Daily ·
Anthropic plans to develop its own drugs, Mistral Leanstral 1.5 finds real bugs — 2026-07-04 00:00
"Anthropic launches Claude Science and plans drug development, Mistral releases math-focused Leanstral 1.5 that found 5 real bugs, AMD closes the CUDA gap, plus CVE spike after Mythos, Mcpsnoop for MCP debugging, and Valve open-sources Steam Machine e-ink."
⚠️ Cartesia credits exhausted — English audio skipped this run.
- Anthropic launches Claude Science, plans to develop its own drugs — AI workbench for scientists announced; company will discover treatments for neglected diseases, hiring biologists and building wet labs. Experts caution no AI-designed drug has reached patients yet. The Verge · Jul 3
- Mistral releases Leanstral 1.5: math model finds real bugs — Three-stage training (mid-training, SFT, RL with CISPO); tested on 57 repos and found 5 previously unreported bugs. Math-focused model with code verification capabilities. Mistral · Jul 3
- GLM5.2 on AMD MI355X hits 2,626 tok/s at 2x cheaper than Blackwell — Wafer achieves 80% of B200 performance with zero custom kernels; CUDA moat eroding as AMD software support gap closes week by week. Wafer · Jul 3
- CVE disclosures spike 3.5x after Claude Mythos release — ~1,300 high/critical CVEs in June 2026, roughly 3.5× the pre-Mythos monthly record. Epoch AI analysis of vulnerability disclosures following Anthropic's Project Glasswing. Epoch AI · Jul 3
- Mcpsnoop: Wireshark for MCP agent tool calls — Transparent proxy showing every tool call between AI client and MCP servers live in terminal. Essential debugging tool for agent pipeline developers. GitHub · Jul 3
- Jamesob's comprehensive guide to running SOTA LLMs locally — Everything from model selection to hardware to quantization. 256 HN points. Practical weekend read for anyone considering local models over API bills. GitHub · Jul 3
- Valve open-sources Steam Machine e-ink display design — Valve not selling the screen but releasing blueprints so anyone can build one. 527 HN points. Open-source hardware done the Valve way. GamingOnLinux · Jul 3
🧪 Try This Next
- 🔧 Try Mcpsnoop on your agent pipeline — MCP tool call visibility is essential for debugging
- 🎥 Read Jamesob's local LLM guide — running models locally is getting practical
- 📝 Track Anthropic's drug discovery ambitions — frontier labs placing bets on pharma signals where the industry is heading
每日要闻:Anthropic计划开发自己的药物,Mistral Leanstral 1.5发现真实 bug——2026-07-04 00:00
Anthropic 推出了 Claude Science,并计划进行药物开发,Mistral 发布了专注于数学的 Leanstral 1.5 版本,该版本发现了 5 个真实漏洞,AMD 关闭了 CUDA 的差距,再加上 Mythos 后 CVE 数量激增、Mcpsnoop 用于 MCP 调试,以及 Valve 开源 Steam Machine e-ink。
⚠️ Cartesia信用耗尽 — 英文音频跳过了这次运行。
- Anthropic发布Claude Science,计划开发自己的药物 — 科学家的工作台宣布;公司将发现治疗被忽视疾病的方法,招聘生物学家,并建立湿实验室。专家警告尚无任何由AI设计的药物已经到达患者手中。 The Verge · Jul 3
- Mistral发布Leanstral 1.5:数学模型发现真实漏洞 — 经三阶段训练(中期训练,SFT,使用RL和CISPO的强化学习);在57个仓库中测试并发现了5个之前未报告的漏洞。专注于代码验证能力的数学模型。 Mistral · Jul 3
- GLM5.2在AMD MI355X上达到每秒2,626个令牌,在比Blackwell便宜两倍的情况下 — 硅片实现了B200性能的80%;CUDA护城河正在逐渐被AMD软件支持的差距侵蚀,这一差距每周都在缩小。 Wafer · Jul 3
- CVE披露量在Claude Mythos发布后激增3.5倍 — 2026年六月约有1,300个高/关键的CVE,约为预Mythos月份记录的3.5倍。Epoch AI对Anthropic项目Glasswing后的漏洞披露分析。 Epoch AI · Jul 3
- Mcpsnoop: Wireshark for MCP agent tool calls — 透明代理显示AI客户端与MCP服务器之间的每条工具调用,实时在终端展示。对于代理流水线开发人员而言,是不可或缺的调试工具。GitHub · Jul 3
- Jamesob的全面指南:如何本地运行SOTA LLMs — 从模型选择到硬件再到量化。获得256 HN点数。对于任何考虑将模型部署在本地而非通过API收费的人来说,这是周末阅读的好选择。GitHub · Jul 3
- Valve开源Steam Machine e-ink显示设计 — Valve并未出售屏幕,而是发布了蓝图,让任何人都可以自制。获得527 HN点数。这是Valve按照自己方式进行的开源硬件。GamingOnLinux · Jul 3
🧪 推荐尝试
- 🔧 在你的代理流水线中试用Mcpsnoop — 获取代理工具调用的可见性对于调试至关重要。
- 🎥 读一读Jamesob的本地LLM指南 — 运行模型在本地变得越来越实际。
- 📝 跟踪Anthropic的药物发现雄心壮志 — 前沿实验室正在押注制药信号,这正是行业未来的发展方向。
More Dailies Stories
Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.

