consensyx.eth

smart contractor

[Kai @hqmank]

Codex now supports non-OpenAI models.

No hacks needed anymore.

You can use local models with Ollama / LM Studio, or plug in third-party model APIs directly.

[Tibo @thsottiaux]

Reminder that you can use the Codex App, CLI and SDK with any open source model, not just with OpenAI models.

https://t.co/spPifB4ck3

[Charlie Marsh @charliermarsh]

At OpenAI, we're continuing to bet on Rust as the future of systems programming.

I'm proud to announce that we're making a $600,000 commitment to the Rust Foundation, which combines our Platinum membership with additional support for maintainer efforts across the Rust ecosystem.

[Dan Liu @danliu]

i'm surprised that someone who joined a $2T company 25 years ago to serve as coo is only worth $1.3B

[vivek @itsreallyvivek]

https://t.co/ZhOgyq7Vgn

[Péter Szilágyi @peter_szilagyi]

The scary part about Anthorpic's Fable nerf is not that it refuses to answer biology or cryptography. It's that it foreshadows what's coming. A world where a couple companies decide what you can and cannot do. They're building a new ruling class and you're not in it...

[Ben Dickson @bendee983]

1- AI-generated code just creates more technical debt.
2- At the end of the day, you (the developer) are responsible for the code, not the LLM that generated it. So the less code you have, the easier it is for you to own it.
3- A good engineer knows what code to write, and equally importantly, what code not to write or to delete.

(BTW, this is what we old-school software engineers have been saying for a long time, and we've been called all sorts of names for it.)

[Duca @big_duca]

Is anyone actually using loops with AI?

Sounds like a really effective way to funnel money out of your pocket.

And into the AI labs.

[Arc @arc]

Quantum computing introduces long-term risk for digital infrastructure, from wallet signatures to validator integrity and more.

Circle’s post-quantum whitepaper explores Arc’s phased approach to resilience across:

→ USDC
→ Smart contracts
→ Validators
→ Infrastructure systems

Planning for long-term security and institutional adoption.

https://t.co/Kt8d1aPIKz

[Leo⟠ @NPC_Leo]

今天恰好见了国内最顶尖的一个形式化验证教授，他说只从 Claude4.6和gpt 5.4 之后，形式化验证一下子就成了真正有潜力大规模使用的技术，原来给航天航空/操作系统内核做的证明，需要20人年的项目，4.6 之后直接能到2人月。但是攻击/防守是不对等的，形式化验证也是在爆发式的利用AI 加速。

[GoPlus中文社区 @GoPlusZH]

1/ Zcash Orchard 电路缺失约束致 zcash:native 无限增发漏洞分析

一、漏洞根本原因

漏洞位于 halo2 库的变基标量乘法电路（variable-base scalar multiplication gadget）中，具体是 https://t.co/zOi84aERBh 的第 309-310 行：

[Certora @Certora]

We’re very proud of our work with @saturn_credit, who chose us to formally verify the contracts for USDat and sUSDat tokens.

The collab really paid off: A critical underflow bug was caught before it could reach users.

More details 🧵

[consensyx.eth @consensyx]

傻逼才会买Minimax的coding plan

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

Minimax has just given me the most braindead, 8B-tier explanation of my test joke, and then when I said "Naive", hit me "Whatever you say" without thinking.
nah, Western frontier can sleep soundly.
Maybe it really is that good on KernelBench tho

[consensyx.eth @consensyx]

同意

[中国政经事实ChinaFacts @ChinaMacroFacts]

花真金白银用过形形色色的国产模型之后，现在很确定真正能有生产力的就是智谱（coding）和deepseek（中文写作）。

minimax我充了年费会员，但至少2.7属于倒给我钱都不用的档次，就算我不了解minimax，上点年纪的人还不了解商汤是个什么样的公司吗？

[squid @squidrouter]

This incident is unrelated to Squid’s core protocol and contracts. All Squid users and integrators are unaffected and no action is needed.

A third-party Gnosis Safe module was exploited today across Base and Ethereum, resulting in approximately $3.2M in losses. The vulnerable contract is verified on Basescan under the name “SquidRouterModule” but this contract was not built, deployed, or operated by Squid. It is a third-party smart-wallet product that chose to integrate with Squid, among other protocols, but has not been in contact with us.

The exploit worked because the third-party module accepted a caller-supplied constant string as proof that a message was secure. If you pass in this string (which is publicly available in the verified contract’s code), then you can execute an array of arbitrary calldata, stealing funds at will. The victims’ Safes had added this faulty contract as a trusted Safe Module, which gives the contract the ability to spend any tokens in the Safe without signatures. Squid’s own router (0xce16F69375520ab01377ce7B88f5BA8C48F8D666) is architecturally different and was not touched. Squid user funds, approvals, and integrations are fully secure.

Early public reporting may reference “SquidRouter” due to the contract’s verified name on Basescan. The accurate framing is: a third-party SquidRouterModule was exploited, not Squid’s Router contract. The contract shares our name but is not our code. We are monitoring the situation and will share updates if anything changes materially.

[qinbafrank @qinbafrank]

细读华为何庭波的署名论文来理解“韬（τ）定律”，从这篇论文能看到“韬（τ）定律”在缩放时间理论上的五个核心要点，但是越细看你越会发现华为、英伟达、台积电几家在未来的演化迭代路径上的底层逻辑是一致的。详细聊聊我的理解。
1、先说“韬（τ）定律”的五个核心要点
1）LogicFolding / 逻辑折叠。
文中给出的定义是：LogicFolding 是一种设计方法，把数字电路、模拟电路和存储电路划分到垂直堆叠的有源层中，通过缩短关键路径走线来优化性能、功耗和面积。它不是简单把封装做厚，也不是普通 2.5D Chiplet，而是更接近“把原本二维平铺的逻辑电路折到垂直方向

2）Unified Bus统一总线，
传统 AI 集群需要多层协议栈：PCIe、NVLink 或专有互联、以太网或 InfiniBand、RDMA、软件消息传递等。每一层协议转换都会增加序列化、DMA 缓冲、握手和延迟。

Unified Bus 的目标是用一个在机箱内和机箱间都运行的单一协议，替代这些协议栈，并在整个系统中暴露原生内存语义。论文称，这能把远程访问延迟从传统协议栈的几十微秒级降到约 100 纳秒，沿主要通信轴线的系统 τ 降低约 500 倍；在机架规模上，使系统更接近一台结构一致的单一机器，内部称为 System-as-One-Chip / 系统即芯片。这其实就是芯片—服务器—机柜—数据中心级别的系统协同。

3）Hi-ONE近封装光引擎
何庭波的论文认为，当单个 AI 芯片带宽进入 Tb/s 级别时，铜缆布线会遇到体积、SerDes、散热、供电和可靠性限制。华为提出的Hi-ONE是一种近封装光互连节点引擎。文中称单 Hi-ONE 模块可提供 8 Tb/s 带宽，把 SerDes 传输距离从约 100 厘米缩短到约 5 厘米，同时把面板间传输距离从不足 1 米扩展到 100 米。

这说明华为的系统级路线也把光互连纳入核心架构，不只是外部光模块，而是更接近 near-package optical I/O / 近封装光 I/O。

4）3D Folding
论文认为，传统 2.5D AI 芯片存在一个几何矛盾：逻辑芯片面积按 N² 增长，所以计算能力按面积增长；但 HBM、SerDes、供电等资源主要沿封装边缘进入，带宽、I/O 和电源能力只按周长 N 增长。计算能力按 N² 扩，边缘资源按 N 扩，二者差距会越来越大。

3D Folding的作用是把原来位于边缘的资源转移到表面或垂直方向：电源通过背面供电和集成电压调节器，高速存储通过与逻辑混合键合，光 I/O 通过近封装 Hi-ONE，从而让内存、互连、电源和逻辑同步扩展。

华为把3D折叠视为 AI 加速器在 2030 年后继续扩展的核心拓扑。

5）逻辑与内存重新融合
论文还有一个重要产业判断：过去几十年，CPU 和内存是解耦发展的；但 AI 时代正在逆转这种解耦。

原因是AI工作负载对内存带宽、延迟、功耗和封装的要求极高。HBM、混合键合、3D 堆叠 SRAM 都说明：数据传输和计算本身一样重要，逻辑和内存正在重新走向紧密物理集成。论文进一步判断，随着逻辑和内存融合，供应链中的影响力会向内存和封装供应商倾斜。

2、论文还列举了“韬（τ）定律”未来的挑战在于：
1）EDA 工具链不够。
传统 EDA 是二维时代的工具，主要在面积、时序、功耗之间优化。全尺寸 LogicFolding 需要工具把多个堆叠芯片视为一个连续设计实体，在单元粒度而不是模块粒度上跨层布局，并对垂直互连、TSV、KOZ 排除区、晶圆间工艺偏差做统一签核。论文称华为已有初步内部工具，但 τ 原生、开放、多物理场、3D 原生工具链仍是未来十年最重要的赋能投资之一。

2）晶圆间工艺偏差。
LogicFolding 可能把不同批次甚至不同节点的晶圆进行键合，阈值电压、驱动电流、互连 RC 偏差会影响时钟分布和保持时间裕量，需要智能冗余、自适应补偿和 τ 感知签核。

3）垂直互连不是免费的。
混合键合和 TSV 都会带来电阻、电容和面积开销，TSV 的 keep-out zone 还会挤占标准单元，所以逻辑折叠必须证明“缩短水平连线获得的收益”大于“增加垂直互连的成本”。

4）τ 是时间定律，不是能耗定律。
论文承认，如果速度提升 10 倍但功耗也提升 10 倍，系统仍可能超出电力约束。因此，τ 缩放必须和能量优化配套，包括近封装/共封装光学、背面供电、内存内计算、动态电压频率调节等。

整体上从这篇论文能看出，“韬（τ）定律”是体系化设计创新 + 三维集成 + 封装/键合工艺，而不是传统意义上的“制程节点突破”。“韬定律”更像是从单点工艺创新，转向体系化、系统级提升。

在先进制程受限、摩尔定律经济性下降的背景下，华为提出一种后摩尔时代的系统级缩放路线：以 τ 时间常数为统一指标，通过 LogicFolding、Unified Bus、Hi-ONE、3D Folding、逻辑—内存融合和 τ 原生 EDA，把性能提升从单点制程竞争转向全栈系统工程。

3、为什么说华为、英伟达、台积电几家在未来的演化迭代路径上的底层逻辑是一致的？
从个人角度这和台积电先进封装、英伟达NVLink/HBM/CPO/AI Factory 的方向本质相通。
大家都在解决同一个问题：数据移动太慢、太贵、太耗电。

这其实是全球半导体巨头共同迈向“后摩尔时代”的终极共识。 无论是华为的“韬定律”，还是国际巨头的动向，都在整体系统上下功夫：
1）台积电： 早就意识到先进制程太贵且良率存在物理瓶颈，因此大力发展 CoWoS 和 SoIC 等先进封装技术，像搭积木一样把多个小芯片（Chiplet）拼在一起。

2）英伟达： 现在的 AI 算力怪兽（如 Blackwell 架构及后续产品），其优势不仅在于单颗 GPU 核心的制程，更在于它通过 NVLink 高速互连技术，把海量的高带宽内存（HBM）和光芯片高度集成在一起，打破了“内存墙”和“通信墙”。

3）华为： 面临外部环境的极限施压，必须在缺乏最尖端制造设备的情况下，依靠先进封装、新材料、光电共封装（CPO）和极其强大的系统工程能力，来硬生生“拼”出等效于 1.4nm 的综合性能。

当然差异也是有的。
例如台积电的先进封装回答的是：“我如何把多个 die、HBM、chiplet、硅中介层、RDL、混合键合做成可量产、可测试、可良率控制的产品？”
华为“韬定律”回答的是：“在制程缩微受限时，我如何从器件、电路、芯片、系统全链路降低 τ，让性能、能效、密度继续提升？”

两者本质相通，因为都在解决：数据搬运太慢、太耗电、太占面积。但它们不是同一层级的东西。台积电更像是底层制造/封装能力平台，华为更像是系统架构与设计方法论。

再看英伟达这几年最典型的路线，就是“不只做 GPU，而是做整套 AI 计算系统”。英伟达 GB200 NVL72 就不是单颗 GPU 的故事，而是 rack-scale 架构：72 颗 Blackwell GPU、36 颗 Grace CPU，通过 NVLink 组成一个 72-GPU domain，对外表现得像一个巨大的 GPU，并通过 NVLink Switch 提供 130TB/s 的低延迟 GPU 通信带宽。

英伟达这套模式和“韬定律”的系统级思路非常接近：不要只看单颗芯片峰值算力，而要看 GPU—GPU、GPU—CPU、GPU—内存、机柜—机柜之间的数据移动效率。

4、投资逻辑：后摩尔时代系统级工程路线的基础设施。
华为“韬定律”不是单一芯片制造突破，而是后摩尔时代的系统级工程路线。它利好的不是单一晶圆厂，而是先进封装、探针测试、EDA、设备材料、高速互连、光互连、散热、电源、系统软件这一整套基础设施。

为什么？因为一旦目标从“晶体管做小”变成“路径做短、系统更快”，产业链价值就会从单点晶圆制造扩散到：
先进封装：把 chiplet、HBM、逻辑芯片靠近；
探针卡/测试设备：多 die、多层、复杂封装对测试要求更高；
封装基板/PCB/连接器：高速信号完整性更关键；
光模块/硅光/CPO：板级和机柜级数据传输从电走向光；
EDA/IP：二维布局不够，需要 2.5D/3D/封装/热/功耗协同设计；
散热/液冷/电源：集成度越高，热和供电越难；
系统软件/总线/互联协议：硬件堆起来还不够，调度和通信协议也要重构。

所以投资上，要去找：谁能让芯片之间、芯片内部、芯片与内存、服务器与服务器之间的数据路径变短，谁就能在产业链上占据优势。

未来先进性更多看“数据走多远、走多快、耗多少电、系统能否协同”。华为“韬定律”、台积电先进封装、英伟达 AI Factory，本质上都在围绕这个问题做文章

[XinGPT🐶 @xingpt]

https://t.co/GULsnZ7dIc

[Huawei @Huawei]

HUAWEI has presented the Tau (τ) Scaling Law, a new principle for guiding the future development of the semiconductor industry. By 2031, HUAWEI's high-end chips based on this law are expected to feature a transistor density that is equivalent to 14 Å (1.4 nm) processes.

[Chubby♨️ @kimmonismus]

Chinas Huawei can't access EUV. So they wrote their own scaling law. The leverage of US export controls erodes.

Huawei just presented the Tau (τ) Scaling Law at IEEE ISCAS, a framework that replaces geometric transistor scaling with time-based optimization across devices, circuits, chips, and systems.

Huawei is not trying to win the nanometer race anymore. They're redefining it.

381 chips designed and mass-produced over six years. Kirin chips with their new LogicFolding architecture ship this fall. Target by 2031: transistor density equivalent to 1.4nm processes - without EUV (ASML embargo still hits hard)

Whether Tau Scaling delivers on that promise remains to be seen. But it shows one thing: US export controls cut Huawei off from cutting-edge lithography, and instead of hitting a wall, they built a parallel road.

China's semiconductor independence isn't hypothetical anymore. It's shipping in phones, running AI workloads, and now has its own scaling law presented at one of the world's top circuit design conferences.

If Huawei can close the performance gap through architecture instead of lithography, the entire leverage of US export controls erodes. The sanctions were designed to keep China two generations behind. Huawei is trying to make that metric irrelevant.

h/t @AndrewCurran_ for finding this interesting blogarticle and @zephyr_z9 for the photo

[GDP @bookwormengr]

https://t.co/KHnPhxJBiz