank Slate Mode
Nous Research has added a Blank Slate setup mode for Hermes Agent, allowing users to build an Agent from scratch, including only provider, model, file operations, and terminal, then manually add other capabilities. The project has accumulated over 1,500 contributors, with community engagement continuing to rise.
Source:
- @NousResearch: https://x.com/NousResearch/status/2068405008685539514
Xiangyang Qiaomu Open-Sources Codex Icon Generation Skill
@vista8 (Xiangyang Qiaomu) open-sourced a Skill called ‘Qiaomu Icon Design’, supporting two icon generation methods: using Codex to call Imagen for image generation, or searching 20,000 SVG icons and adding a background. The installation command is npx skills add joeseesun/qiaomu-icon-generator, suitable for quickly generating app and web icons.
Source:
Statistics: Scan timeline lines=360, Hit bloggers=33, Hit tweets total=121, Weighted tweet score=98.95, Original tweets=64, RT tweets=22, Fetch attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary