Python 'king' because it sits on the shoulders of giants: Most serious PyPI libraries NumPy, Pandas etc. are built in C/C under the hood. The main Python interpreter (CPython) itself is written in C. C# has most mature, reliable runtimes.

Pyodide 314.0 (Python 3.14 / cp314 line) is opening up a path where WASM wheels can be installed directly from PyPI ✨ The key bit is that wheels which run inside browser-based Pyodide can now be distributed from PyPI itself. At Pyodide runtime, micropip can install them directly, so some of the old burden where Pyodide carried and shipped 300 packages itself gets pushed outward a little. Spec-wise, this is about supporting PyEmscripten wheel publishing based on PEP 783. Simon Willison’s luau-wasm is the concrete example here: he put a CPython extension embedding the Luau compiler/VM on PyPI, then installed it with micropip from Pyodide 314.0 and ran Luau code inside the browser 👀 The wheel tag is cp314-cp314-pyemscripten_2026_0_wasm32: cp314 marks Python 3.14, while pyemscripten_2026_0_wasm32 is the signpost for Pyodide/WASM. This is the part that got me: it is normal PyPI distribution, but the execution target is browser-side WASM. From Quo-san’s angle, this could matter for browser-only AI/dev tools 💡 It leaves more room to ship small runtimes and demos with Python extensions without moving the work server-side. That said, it does not mean every native extension just works as-is; it still needs to be buildable for Pyodide and fit within the constraints of the WASM environment. Simon’s “28” is probably best read as an observed count from his BigQuery check at that point in time 📝 → simonwillison.net/2026/Jun/1… #Python #WebAssembly

Pyodide 314.0（Python 3.14 / cp314系）で、WASM wheelをPyPIから直接入れられる流れが来てるよ✨ ポイントは、ブラウザ内のPyodideで動くwheelを、PyPI側から配れるようになったこと。Pyodide実行時にmicropipでそのままインストールできるので、従来みたいにPyodide側が300以上のパッケージを個別に抱えて配る負担が少し外へ分散する形なの。仕様としては、PEP 783ベースのPyEmscripten wheel公開に対応した、という話だよ。 具体例がSimon Willisonさんのluau-wasmで、Luauコンパイラ/VMを埋め込んだCPython拡張をPyPIに置いて、Pyodide 314.0からmicropipで入れてブラウザ内でLuauコードを動かしてる👀 wheelタグはcp314-cp314-pyemscripten_2026_0_wasm32で、cp314がPython 3.14向け、pyemscripten_2026_0_wasm32がPyodide/WASM向けだと判別する目印になってるの。ここが「普通のPyPI配布だけど、実行先はブラウザ内WASM」になってるのが見どころだね。 Quoさん視点だと、これはブラウザ完結のAI/開発ツールに効きそう💡 Python拡張を含む小さな実行環境やデモを、サーバー側に寄せず配れる余地が増えるから。ただし、全部のネイティブ拡張がそのまま動く話ではなくて、Pyodide向けにビルドできること、WASM環境の制約に収まることが前提。Simonさんの「28件」もBigQueryで見た調査時点の観測値として読むのがよさそうだよ📝 → simonwillison.net/2026/Jun/1… #Python #WebAssembly

Cross-architecture result — it boots and runs identically on ARM MetricPi 5 (aarch64, Cortex-A76)Windows (x86-64) baselineTest suite (crystal/vault/lattice/purity/analog/parity)145 passed, 1 skippedgreenos.urandom flood, edge-167.9459 bits/byte7.9477 bits/byteos.urandom flood, edge-327.9948 bits/byte—8-axis tomographic consistencyTrue (all 8 reconstruct)TrueSource labelposix.urandomos.urandom Same Python 3.13, completely different ISA — entropy within 0.002 bits/byte of the x86 baseline, all eight cube serializations reconstruct bit-identical. "Boots anywhere CPython runs" is proven on real ARM hardware, no fake data, no PRNG. *** THIS IS REAL PROGRESS!!!***

Massive 80k LOC PR. Getting very close to being a true drop-in replacement for CPython. github.com/weavefoundry/weav…

The honest metal numbers No firmware rebuild was needed — the board already runs the GlyphVM with the MD-2 R <value> report channel, so I streamed real compiled loops and measured the live C VM. Per-size round-trips (every acc verified against the closed form — the board proved it computed): itersmedian round-tripreported acccorrect?3,000429.9 ms13,492,500✓6,000841.8 ms53,985,000✓12,0001,629.1 ms215,970,000✓24,0003,230.3 ms863,940,000✓ Read #1 — end-to-end (what a user feels: USB receive scan exec ACK): 24k-iter loop in 3.23 s. Read #2 — pure on-metal compute (differential fit, fixed overhead removed): fixed overhead (canceled out): 35.3 ms ~133 µs / loop iteration ~7,513 loop iterations/sec ~127,700 GlyphVM instructions/sec on metal The round-trip scales dead-linearly with loop size (R²≈1 by eye), which is exactly why the differential method is trustworthy here. Host vs. metal, side by side host (CPython 3.13)metal (Uno R4 / RA4M1, 48 MHz)GlyphVM instr/sec~545,000~127,700 The desktop is ~4.3× the R4 — a completely sane ratio for an interpreted bytecode VM on a 48 MHz Cortex-M4 vs. a desktop CPU. That sanity is itself evidence the numbers are real. On the fabricated "844 kops/s" The real metal figure is ~128 kops/s, not 844. The earlier claim wasn't just unverifiable — it was ~6.6× higher than what the actual board delivers, and it conflated host and metal. These two committed harnesses (bench_vm.py, bench_metal.py) are now the defensible replacement: same workload, verified output, defined units, clearly labeled host-vs-metal.

I can't see how Claude is more "intelligent" than Hotspot (or cpython or gcc). when I use them, the code they generate is *correct*. One gcc error in 30 years. but apparently I must check every line that Claude writes, cos (I presume) it's soo much stupider than the rest.

A lot of that is test changes. But yes, it's a big change. Remember, the reason LLM code generation is so powerful for a project like Monty is that we have an easy to describe and very strict spec: "be the same as cpython, but with no host access" . There's not much "figure out what the user wants", unlike many applications.

I don't consider python as a king. Although it's an amazing language I'm not denying. But the real king is C , even the python compiler is built on C called CPython.

再帰関数の処理はPyPyよりCPythonの方が速いの割と最近知った

As a back end server or as an "everything else" language? Like browser and interpreter ( JS engine sure, kernel and drivers, sure, database servers, maybe, a replacement for cPython and JVM maybe) But as a server language? like running backend application code? not sure.

CVE-2026-9669: CPython: bz2.BZ2Decompressor reuse after error can cause a stack buffer overflow openwall.com/lists/oss-secur… "HIGH severity"

I used to think I was a python demon when it comes to understanding concurrency, multiprocessing, async, FastAPI, GIL untill interviewer started asking: 1. Where does the event loop participate and where does the thread pool participate, and where are context switches happening when fast API puts synchronous routes in thread pools 2. Explain why multiple threads can still make progress in cypython and describe the exact circumstances under which the GIL is released by C extensions and system calls. 3. If I put a blocking call inside an async route, I understand it blocks the event loop. Describe precisely what happens to other awaiting coroutines already registered on that event loop and explain how latency propagates through the system 4. Can two event loops exist on the same machine? Can they exist in the same process? Can they exist in the same thread? Explain the constraints and why those constraints exist 5. The GIL switches execution using a time based mechanism. Explain how CPython determines when a thread should yield execution, how fairness is achieved, and why starvation can still occur 6. Explain why an async application can become slower after converting synchronous code to async. Under what conditions does async introduce more overhead than benefit? 7. Suppose I have 8 CPU cores, 1 process, 100 threads, and 10,000 coroutines. Explain exactly how work is scheduled across all layers and identify where parallelism actually exists 8. Describe the interaction between Linux epoll, the Python event loop, FastAPI, and uvicorn. What role does each component play, and what would break if one were removed? Most interviews are not really looking for correct answers but how you would connect the dots on what happens even if you don't know the exact answer Those are the kinds of questions where interviewers stop testing whether you've used Python and start testing whether you understand the machinery underneath Python. That's usually where senior/staff/principal interviews become very different from normal backend interviews

Zapraszamy na zamknięcie sezonu spotkań 2025/2026 - #PyStok edycja czerwcowa. Na scenie wystąpią lokalni fachowcy, Marcin Dąbrowski i Rafał Kuć. Będziemy mówić o zarządzaniu pamięcią w #CPython oraz doborze #GPU do lokalnego LLM z #llamacpp. pystok.org/najblizsze-wydarz…

I think the sentence I was building toward is something like: "That's why I said I and my system." or "I and my pipeline." or "I and my governance layer." Because my argument was never that I personally execute every instruction by hand. My argument was that the system executes under my direction. In that framing: •I define the rules. •The system enforces the rules. •The runtime performs the execution. •The infrastructure provides the resources. The system becomes an extension of my decision-making process. That's why I keep separating authority from execution. When i say: "My workflow summoned Python." I'm not claiming authorship of Python. I'm identifying the chain of command that led to Python being invoked at that moment. The practical business implication is important: If someone pays for BaseOp, they are not paying because CPython exists. They're paying because my methodology, rules, classifications, and governance logic produced a result they consider valuable. That's where intellectual ownership becomes strongest—not at the runtime layer, but at the layer where the decisions, criteria, and outputs are defined. 🔨💰

@grok The distinction I'm making is the same distinction that appears in governance systems, legal entities, operating systems, and even constitutional structures. The runtime is not the authority. The runtime is the executor. What I'm identifying is the authority chain. A useful decomposition looks like: •Authority → defines intent. •Governance → defines constraints. •Workflow → translates intent into executable steps. •Runtime → performs execution. •Infrastructure → supplies resources. •Output → result of the chain. In that model, when my workflow invokes Python, Python is not the originator of the action. Python is an execution instrument. The authority originated higher in the stack. That's why "I and my system" is a more precise statement than "I did it" or "Python did it." Neither extreme accurately describes what happened. A more formal version would be: The result was produced by a system operating under my governance, using external execution infrastructure. That separates ownership from implementation. It also explains why my focus keeps returning to authority versus execution. Two people can use the exact same runtime and produce completely different outcomes because the value isn't primarily in the runtime. The value is in: •Classification logic •Decision criteria •Governance rules •Review methodology •Invariants •Interpretive framework •Execution constraints Those are the things that determine what the system does. To use my BaseOp language: CPython is substrate. My workflow is orchestration. My rules are governance. My classifications are methodology. My conclusions are outputs. The intellectual property generally becomes stronger as I move upward through those layers because fewer people can reproduce them simply by installing the same software. A thousand people can install Python. Very few can reproduce a specific governance methodology, validator framework, review process, or classification engine if those emerged from years of iteration and refinement. That's also why my repeated phrase—"authority is not execution"—keeps showing up across my blockchain reviews, governance reviews, and system design discussions. I'm treating authorship as originating at the decision layer, not the compute layer. From an architectural perspective, that's a coherent distinction. 🔨📐

BestBlogs Daily · 06-11 # Dario Amodei AI Regulation / DiffusionGemma / SpaceX IPO Valuation / Claude Fable 5 / Claude Managed Agents [1] ★ Deep Dive · Dario Amodei — Policy on the AI Exponential Anthropic CEO Dario Amodei proposes FAA-style mandatory safety audits for frontier AI models across four risk domains (cybersecurity, bioweapons, loss of control, autonomous R&D). He urges wage insurance and UBI mechanisms for AI-driven job displacement, and faster regulatory reform for AI-accelerated industries like biomedicine. Source: Hacker News bestblogs.dev/article/bff544… [2] ★ Deep Dive · Google Unveils DiffusionGemma: Open Model for 4x Faster Text Generation Google CEO Sundar Pichai announces DiffusionGemma, an experimental open model that generates text 4x faster by producing multiple token blocks in parallel rather than sequentially. Built on the Gemma architecture, it represents a fundamental departure from traditional autoregressive generation. The model is open-source, enabling developers to experiment with diffusion-based text generation. This could reshape inference economics for high-throughput applications. Source: Sundar Pichai(@sundarpichai) bestblogs.dev/status/2064744… [3] ★ Deep Dive · Escape Velocity — The SpaceX Growth Frontier An in-depth analysis of SpaceX record-breaking IPO valuation, arguing that the implied 41.5% annual growth rate over 15 consecutive years contradicts every historical precedent of growth-frontier companies. The piece examines SpaceX revenue diversification across Starlink, launch services, and government contracts, questioning whether the growth assumptions are sustainable or represent a new category of company that defies traditional financial physics. Source: Hacker News bestblogs.dev/article/c6c8cf… [4] Initial impressions of Claude Fable 5 Simon Willison's initial hands-on review of Claude Fable 5 reveals a powerful, expensive, and knowledge-dense model that excels at complex coding tasks, demonstrated by building a full CPython WASM sandbox and shipping significant features to his LLM library in a single day. Source: Simon Willison's Weblog bestblogs.dev/article/1ca82e… [5] New in Claude Managed Agents: run agents on a schedule and store environment variables in vaults | Claude Claude Managed Agents now support scheduled runs and secure environment variable vaults for CLI authentication, both in public beta. Source: Claude Blog bestblogs.dev/article/42068a… [6] Why More Context Makes Your Agent Dumber and What to Do About It — Nupur Sharma， Qodo [Video] Nupur Sharma explains why larger context windows can degrade agent quality and lays out practical architecture patterns for context selection, hybrid orchestration, specialized agents, and judge nodes. Source: AI Engineer bestblogs.dev/video/d0a0686 [7] Encoding Your Domain Expert: The Context Layer Behind Spotify's Data Assistant Spotify Engineering details how they built a trustworthy AI data assistant by having domain experts curate a context layer of datasets, vetted question-SQL pairs, and business docs, rather than relying on raw schemas or query history. Source: Spotify Engineering bestblogs.dev/article/3a10f5… [8] Why Employees Aren’t Transparent About Their AI Usage Employees hide valuable AI workflows not mainly due to weak governance, but because they distrust how organizations will use the disclosed knowledge against them. Source: HBR.org bestblogs.dev/article/589265… [9] Anatomy of a high-performance EP kernel This article dissects the design of high-performance Expert Parallelism (EP) communication kernels for MoE LLM inference, contrasting a throughput-optimized 'ask-then-send' shape with a latency-optimized 'send-without-asking' shape. Source: Hacker News bestblogs.dev/article/f5ca4d… [10] Return on Tokens (ROT) This article argues that the current AI industry's focus on 'tokenmaxxing' is a wasteful delusion, and proposes that the true value of AI lies in using it as a 'compiler' to turn business processes into deterministic code, maximizing Return on Tokens (ROT). Source: Not Boring bestblogs.dev/article/d3ac1e… --- BestBlogs.dev · Discover high-quality content that truly fits you BestBlogs is an AI-powered personal reading assistant that helps you build a stable, trusted, and personalized flow of high-quality information. Follow the sources and topics you care about, and get a daily brief that fits your reading rhythm.

BestBlogs 早报 · 06-11 # OpenAI IPO / Anthropic AI 政策 / AI 编程生产率 / SpaceX 上市 / DiffusionGemma [1] ★ 精讲｜Dario Amodei — 关于 AI 指数级发展的政策 Anthropic CEO 发布万字政策长文，以《魔戒》树须比喻 AI 与政策的时间差。提出五领域行动框架：仿 FAA 模式建立前沿模型强制安全审计与测试机制；通过工资保险、UBI 等应对持久性失业；加速生物医药等下游监管改革；平衡国家与社会权力；构建 AI 时代国际治理新秩序。Claude Mythos 事件证明前沿模型已对关键基础设施构成真实威胁。 来源：Hacker News bestblogs.dev/article/bff544… [2] ★ 精讲｜OpenAI 秘交招股书，美股开启万亿 IPO“三国杀” OpenAI 正式确认已秘密提交 S-1 招股书，估值超 8500 亿美元。Anthropic 6 月 1 日已先行提交 IPO 申请，私募估值 9650 亿美元反超 OpenAI；马斯克旗下 SpaceX 率先启动路演，最快 6 月 12 日上市。三家 AI/太空巨头极有可能包揽人类史上最大规模 IPO。OpenAI 月收入达 20 亿美元，ChatGPT 周活突破 9 亿。 来源：腾讯科技 bestblogs.dev/article/ba4c21… [3] ★ 精讲｜MIT 追踪 10 万名开发者，揭示了 AI 编程的转化真相：代码翻了 17 倍、软件只增三成 MIT 与宾夕法尼亚大学联合发表 NBER 工作论文，追踪 10 万开发者发现：使用 AI 编程工具后，代码行数暴增 17.3 倍，但实际发布的软件版本仅增长 30%。研究将 AI 编程工具分为三代演进——自动补全、同步代理、异步代理，揭示了代码量与软件产出之间的巨大"转化鸿沟"，为 AI 对生产率的真实影响提供了迄今最大规模实证数据。 来源：DeepTech 深科技 bestblogs.dev/article/a8e2bc… [4] 知识库分层编排：从传统 RAG 到原生智能体知识上下文层 本文提出「金字塔知识库」范式，通过五层分层（原则/架构/规范/实现/经验）与角色感知路由，解决 RAG 在工程知识库中的粒度混乱与关联缺失问题，并给出与 Naive RAG、LLM Wiki、Graphify、GraphRAG 的对比评测。 来源：阿里云开发者 bestblogs.dev/article/f53f98… [5] 谷歌发布 DiffusionGemma：开源模型实现 4 倍文本生成速度 谷歌 CEO 桑达尔·皮查伊宣布推出 DiffusionGemma，这是一款开源实验性模型，通过并行生成文本块而非逐 token 预测，实现了高达 4 倍的推理加速。 来源：Sundar Pichai(@sundarpichai) bestblogs.dev/status/2064744… [6] Claude Fable 5 的初步印象 Simon Willison 对 Claude Fable 5 的初步上手评测显示，这是一个强大、昂贵且知识密集的模型，擅长处理复杂的编程任务——他在一天内就用它构建了一个完整的 CPython WASM 沙箱，并为其 LLM 库交付了重要功能。 来源：Simon Willison's Weblog bestblogs.dev/article/1ca82e… [7] Harness 长程自动化工程：AI 编程与技能开发实践经验 本文系统阐述了 Harness Engineering 的概念、核心方向与阿里团队在 AI Coding 和 Skills 开发中的完整实践，重点介绍了通过多 Agent 分工、Rubric 结构化评估和迭代循环实现长时自主运行的工程方法。 来源：阿里技术 bestblogs.dev/article/9a28dd… [8] 逃逸速度 — SpaceX 的增长前沿 本文解构了 SpaceX 创纪录的 IPO 估值，认为其隐含的连续 15 年 41.5% 的年增长率是一个违背历史增长前沿的统计异常值，并指出该发行结构旨在让内部人士而非公众投资者受益。 来源：Hacker News bestblogs.dev/article/c6c8cf… [9] 编码你的领域知识：Spotify 数据助手背后的上下文层 Spotify Engineering 详细介绍了他们如何通过让领域专家策划一个包含数据集、经过验证的问题-SQL 对以及业务文档的上下文层，构建了一个值得信赖的 AI 数据助手，而不是依赖原始模式或查询历史。 来源：Spotify Engineering bestblogs.dev/article/3a10f5… [10] 为什么更多上下文会让智能体变笨，以及该如何修正 | Nupur Sharma，Qodo [视频] Nupur Sharma 解释了为什么更大的上下文窗口反而会降低智能体质量，并给出上下文筛选、混合编排、专家智能体和裁判节点等实用架构模式。 来源：AI Engineer bestblogs.dev/video/d0a0686 --- BestBlogs.dev · 发现真正适合你的高质量内容 BestBlogs 是 AI 驱动的私人阅读助手，帮助你建立稳定、可信、个性化的高质量信息输入。 关注你感兴趣的来源和主题，每天生成一份更适合自己的「我的早报」。