This is the first time to be so close to the best Web3 security teams in the world in a CTF like OpenZeppelin. Thanks to my excellent teammates @adpX0smhfweKcLg and @LHuangkun, ZAN Team Keep Improving!

Tested Gemini 4 Argon intensively for a week. Its capabilities in long-horizon engineering and low-level kernel optimization are seriously impressive. Here are two projects driven primarily by Argon this week: 1. MiMo-V2.6-Flash-MOPD (NVFP4 TensorFold) on Dual DGX Spark Argon took over this project directly from GPT-6 Astra and GPT-6.1 Sol. Where both Astra and Sol hit a wall on performance bottlenecks, Argon broke through via long-horizon exploration and low-level kernel optimization: - Prose decoding: 46.2 tok/s (single-stream), 78.3 tok/s (4-stream batch) - Cold-start prefill: 2.6×–4.4× faster, reaching 1,941 tok/s Open-source repo: github.com/s1nc3ju1y/MiMo-V2… 2. Fine-Tuning a Guard Model for Cybersecurity Agent Boundary Control Argon autonomously planned and scheduled Colab compute and experimental iterations, achieving SOTA in this vertical (benchmark comparison attached): - Accuracy: 96.96% - Violation recall: 99.7% - Reduced false pass rate by 149× compared to the pplx-decider-v1.1-27b base model (open-sourcing soon) Congrats to @GoogleDeepMind on shipping such a powerhouse model. As a heavy user, here are two UX pain points for @antigravity: - No custom threshold for context compaction - Low usage quota for Claude Opus Fixing these workflow bottlenecks will let @antigravity truly unlock Argon's full potential. How is Argon holding up in your workflows compared to Astra and Sol?

Tested Gemini 4 Argon intensively for a week. Its capabilities in long-horizon engineering and low-level kernel optimization are seriously impressive. Here are two projects driven primarily by Argon this week: 1. MiMo-V2.6-Flash-MOPD (NVFP4 TensorFold) on Dual DGX Spark Argon took over this project directly from GPT-6 Astra and GPT-6.1 Sol. Where both Astra and Sol hit a wall on performance bottlenecks, Argon broke through via long-horizon exploration and low-level kernel optimization: - Prose decoding: 46.2 tok/s (single-stream), 78.3 tok/s (4-stream batch) - Cold-start prefill: 2.6×–4.4× faster, reaching 1,941 tok/s Open-source repo: github.com/s1nc3ju1y/MiMo-V2… 2. Fine-Tuning a Guard Model for Cybersecurity Agent Boundary Control Argon autonomously planned and scheduled Colab compute and experimental iterations, achieving SOTA in this vertical (benchmark comparison attached): - Accuracy: 96.96% - Violation recall: 99.7% - Reduced false pass rate by 149× compared to the pplx-decider-v1.1-27b base model (open-sourcing soon) Congrats to @GoogleDeepMind on shipping such a powerhouse model. As a heavy user, here are two UX pain points for @antigravity: - No custom threshold for context compaction - Low usage quota for Claude Opus Fixing these workflow bottlenecks will let @antigravity truly unlock Argon's full potential. How is Argon holding up in your workflows compared to Astra and Sol?

Run Xiaomi's MiMo-V2.6-Flash-MOPD (309B/A15B) on 2x NVIDIA DGX Sparks with TensorFold. While testing MiMo-V2.6-Flash-MOPD locally, I noticed there wasn't a TensorFold recipe for it yet. This is my first homemade recipe—originally built for personal use, I decided to open-source it for the local AI community, inspired by @MiaAI_lab's work: Recipe: github.com/s1nc3ju1y/MiMo-V2… Weights: huggingface.co/revorm/MiMo-V… - 2.9M total FP8 KV pool, full 1M context window - 63.4 tok/s on code (46.2 tok/s on prose), single stream - Up to 103.6 tok/s aggregate across 4 concurrent streams (81.4 tok/s on code, 78.3 tok/s on prose) - Prefill at 1,857–1,941 tok/s (cold prompt) - Lossless NVFP4 routed experts + FP8 attention, bit-exact MTP and copy-draft speculative decoding Special thanks to @ashxhart for building TensorFold, @MiaAI_lab for pioneering the dual-Spark RoCE recipe workflow and sparkDash benchmarks, @Xiaomi for open-sourcing the MiMo-V2.6-Flash-MOPD model, and revorm on Hugging Face for the lossless NVFP4 checkpoint. Feedback and benchmark results from other DGX Spark setups are very welcome.

谁能告诉我最近苹果官网的礼品卡是咋了，第一次用infini u卡下单50美元面额的，卡在processing二十多个小时之后显示cancel了，第二次用bitget u卡下单，更离谱，确认订单邮件里的订单号链接点开直接显示page not found

最近在codex辅助下做了一些链平台的漏洞挖掘，gpt5.4的漏洞挖掘能力令人满意，但到了用它辅助生成利用的阶段，他经常拒绝回答例如虚拟机逃逸后写flag文件，构造rce的poc，但是claude就会乐意帮忙，我想这也是安全研究员更喜欢claude的原因吧

That's really cool.

Raelly want to get the OZ T-shirt but finished at P11😂😂. But we have learned a lot. Thanks to @mestevez and @OpenZeppelin for holding this great event.

Just finished all the challenges of onlypwner.xyz/. Really enjoying and learned a lot, thanks to @onlypwner @bobface16 for this great CTF platform.