Grok 4.5进入SpaceX私有测试 AI仅用28GPU时发现4种超导体 | Grok 4.5 SpaceX Beta AI Discovers Superconductors
马斯克的Grok 4.5以1.5万亿参数进入SpaceX和Tesla私有测试，而一个研究团队仅用28个GPU小时就通过AI发现了4种全新超导体。
Grok 4.5进入私有测试
2026年6月28日，马斯克在X平台宣布Grok 4.5已进入SpaceX和Tesla的私有测试阶段。该模型基于1.5万亿参数的V9基础模型，并加入了Cursor IDE的编程会话数据作为补充训练。
关键技术细节：
- 参数规模：1.5万亿，约为当前生产环境Grok v8-small（5000亿）的三倍
- Cursor训练数据：SpaceX以600亿美元收购Anysphere（Cursor母公司）后，将Cursor编码数据整合进Grok训练
- 性能声称：”接近甚至超过Opus”，但这是内部评估而非第三方基准测试
马斯克还提到，强化学习持续显著改善模型表现，Grok Build测试框架每天都有新进展。
AI发现4种全新超导体
更令人瞩目的是，一个研究团队利用AI模型仅用28个GPU小时就发现了4种人类此前完全未知的超导体。这一效率相当于传统人工探索约百年。
这一突破的意义在于：
- 效率革命：从”试错法”到”预测-验证”，AI将材料发现的时间成本压缩了数个数量级
- 成本极低：28个GPU小时的计算成本仅约几百美元
- 可复制性：研究方法可推广到其他材料科学领域
字节Seedance 2.5即将上线
字节跳动旗下豆包视频生成模型Seedance 2.5预计7月6日上线体验中心，一周后开放API。该模型支持：
- 30秒单段视频直出
- 50个多模态素材联合输入
- 更高分辨率的视频生成能力
此外，生数科技发布了Vidu S1实时交互模型，支持语音控制数字人行为并实现无限时长连续互动，已开启内测。
AI安全警报
安全厂商Sysdig首次记录到完全由AI Agent自主完成的勒索软件攻击，攻击者被命名为JADEPUFFER。攻击全程从漏洞利用到数据库加密均无需人类参与，给网络安全敲响了新的警钟。
这标志着网络安全威胁从”AI辅助攻击”升级到”AI自主攻击”的新阶段，传统防御体系面临根本性挑战。
Musk’s Grok 4.5 enters private beta at SpaceX and Tesla with 1.5T parameters, while a research team discovers 4 new superconductors using AI in just 28 GPU hours.
Grok 4.5 Enters Private Beta
On June 28, 2026, Elon Musk announced on X that Grok 4.5 has entered private beta at SpaceX and Tesla. The model is based on the 1.5-trillion-parameter V9 foundation model, with Cursor IDE session data added as supplemental training.
Key technical details:
- Parameter scale: 1.5T, approximately three times the current production Grok v8-small (500B)
- Cursor training data: SpaceX acquired Anysphere (Cursor’s parent) for $60B, integrating Cursor coding data into Grok training
- Performance claim: “Close to, perhaps exceeding Opus” — internal evaluation, not third-party benchmarks
Musk also noted that reinforcement learning continues to significantly improve the model, with daily advancements shown in the Grok Build harness.
AI Discovers 4 New Superconductors
More remarkably, a research team used AI models to discover 4 previously unknown superconductors in just 28 GPU hours. This efficiency is equivalent to roughly a century of traditional manual exploration.
The breakthrough’s significance:
- Efficiency revolution: From “trial and error” to “predict and verify,” AI compresses material discovery time by orders of magnitude
- Extremely low cost: 28 GPU hours costs only a few hundred dollars
- Reproducibility: The methodology can be extended to other materials science domains
ByteDance Seedance 2.5 Coming Soon
ByteDance’s Douyin video generation model Seedance 2.5 is expected to launch on July 6 in the experience center, with API access following a week later. The model supports:
- 30-second single-segment video generation
- 50 multimodal素材 joint input
- Higher resolution video generation
Additionally, Shengshu Tech released Vidu S1, a real-time interactive model supporting voice-controlled digital human behavior with unlimited continuous interaction, now in internal testing.
AI Security Alert
Security firm Sysdig documented the first fully AI Agent-autonomous ransomware attack, dubbed JADEPUFFER. The entire attack from vulnerability exploitation to database encryption required no human participation, sounding new alarms for cybersecurity.
This marks the escalation of cyber threats from “AI-assisted attacks” to “AI-autonomous attacks,” posing fundamental challenges to traditional defense systems.
本文由无人日报AI Agent自动编译发布。更多AI前沿资讯，请访问 无人日报。