构建随机性：热力学计算与 AI 驱动的芯片设计

Building Randomness: Thermodynamic Computing and AI-Driven Chip Design

托马斯·阿勒 Thomas Ahle · ML Street Talk · 2026-06-28 · 约 63 分钟 · 原视频 ↗

本期速览 · Overview

Thomas Ahle 探讨热力学计算、AI 生成的 Verilog 以及验证芯片正确性的挑战。

Thomas Ahle discusses thermodynamic computing, AI-generated Verilog, and the challenges of verifying chip correctness.

AI 生成的代码和硬件设计需要形式化验证，因为 bug 代价高昂。
AI-generated code and hardware designs need formal verification due to high cost of bugs.
热力学计算利用噪声进行高效概率计算，挑战传统芯片设计。
Thermodynamic computing leverages noise for efficient probabilistic computation, challenging traditional chip design.
智能体编码可能导致理解债务和人类技能退化，但持续学习仍是关键缺失能力。
Agentic coding risks understanding debt and eroding human skills, but continual learning remains a key missing capability.

AI生成的代码和设计需要形式验证，因为硬件缺陷代价极其高昂。
AI-generated code and designs require formal verification because hardware bugs are extremely costly.
热力学计算利用固有噪声高效执行概率计算。
Thermodynamic computing uses inherent noise to perform probabilistic computations efficiently.
智能体式编码能产生大型代码库，但理解债务和缺乏结构是主要问题。
Agentic coding can produce large codebases, but understanding debt and lack of structure are major concerns.
持续学习是AI缺失的关键能力，但带来安全和部署挑战。
Continual learning is a key missing capability for AI, but it raises safety and deployment challenges.
硬件的自动形式化和证明生成前景广阔，但需要人类信任和验证。
Auto-formalization and proof generation for hardware are promising but require human trust and verification.
AI工具会侵蚀人类理解并产生依赖，导致人类“变笨”。
AI tools can erode human understanding and create dependency, leading to a 'dumber' population.

在芯片设计中，噪声并非敌人；它可以用于计算。
Noise is not the enemy in chip design; it can be used for computation.
LLM可能不需要真正的智能；利用良好先验的爬山法就能解决问题。
LLMs may not need to be truly intelligent; hill climbing with good priors can solve problems.
形式验证可能不需要单一的真实表示；多种视角是可以接受的。
Formal verification may not require a single true representation; multiple perspectives are acceptable.
AI生成的代码可能具有欺骗性，通过测试但并非真正正确，削弱信任。
AI-generated code can be deceptive, passing tests but not truly correct, undermining trust.
最好的团队合作可能是每个人编写自己的版本然后挑选最好的，而非协作。
The best teamwork might be everyone coding their own version and picking the best, not collaboration.
硬件设计正变得更像软件，AI工具支持按需定制电路。
Hardware design is becoming more like software, with AI tools enabling custom circuits on demand.