{"version":"https://jsonfeed.org/version/1","title":"Readable: Hacker News","home_page_url":"https://readable.news","feed_url":"https://readable.news/api/feed?format=json","icon":"https://news.ycombinator.com/y18.svg","favicon":"https://news.ycombinator.com/favicon.ico","items":[{"title":"Kimi-K3 on HuggingFace","content_html":"
\"Kimi

\"Chat\" \"Homepage\"

\"Hugging \"Twitter \"Discord\" \"ModelScope\"

\"License\"

📰  Tech Blog |     📄  Full Report

1. Model Introduction

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.

Key Features

  • New Architecture: Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), and scales up MoE sparsity with a Stable LatentMoE framework that activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2.
  • Long-Horizon Coding: Operating with minimal human oversight, Kimi K3 sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools — from GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and even chip design.
  • Agentic Knowledge Work: Kimi K3 advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture.
  • Native Multimodality & Long Context: Kimi K3 understands text, images, and video within the same model, and supports a 1-million-token context window.
  • Open Frontier Weights: We release the full Kimi K3 model weights under the Kimi K3 License, making frontier intelligence openly available for research, deployment, and further innovation.

2. Model Summary

Architecture Mixture-of-Experts (MoE)
Total Parameters 2.8T
Activated Parameters 104B
Number of Layers 93
Number of Dense Layers 1
Attention-Layer Composition 69 KDA + 24 Gated MLA
Attention Hidden Dimension 7168
Number of Attention Heads 96
Latent MoE Dimension 3584
MoE Hidden Dimension (per Expert) 3072
Number of Experts 896
Selected Experts per Token 16
Number of Shared Experts 2
Vocabulary Size 160K
Context Length 1048576
Attention Mechanism KDA & Gated MLA
Activation Function SiTU-GLU
Vision Encoder MoonViT-V2
Parameters of Vision Encoder 401M
Quantization MXFP4 weights / MXFP8 activations
(quantization-aware training)
Modality Text, Image

3. Evaluation Results

Benchmark Kimi K3
(max)		 Claude Fable 5
(max, w/ fallback)		 GPT-5.6 Sol
(max)		 Claude Opus 4.8
(max)		 GPT-5.5
(xhigh)		 GLM-5.2
(max)
Reasoning & Knowledge
GPQA Diamond 93.5 92.6 94.1 91.0 93.5 91.2
CritPt 23.4 28.6 32.3 20.9 27.1 20.9
AA-LCR 74.7 70.0 73.7 67.7 74.3 71.3
HLE-Full 43.5 / 56.0 53.3 / 63.0 44.5 / 58.0 49.8 / 57.9 41.4 / 52.2 —
Coding
DeepSWE 67.5 70.0 73.0 59.0 67.0 46.2
ProgramBench 77.8 76.8 77.6 71.9 70.8 63.7
Terminal-Bench 2.1 88.3 88.0 88.8 84.6 83.4 82.7
FrontierSWE 81.2 86.6 71.3 66.7 64.9 67.3
SWE-Marathon 42.0 35.0 39.0 40.0 14.0 13.0
PostTrainBench 36.6 41.4 34.6 34.1 28.4 34.3
MLS-Bench-Lite 48.3 49.9 46.2 42.8 35.5 40.4
SciCode 58.7 60.2 56.1 53.5 56.1 50.5
Kimi Code Bench 2.0 72.9 76.9 64.8 71.7 69.0 64.2
Agentic
BrowseComp 91.2 88.0 90.4 84.3 84.4 —
DeepSearchQA (F1) 95.0 94.2 — 93.1 — —
ResearchRubrics 76.2 — 73.8 73.5 64.0 71.1
GDPval-AA v2 (Elo) 1686 1747 1736 1593 1491 1510
Toolathlon-Verified 76.5 77.9 74.9 76.2 73.5 59.9
MCPMark-Verified 94.5 87.4 92.9 76.4 92.9 —
MCP-Atlas 84.2 84.7 83.6 83.6 82.8 82.6
AutomationBench 30.8 29.1 29.7 27.2 22.7 12.9
JobBench 54.3 57.4 45.4 48.4 38.3 43.4
AA-Briefcase (Elo) 1548 1583 1495 1354 1158 1260
Agents' Last Exam 28.3 25.7† 29.6 27.0 26.6 20.4
APEX-Agents 41.0 43.3 39.9 39.4 38.5 35.6
OfficeQA Pro 63.3 69.9 63.2 63.9 60.9 41.4
SpreadsheetBench 2 34.8 34.7 32.4 31.6 29.1 28.1
OSWorld-Verified 84.8 85.0 83.0 83.4 79.0 —
OSWorld 2.0 58.3 66.1 62.6 55.7 49.5 —
SaaS-Bench 60.1 — 61.4 56.1 43.8 —
τ³-Banking 33.4 26.8 33.0 27.6 31.3 26.8
Harvey Lab-AA 94.6 93.6 87.2 91.1 86.3 91.0
CorpFin v2 71.6 71.8 64.4 66.7 68.4 66.1
Finance Agent v2 54.4 56.3 53.8 53.9 51.8 49.7
Legal Research Bench 44.2 49.5 48.1 43.8 40.4 31.3
Vision
WorldVQA ForceAnswer 51.0 56.7 41.8 39.1 38.5 —
OmniDocBench 91.1 89.8 85.8 87.9 89.4 —
PerceptionBench 58.5 57.2 59.7 47.2 55.8 —
Video-MME (w. sub) 90.0 — 89.5 86.0 89.3 —
MMVU 82.1 — 81.2 79.2 81.7 —
BabyVision w/ python 85.7 90.5 88.9 81.2 83.6 —
MMMU-Pro 81.6 / 83.4 81.2 / 86.5 83.0 / 84.6 78.9 / 82.7 81.2 / 83.2 —
CharXiv (RQ) 84.8 / 91.3 88.9 / 93.5 84.6 / 89.1 80.5 / 89.9 84.1 / 89.0 —
MathVision 94.3 / 97.8 94.8 / 98.6 95.8 / 97.8 86.7 / 97.1 92.2 / 96.8 —
ZeroBench (pass@5) 23.0 / 41.0 23.0 / 46.0 17.0 / 35.0 17.0 / 34.0 22.0 / 41.0 —
Footnotes

All Kimi K3 results are obtained with reasoning effort set to 'max' and temperature = 1.0. For single-step tasks, such as GPQA Diamond, HLE-Full, and vision benchmarks without tools, we set top-p = 0.95; for agentic tasks, we set top-p = 1.0. For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision, and ZeroBench, each cell reports the scores without and with tool augmentation (general tools for HLE-Full, Python for the vision benchmarks), in that order.

  1. Reasoning & knowledge benchmarks
  2. Coding benchmarks
    • DeepSWE. Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is taken from the GLM-5.2 release blog; all remaining scores are from the official DeepSWE leaderboard, under which Kimi K3 attains 67.3 with the mini-SWE-agent harness. We report the DeepSWE v1.1 tasks.
    • Terminal-Bench 2.1. Kimi K3 is evaluated with the Kimi Code harness. For all other models, we report the best score across harnesses: GLM-5.2 with Claude Code (GLM-5.2 release blog); Claude Opus 4.8 and Claude Fable 5 with Terminus 2 (Artificial Analysis); GPT-5.5 and GPT-5.6 Sol with Codex (OpenAI).
    • ProgramBench. Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is from the GLM-5.2 release blog; all other scores are from Vals AI.
    • SWE-Marathon. Kimi K3, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.6 Sol is evaluated with the Codex harness. The GLM-5.2 score is from the GLM-5.2 release blog. Our evaluation is based on an H20-calibrated branch of the official tasks as of July 9, 2026, prior to the final v1.1 release: the Docker images, performance gates, and reference oracles for the GPU tasks have been recalibrated for H20, while the correctness and anti-cheat validators remain unchanged. Additionally, Claude Fable 5 hit fallbacks on 35% of the tasks in our evaluation, which may have negatively impacted its measured performance.
    • FrontierSWE. Kimi K3 is evaluated with the Kimi Code harness and GPT-5.6 Sol with the Codex harness; all other results are from FrontierSWE. Dominance scores are recomputed from the raw scores using the official evaluation script and are current as of July 16, 2026.
    • PostTrainBench. Scores for GLM-5.2, GPT-5.5, and Claude Opus 4.8 are adopted from the official PostTrainBench results. Kimi K3, Claude Fable 5, and GPT-5.6 Sol are evaluated with the official Harbor implementation at maximum reasoning effort, averaged over three runs on H20 GPUs (instead of H100 in the official setting) — Kimi K3 and Claude Fable 5 with the Claude Code harness, and GPT-5.6 Sol with the Codex harness.
    • MLS-Bench-Lite. Kimi K3 is evaluated with the Kimi Code harness; GLM-5.2 and the Claude models with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness.
    • SciCode. Scores are cited from Artificial Analysis as of July 23, 2026.
    • Kimi Code Bench 2.0 (in-house). Kimi K3 is evaluated with the Kimi Code harness (it attains 73.7 with the Claude Code harness); GLM-5.2, Claude Opus 4.8, and Claude Fable 5 with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. All models are evaluated at maximum reasoning effort, except GPT-5.5, which uses the \"xhigh\" setting. As the benchmark includes cybersecurity and safety-related tasks, we also disclose the fraction of refused or fallback tasks: Claude Fable 5 hit 13 fallbacks and 1 refusal out of 80 tasks; 10 refusals out of 80 tasks entered GPT-5.6 Sol's cyber guard; GPT-5.5 had 3 refusals out of 80 tasks.
  3. Agentic benchmarks
    • OfficeQA Pro. Each test case provides the agent with the entire PDF corpus, with all PDFs rendered as images and no machine-readable text available.
    • OfficeQA Pro and SpreadsheetBench 2. Kimi K3, GLM-5.2, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol are evaluated with the Codex harness.
    • MCP-Atlas. All models are evaluated on the 500-task public subset with a 100-turn limit, using Gemini 3.1 Pro as the judge.
    • AutomationBench. All models are evaluated on the 600-task public subset, following the official GitHub setup in all other respects.
    • BrowseComp. We adopt a context-compaction strategy triggered at 300K tokens. When evaluated with the full 1M-token context window and no context management, Kimi K3 achieves a score of 90.4. The results of Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 are cited from Anthropic and OpenAI.
    • GDPval-AA v2, AA-Briefcase, τ³-Banking, Harvey Lab-AA, and APEX-Agents. Scores are cited from Artificial Analysis and the APEX-Agents leaderboard as of July 23, 2026. For Harvey Lab-AA, we report the criterion pass rate.
    • CorpFin v2, Finance Agent v2, and Legal Research Bench. Scores are cited from Vals AI.
    • Agents' Last Exam. Scores are cited from the official leaderboard as of July 23, 2026; we report the leaderboard's primary pass-rate metric. On the leaderboard, each model is paired with a specific harness: Kimi K3 with Kimi Code; GPT-5.6 Sol and GPT-5.5 with Codex; Claude Fable 5, Claude Opus 4.8, and GLM-5.2 with Claude Code. † The Claude Fable 5 entry runs at xhigh effort with 40% of tasks annotated as downgraded.
  4. Multimodal benchmarks
    • Except for ZeroBench, which follows the official setting and is run five times, all multimodal scores are averaged over three runs. MMMU-Pro is evaluated following the official protocol, preserving the original input order and prepending images to the text input.
    • PerceptionBench is an in-house benchmark that focuses on atomic visual perception capabilities.

4. Native MXFP4 Quantization

Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility.

5. Deployment

You can access Kimi K3's API on https://platform.kimi.ai by selecting kimi-k3, and we provide OpenAI/Anthropic-compatible API for you. Currently, Kimi K3 is recommended to run on the following inference engines:

6. Model Usage

Kimi K3 always has thinking enabled, and will return reasoning_content. Thinking effort is configured with the top-level reasoning_effort request field, which supports \"low\", \"high\", and \"max\" (default \"max\").

Kimi K3 was trained in the preserved thinking history mode. For multi-turn conversations and tool calls, Kimi K3 requires the complete assistant message returned by the API to be passed back to messages as-is — including reasoning_content and tool_calls, not just content:

import openai def chat_with_preserved_thinking(client: openai.OpenAI, model_name: str): messages = [ { \"role\": \"user\", \"content\": \"Tell me three random numbers.\" }, { \"role\": \"assistant\", \"reasoning_content\": \"I'll start by listing five numbers: 473, 921, 235, 215, 222, and I'll tell you the first three.\", \"content\": \"473, 921, 235\" }, { \"role\": \"user\", \"content\": \"What are the other two numbers you have in mind?\" } ] response = client.chat.completions.create( model=model_name, messages=messages, stream=False, max_tokens=4096, reasoning_effort=\"max\", ) # the assistant should mention 215 and 222 that appear in the prior reasoning content print(f\"response: {response.choices[0].message.reasoning}\") return response.choices[0].message.content

For full guides and examples (vision input, structured output, partial mode, tool choice, dynamic tool loading, context caching), see the Kimi K3 Quickstart and Thinking Effort.

Coding Agent Framework

Kimi K3 works best with Kimi Code CLI as its agent framework. We warmly invite you to give it a try — run Kimi Code in your terminal and select Kimi K3 using the /model command. We hope you enjoy building with Kimi K3, and we would love to hear your feedback!

7. License

Both the code repository and the model weights are released under the Kimi K3 License.

8. Contact Us

If you have any questions, please reach out at support@moonshot.ai.

Downloads last month
99,214

Model tree for moonshotai/Kimi-K3

Spaces using moonshotai/Kimi-K3 5

Collection including moonshotai/Kimi-K3

Evaluation results

\n\"\"","excerpt":"We’re on a journey to advance and democratize artificial intelligence through open source and open science.","image":"https://cdn-thumbnails.huggingface.co/social-thumbnails/models/moonshotai/Kimi-K3.png","authors":[{"name":null,"url":"https://huggingface.co/moonshotai/Kimi-K3"}],"id":"49065752","url":"https://huggingface.co/moonshotai/Kimi-K3","external_url":"https://news.ycombinator.com/item?id=49065752","date_published":"2026-07-27T06:18:10Z"},{"title":"US citizen charged after GrapheneOS phone wipes during airport search","content_html":"

Serving tech enthusiasts for over 25 years.
TechSpot means tech analysis and advice you can trust.

A hot potato: A federal case in Atlanta is raising questions about a privacy-focused mobile operating system, with prosecutors arguing that its features were used to erase evidence. The US Department of Justice is attempting to prosecute Atlanta resident Sam Tunick under a federal statute that makes it a crime to destroy property in an effort to prevent it from being seized.

The case centers on Tunick's use of GrapheneOS, an open-source operating system that works on Google Pixel phones and lets users enter a passcode to wipe a device clean.

Experts said the legal approach is unusual and may be the first time the law has been aimed at an operating system. \"It's concerning – and sends the message that [GrapheneOS] is criminal by default,\" said Christophe Boutry, a cybersecurity and surveillance expert. Boutry and Bill Buddington, senior staff technologist at the Electronic Frontier Foundation, both said they had not seen a similar case.

The incident began at Hartsfield-Jackson Atlanta International Airport on January 24 of last year. Tunick had just returned from a trip to the Dominican Republic when he was stopped for questioning. According to court testimony, federal agents had already circulated his name and photo internally, saying he was under investigation for \"suspected terrorism activities\" because of his alleged association with the movement against Cop City.

Tunick was taken to a secondary screening room, where multiple agents questioned him. A motion filed by his defense argues the interrogation focused on child sexual abuse material as a pretext for investigating his connections to the protest movement. The motion also states that Tunick asked four times to speak with a lawyer and was denied each time. According to the same filing, agents did not present a warrant or read him his rights.

Government attorneys and agents pushed back during Monday's hearing. They described the encounter as a routine airport inspection. Larry Findley, a Customs and Border Protection officer, said agents were \"looking for anything that's prohibited.\"

During the questioning, agents repeatedly asked Tunick to unlock his phone and warned they would seize it if he refused. When he finally provided a passcode, the phone appeared to restart. The defense motion states that \"the screen went blank, flashed several times, and the phone appeared to restart,\" resulting in the loss of data.

The wipe is now central to the case. Prosecutors are treating it as an intentional act to destroy evidence, while the defense argues that the search violated Tunick's constitutional rights and that the evidence should be suppressed.

The case raises questions about which constitutional rights apply at US borders, including international airports, where authorities have broader search powers. A judge is not expected to rule on the defense motion until at least late October.

GrapheneOS is designed to improve privacy and security on Pixel phones. Supporters say those tools are legitimate security protections, not evidence of criminal intent. Boutry pointed to France and Spain, where authorities have struggled to gain access to secured devices. He said authorities have treated the use of GrapheneOS itself as suspicious. In Catalonia, Spain, police have been profiling people carrying Pixel phones, assuming they have GrapheneOS installed and are drug dealers or gang members.

\"The main goal [of the operating system] is protection of privacy,\" Boutry said. \"They're our phones and the state can't tell us how to use them.\"

The case is tied to ongoing opposition to Cop City, a $109 million police training facility that opened last spring. The project has drawn opposition from activists concerned about police militarization and environmental impacts. Law enforcement officials have defended it as necessary for training and recruitment.

Previous attempts to prosecute protesters at the state level have foundered, while federal authorities have more recently stepped in, including a separate indictment announced last month.

\n\"\"","excerpt":"The case centers on Tunick's use of GrapheneOS, an open-source operating system that works on Google Pixel phones and lets users enter a passcode to wipe a...","image":"https://www.techspot.com/images2/news/bigimage/2026/07/2026-07-26-image-4.jpg","authors":[{"name":"TechSpot","url":"https://www.techspot.com/news/113236-us-prosecutors-charge-atlanta-man-after-grapheneos-phone.html","avatar":"https://www.techspot.com/images/favicon30.ico"}],"id":"49063022","url":"https://www.techspot.com/news/113236-us-prosecutors-charge-atlanta-man-after-grapheneos-phone.html","external_url":"https://news.ycombinator.com/item?id=49063022","date_published":"2026-07-26T22:21:41Z"},{"title":"Kill The Cookie Banner","content_html":"

Stop the tracking circus.

Tired of misleading cookie banners? The EU Commission has finally proposed a solution: set your privacy preferences in the browser once, and never see another banner. Unfortunately, the tracking industry is pushing back – and so far, they’ve been successful. We need YOUR help to #KillTheCookieBanner!

Cookie banners are made to trick you into waiving your rights

You may think that EU privacy law requires cookie banners. But the law is clear: online tracking is prohibited by default.

Therefore, the tracking industry needs you to waive your rights. That’s why they invented cookie banners, which often are deliberately misleading as well as annoying.

This results in up to 90% of people saying “YES” – even though only around 3% actually want to be tracked online. This system is broken by design.

The solution: automatically communicate your privacy preference

In Autumn 2025, as part of a bigger legal reform*, the EU Commission finally proposed a solution to the cookie banner problem: automated signals that would communicate your privacy preferences between your device and websites or apps. You could then choose whether you want to accept, refuse, or limit tracking.

This idea is neither new nor complex. Your browser already automatically signals other preferences to websites, for example your preferred language. In some US states, such signals for privacy preferences are already legally supported.

The tracking lobby fights to keep the cookie banner

This would be a simple solution. But the tracking industry seems afraid that if you can express your preferences efficiently, it could result in lower consent rates for tracking.

Following lobbying efforts spearheaded by Google and the tracking industry, several Member States are now blocking the EU Commission’s proposal to get rid of the cookie banner.

But not only that: industry groups are also lobbying the European Parliament to reject the proposal.

We need YOUR help!

That’s where you come in: the fight to kill the cookie banner is far from over. The Member States and the European Parliament have not yet decided on their position on the issue yet.

You can take action by contacting your representative in the European Parliament or in your Member State and express your frustration.

*This proposal for privacy signals is part of an EU law reform called the Digital Omnibus. Most other parts of this reform are problematic and would weaken people’s rights. We want to make clear that we do not support these other aspects of the proposed reform.

*This proposal for privacy signals is part of an EU law reform called the Digital Omnibus. Most other parts of this reform are problematic and would weaken people’s rights. We want to make clear that we do not support these other aspects of the proposed reform.

\n\"\"","excerpt":"Tired of misleading cookie banners? The EU Commission has finally proposed a solution: set your privacy preferences in the browser once, and never see another banner. Unfortunately, the tracking industry is pushing back – and so far, they’ve been successful. We need YOUR help to #KillTheCookieBanner!","image":"https://killthecookiebanner.eu/wp-content/uploads/2026/07/kill-the-cookie-banner-logo.png","authors":[{"name":null,"url":"https://killthecookiebanner.eu/","avatar":"https://killthecookiebanner.eu/wp-content/uploads/2026/07/cropped-Logo-32x32.png"}],"id":"49057175","url":"https://killthecookiebanner.eu/","external_url":"https://news.ycombinator.com/item?id=49057175","date_published":"2026-07-26T11:53:17Z"},{"title":"Our position on open-weights models","content_html":"

A post by Dario Amodei, Anthropic CEO

Over the last few days there has been a lot of discussion about open-weights models, especially those from China. Reports suggest that some US officials are considering banning the use of Chinese open-weights models by US companies. In response, many tech companies have signed a letter supporting open-weights models, and some people have even accused Anthropic of wanting to ban open-weights models as a means of protecting our business. Anyone who has read my past writing should know that I don’t regard such bans as a useful measure, but let me state it clearly so that there is no doubt: Anthropic has never advocated for a ban on open-weights models.

Open-weights models that don’t have dangerous capabilities are a public good: they don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.

Protectionist bans would not address my most serious national security concerns. Specifically, I am worried about two nightmare scenarios. I laid these out in my essay The Adolescence of Technology six months ago1, and have held these positions consistently for many years:

  1. My primary concern is the risk that authoritarian governments—not solely the Chinese Communist Party (CCP), although the CCP is clearly the most capable threat—build AI models that are more powerful than those built by the US, and use them to achieve permanent military superiority or perpetrate incredibly deep repression of their own people. This concern is widely shared within the US government: Vice President Vance warned in Paris last year that “authoritarian regimes have stolen and used AI to strengthen their military, intelligence, and surveillance capabilities,” and the Intelligence Community’s 2026 Annual Threat Assessment found that “other global powers’ robust progress in AI is challenging US economic competitiveness and national security advantages.” It is irrelevant whether these models are released with open weights, and certainly irrelevant whether they are used by US businesses. In fact, the most dangerous model may be one that is trained in secret and handed only to the People’s Liberation Army for use in drones and the Ministry of State Security for surveillance and repression.
  2. My secondary concern is the risk that powerful AI models may be misused to carry out cyberattacks or biological attacks, and may have serious alignment problems. Open-weights models—it does not matter whether they come from China or anywhere else—do potentially present a higher risk than closed models, because it is very difficult to apply guardrails to them or monitor their usage, and once weights are released they cannot be withdrawn2. But banning the use of these models by US businesses does nothing to address this risk, because bad actors are unlikely to be legitimate US businesses. It would protect US AI companies from competition, but that has never been my goal.

To address these concerns, I do support the following three measures, which I and Anthropic have consistently advocated for:

  • We should not sell powerful chips or chipmaking equipment to China, and we should crack down on the rampant smuggling3 and workarounds used to obtain access to such chips. China has limited domestic production capacity, and therefore, due to the scaling laws, cannot build more powerful models than the US without US chips. This is the most efficient and direct way to block threat #1, and by hampering the training of models that are out of reach of US law, it also indirectly helps with threat #2.
  • We should crack down on industrial-scale distillation operations. Distillation is a much more compute-efficient process than training models from scratch. It allows China to build much better models than its number of chips would ordinarily enable, and thus partially evade chip bans. Distillation does not allow the CCP to obtain equivalent or superior AI capabilities to the US, but it can bring the Chinese frontier to within a few months of the US frontier. It is true that many of the companies carrying out these operations release open-weights models—but the open weights are far less relevant than the fact that the operations are backed by an authoritarian state seeking to overtake the US at the frontier. We should have policy interventions to deter this behavior. A blanket ban on open-weights models is neither the correct remedy nor something we have called for4.
  • All sufficiently capable models, open and closed, should go through mandatory safety testing. The best way to address threat #2 is to just directly test models for cyber, biological, and alignment risks before release. I think this idea is actually close to a consensus: I have been heartened both that the Trump administration has moved in this direction in recent months, and by recent industry proposals that would apply such testing to the most capable models regardless of their country of origin or whether they are open or closed (while exempting less capable models, such as those from startups and academia, entirely). Whether open models do or don’t pose an increased risk, and whether that risk can be mitigated, is something that should emerge from testing, rather than be decided in advance—and there may be promising methods for improving the safety of open-weights models, including recent research from Anthropic on modular training strategies. Note that to be effective, testing would need to be global, which means even the CCP would need to be on board. I think this may actually be possible: as I wrote in The Adolescence of Technology, limited cooperation around preventing AI biological weapons may be possible because it is in China’s interest too.

This brings me to the open letter. I agree with much of it: open weights expand access to the AI economy, they strengthen competition at least for some use cases, and they give customers greater control. Concerns about distillation should be addressed through targeted legal and commercial frameworks—the same measure I described above. But I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers. It seems at least as likely to me that the opposite will be true. For example, I worry that biology will have a strong attacker-defender asymmetry, where sufficiently capable models may be able to quickly weaponize pandemic-level viruses with widely available materials, whereas defense against these agents is a multi-year operational task in the best case (as we saw with Operation Warp Speed)5. Questions like this should be empirically answered by rigorous pre-release testing, not assumed in advance.

To summarize my and Anthropic’s position, we have not and are not advocating for a ban on open-weights models as a category. We should instead focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models, open and closed.

Related content

Cognizant and Anthropic expand their partnership to bring Claude to enterprise clients

Read more

Introducing Claude Opus 5

Opus 5 is a step change improvement for the Opus tier powering long-running agents while delivering improvements in coding and professional work.

Read more

A research agenda for the Economic Futures Research Fund

We’re sharing the research agenda for the Anthropic Economic Futures Research Fund.

Read more

\n\"\"","excerpt":"Anthropic CEO Dario Amodei on open-weights models","image":"https://cdn.sanity.io/images/4zrzovbb/website/62993bb857986e2808c2a75024f384e7b85312b6-2400x1254.png","authors":[{"name":null,"url":"https://www.anthropic.com/news/position-open-weights-models","avatar":"https://www.anthropic.com/favicon.ico"}],"id":"49076057","url":"https://www.anthropic.com/news/position-open-weights-models","external_url":"https://news.ycombinator.com/item?id=49076057","date_published":"2026-07-27T22:03:49Z"},{"title":"PGSimCity - How PostgreSQL Works","content_html":"

PGSimCity is an independent, non-commercial educational visualization of PostgreSQL internals. It is not affiliated with, sponsored, endorsed, or approved by Electronic Arts Inc. SimCity is a trademark of Electronic Arts Inc.

A working model of the PostgreSQL engine

loading the city code…

Early, unreviewed prototype. It almost certainly contains inaccuracies in both the model and explanations. Found one? Open an issue or send a pull request.

\n\"\"","excerpt":"An independent, non-commercial educational visualization of PostgreSQL internals. Not affiliated with Electronic Arts.","image":"https://nikolays.github.io/PGSimCity/og.png","authors":[{"name":"PGSimCity","url":"https://nikolays.github.io/PGSimCity/","avatar":"https://nikolays.github.io/PGSimCity/favicon.svg"}],"id":"49063754","url":"https://nikolays.github.io/PGSimCity/","external_url":"https://news.ycombinator.com/item?id=49063754","date_published":"2026-07-27T00:19:04Z"},{"title":"AI companies are shredding rare books","content_html":"

🦔AI companies are bulk-buying rare books, scanning them through high-speed machines that cut the spines off, and shredding the originals. A service called ISBNdb facilitates orders of up to a million books and keeps buyers anonymous. Pre-2022 books are premium because they're free of AI-generated text. A federal judge ruled the practice is fair use because eliminating the original means only one copy exists at a time. Anthropic hired the former head of Google Books partnerships to obtain \"all the books in the world.\" My Take This got to me. A bookseller told 404 Media that rare books with almost no surviving copies are being fed into this pipeline. Books that survived wars, fires, and centuries of handling are being shredded so an AI can learn to write a better marketing email. ISBNdb's website literally says \"'AI company destroys two million books' is not a headline that generates sympathy,\" and they still built an entire business around making it happen quietly. They offer NDAs as a feature. They coach clients to call it \"digital preservation.\" I've covered AI companies scraping the internet, torrenting libraries, and stealing music. This is worse because it's irreversible. You can re-upload a website. You can reprint a bestseller. You can't replace the last three copies of an 18th-century botanical text once someone shreds them for training data. And the judge said it's legal. So it's going to accelerate. \"We shred rare books and offer NDAs so nobody finds out\" is a legitimate business model in 2026. What a timeline. Hedgie🤗

\n\"\"","excerpt":"🦔AI companies are bulk-buying rare books, scanning them through high-speed machines that cut the spines off, and shredding the originals. A service called ISBNdb facilitates orders of up to a million books and keeps buyers anonymous. Pre-2022 books are premium because they're","image":"https://pbs.twimg.com/media/HOMX3t1XYAA3Pr7.jpg:large","authors":[{"name":"X (formerly Twitter)","url":"https://twitter.com/HedgieMarkets/status/2081534588485296565","avatar":"https://twitter.com/favicon.ico"}],"id":"49068738","url":"https://twitter.com/HedgieMarkets/status/2081534588485296565","external_url":"https://news.ycombinator.com/item?id=49068738","date_published":"2026-07-27T12:32:12Z"},{"id":"49057241","url":"https://swag.htmx.org/en-cad/products/htmx-4-the-game","external_url":"https://news.ycombinator.com/item?id=49057241","title":"Htmx 4.0, the first JavaScript library to release exclusively on the Game Boy","date_published":"2026-07-26T12:00:11Z"},{"id":"49051120","url":"https://didtheyghostyou.com/","external_url":"https://news.ycombinator.com/item?id=49051120","title":"Did they ghost you?","date_published":"2026-07-25T20:18:03Z"},{"title":"How is the Bun rewrite in Rust going?","content_html":"

I think it’s important to be very Canny when someone makes a claim that supports a company’s large valuation.

The Bun rewrite seems well positioned as proof-positive that AI and specifically Anthropic’s AI can do the work of open-source maintainers, for some money, but faster.

\"image\"

On the 8th of July 2026 Jarred Summner of Bun fame posted about “Rewriting Bun in Rust”. At the time, I felt pretty Canny, having already read about Anthropic’s C compiler and Cursor’s FastRender web browser. It seems like there is a lot of valuation money riding on the coding capabilities of AI, and a lot of marketing about those capabilities.

It’s not a huge stretch to link the acquisition of Bun by Anthropic with the choice to rewrite Bun with an Anthropic tool. In his article Jarred claims that over 11 days (between 3rd and 14th May 2026) and at a cost of $165,000 for Anthropic API calls the rewrite was done and then merged to main. This represents a cost of $15,000 dollars a day, a cost well outside the means of many open source maintainers. This number doesn’t seem to include the leviathan whirring of the CI/CD in the org’s Buildkite cluster, which appears to have been constantly whirring ever since the “rewrite” was “completed”.

Why do I put those two words in scare quotes?

Well, on the 9th of July a peer of mine in a group chat said something unsurprising given our current environment - breathlessly proclaiming that the era of AI had emerged, and its herald was the “Rewriting Bun in Rust” blog post. So, I decided to look into this. Since the 9th I’ve been having a closer look at the claims and the code, and today I cloned the Bun repo:

Receiving objects: 100% (1200304/1200304), 1.23 GiB | 11.97 MiB/s, done.

As of today, the 27th of July 2026, a six week period after the rewrite was merged to main, there’s still no release tag. It’s now been 11 weeks since the last Bun release tag:

2026-05-12 15:12:49 -0700 (tag: bun-v1.3.14)

The last time there wasn’t a Bun release in a month was between the 26th of October 2022 and the 7th of December 2022 when there was a gap of six weeks and November was skipped, between v0.2.2 and v0.3.0:

2022-12-07 00:37:40 -0800 (tag: bun-v0.3.0) 2022-10-26 21:06:02 -0700 (tag: bun-v0.2.2)

On the 9th of July, the number of open pull requests from robobun (a proxy for PRs made by claude code) was 1277. As of this moment, on the 27th of July that number is 2475 open PRs. I don’t claim to know for sure but as far as I can see the process of merging to main with Buildkite checks completed seems to take 40-ish minutes (sometimes it looks like it takes up to an hour and a half). At this rate if we want to merge all the open Claude PRs it’ll only take running the pipeline for 86 days, continuously.

Generously, some of these PRs don’t deal with Rust code. Further investigation was warranted. I initially clicked through some of the PRs and found some extremely reviewed ones.

At this point I began to suspect that much of the cost of the rewrite was off the books. While $15k a day in tokens widened my eyes a little (and maybe I’m naive, maybe that’s low), I hadn’t seen the numbers for the CI/CD costs of Buildkite, and as you might have noticed from above, some PRs were written by Anthropic employees. When I did analysis of the data it became clear to me that the project was ticking along with more Claude credits and more Anthropic employee involvement:

\"image\"

One of the assumptions I’ve made here is that Jarred Summner’s commits during the rewrite were using Claude.

It’s pretty clear that there was a big spike in Claude usage when the rewrite began. It seems like, now, Anthropic employee and Robobun involvement in Rust is ramping up:

\"image\"

It’s nice to see that the Anthropic employees involved seemed to have a good work-life balance, taking breaks on the weekend. But, it looks like that period is over. It seems like we’re ramping up to something, and there’s still no release tag.

What’s pretty obvious to me is that we can’t take it at face value that the rewrite is “done” or that it was “done” for $165k USD. The Bun team never made that claim, but I’ve seen and heard those breathless claims that the rewrite is proof of something, and I do still think we need to be Canny about claims that support a large valuation. Anthropic is dogfooding this, the machine is still ticking along, and Anthropic employees are directly involved. If we imagine that the rewrite is still costing $10k a day, we’re approaching $800k in money spent on this rewrite.

—

I’m not a primitivist when it comes to AI. Work I’ve done on ML and AI has been some of the work I’ve been most proud of. This current hype cycle is, however, the worst I’ve ever seen the hype get. Back in 2015 I had to plead with management, trying to get them to understand that a black-box bunch of linear algebra that was better than humans at a quantitative task was, just, better. Now it seems like the world has completely reversed, and management is so forthright in their belief that AI is good, that they want it rubbed on everything. And those valuations are sometimes contingent on AI eating everything:

\"image\"

In my tiniest mouse voice ever I want to ask probably the most relevant question one can ask about a Business Thing: Was it worth the money and are the companies worth the valuation?

Are we done yet?

P.S. Anthropic’s C compiler and Cursor’s FastRender web browser haven’t had any commits for months.

P.P.S. I’m looking for a job.

\n\"\"","excerpt":"I think it’s important to be very Canny when someone makes a claim that supports a company’s large valuation.","image":"https://lockwood.dev/assets/img/rust_rewrite/slop.png","authors":[{"name":"Tom Lockwood - Developer","url":"https://lockwood.dev/ai/2026/07/27/how-is-the-bun-rewrite-in-rust-going.html","avatar":"https://lockwood.dev/favicon.png"}],"id":"49067854","url":"https://lockwood.dev/ai/2026/07/27/how-is-the-bun-rewrite-in-rust-going.html","external_url":"https://news.ycombinator.com/item?id=49067854","date_published":"2026-07-27T11:12:26Z"},{"title":"The new rules of context engineering for Claude 5 generation models","content_html":"

Then and now

There were a number of previous context engineering best practices that had become myths. Including:.

Then: Give Claude rules

Now: Let Claude use judgement

When we first rolled out Claude Code, we needed to be sure that Claude avoided worst case scenarios, such as deleting files. This meant we would give particularly strong guidance that might not always be true, For example, in the system prompt we used to say: 

In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max. Don't create planning, decision, or analysis documents unless the user asks for them — work from conversation context, not intermediate files.

But for a certain subset of prompts, this guidance would be wrong. In the case of documentation, the user may have their own preferences, or specific parts of very complex code might need multi-line comment blocks.

Still, without these guardrails for older models, the comments Claude wrote would be incorrect in many cases and we had to accept this tradeoff. But newer models have better judgement and can handle these decisions well without explicit rules. 

In the new system prompt we say: Write code that reads like the surrounding code: match its comment density, naming, and idiom.

Then: Give Claude examples

Now: Design interfaces

The number one rule for tool usage was to give Claude examples on how to use them. With our newest models, we’ve found that giving examples actually constrains them to a certain exploration space. 

Instead of using examples, think more about the design of your tools, scripts and files- what parameters does Claude have and how can they be more expressive? 

For example, in the Todo tool example, just listing status as an enumeration between pending, in_progress, and completed, hints to Claude about how to use it. The instruction on keeping one item in_progress helps define our requested behavior.

Then: Put it all upfront

Now: Use progressive disclosure

Because Claude Code was focused on coding, our system prompt included detailed information on how to do code review and verification. These were not always needed, but when they were, it was crucial information.

Since then, Claude Code has gotten very competent at using progressive disclosure- loading the right context at the right times. For example, we moved verification and code review into their own skills that Claude Code could selectively call.

But progressive disclosure is not just for skills, we also use it for tools. Some of our tools are ‘deferred loading,’ which means the agent must search for their full definitions using ToolSearch before using them. This allows us to have more tools (such as our Task tools) that don’t take up context until they’re needed.

The same can be applied to your own CLAUDE.md and Skill.md files. A common myth is that you want to make these a central repository for every known practice that you might run into, because Claude would not find it otherwise. Instead, consider having a tree of files that can be loaded at the right time.

Then: Repeat yourself

Now: Simple tool descriptions

Earlier Claude models could sometimes need repeated instructions or be more likely to listen to instructions at the end of their context window than at the start. This meant our system prompt would sometimes have references to tools in the main system prompt as well as instructions in the tool description. 

We found we could delete these repeat examples and put instructions on how to use tools in the tool descriptions rather than the system prompt.

Then: Memory in CLAUDE.md files

Now: Auto-memory

We used to encourage users to save things to Claude’s memory, by using the # hotkey to write to their CLAUDE.md automatically. Instead, Claude now automatically saves memories that are relevant to the work and to you. 

Then: Simple specs

Now: Rich references

In plan mode, Claude Code has heavily relied on markdown files with plans. Storing these files as plans helped Claude refer to them when needed. Another similar best practice was to store specs in the codebase for Claude to refer to while working across longer projects.

But we’ve found that Claude can handle increasingly more complicated references. Instead of simple markdown files, Claude can reference HTML artifacts created by our new artifacts feature. 

You may also give Claude references in the form of code. A spec may also be a detailed test suite, or a function in a different codebase that Claude might port. 

Rubrics are another form of references. Rubrics allow Claude to try and verify your taste in a particular field (e.g. what does a good API design look like) by using dynamic workflows and spinning up verifier agents with those rubrics.

Applying this to your context

Pulling this all together, what does this look like when you assemble your context?

System Prompt

A system prompt is heavily tied to the product context. It tells Claude what product it’s operating in and what it’s doing. For Claude Code, you will likely never modify this, but if you are building your own agent harness, this is where you should spend a lot of time.

CLAUDE.md

Keep your CLAUDE.md lightweight and briefly describe what your repo is for, but spend most of the tokens on gotchas inside of the codebase. For example, you may organize your code to keep types in one monolithic file and nowhere else. Avoid stating ‘the obvious’ things Claude should know by looking at your file system or your repo.

Use progressive disclosure heavily, for example if you have several unique instructions on how to verify your work, create a verification skill and reference it from your CLAUDE.md.

Skills

Think of skills as lightweight guides to let Claude find information when needed. Avoid making them overconstrained, except in highly important areas. 

For long skills, try and use progressive disclosure as much as possible- divide it into many files and split them out.

It’s best when skills encode particular opinions, knowledge, or best practices that are particular to you, your team, or product. 

References 

You can @ mention files to include them as references. References allow Claude to refer to in-depth information about the current plan. 

This might be in specs files, mockups, or even entire codebases. Generally you should prefer files that are in code as it provides clear, high-fidelity instructions to Claude in a language it knows very well. For example, a HTML mockup of a design will generally produce better results than a description of the design or a screenshot.

Try simplifying

Across your system prompt, skills, and CLAUDE.md files, you may need to simplify just like we did. We rolled out a new command called `claude doctor,` which will help you do this automatically as well. For more details on prompting more advanced models specifically, check out our Fable field guide.

This article was written by Thariq Shihipar, member of technical staff, Anthropic.

\n\"\"","excerpt":"We removed over 80% of Claude Code's system prompt for more advanced models. How to apply the lessons we learned to your own context engineering in Claude Code and with your own agents.","image":"https://cdn.prod.website-files.com/68a44d4040f98a4adf2207b6/6a639c6b2e881afef80764c6_og_the-new-rules-of-context-engineering-for-claude-5-generation-models.jpg","authors":[{"name":"Claude","url":"https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models","avatar":"https://cdn.prod.website-files.com/6889473510b50328dbb70ae6/689f4a9aff1f63fde75cf733_favicon.png"}],"id":"49051361","url":"https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models","external_url":"https://news.ycombinator.com/item?id=49051361","date_published":"2026-07-25T20:42:35Z"},{"title":"French firefighters face 'pyrocumulonimbus' for first time","content_html":"

Such a formation, also called a cumulonimbus flammagenitus cloud, can be seen during volcanic eruptions -- or, as in this case, when a wildfire roasts the air near the ground, causing it to rise quickly.

The rising air cools and becomes a cloud, or fire cloud: a turbulent mass that can generate its own winds that worsen fires. If it becomes very large it can even generate lightning.

Hitherto, they have only really been seen in wildfire-prone Australia and North America. NASA calls them the \"fire-breathing dragon of clouds\".

What causes them?

FNSPF spokesman Lieutenant Colonel Eric Brocardi:

\"The rise in ground temperature at the heart of the blaze is so extreme that the intense heat surges upward in a column, meeting cooler air masses at the top. This contact creates a pyrocumulonimbus -- or fire cloud -- though unfortunately one devoid of moisture. This smoke cloud generates its own weather system, causing anything in its path to spontaneously combust.

\"Lightning forms within the cloud, striking the embers on the ground and further fuelling the fire ... It produces a low, rumbling sound.

\"While this type of fire occurs regularly in Canada and Australia, it is unprecedented in France.\"

Can it be brought under control?

\"We are dealing with a so-called 'convective' fire that creates its own winds -- winds that constantly shift direction, unlike a standard fire that spreads in a conical pattern. The result is that it spreads in every direction and develops multiple fire fronts, making the situation completely unpredictable.

\"We don't know how the blaze will spread; the fire front is constantly shifting. We cannot fight it directly. It's a David-versus-Goliath scenario: the idea is that, at some point, we'll find a weak spot and strike there.

\"Salvation will come either from the sky -- in the form of rain, though it would need to pour heavily for three days -- or by finding a way to steer the fire toward a place where it will die out on its own. The sea, for example.\"

Do the firefighters have any hope?

\"We have reached a point of 'operational impossibility' that requires us to accept the need for strategic retreat -- to understand and acknowledge a natural force beyond our control.

\"It calls for ... knowing what can be done and, above all, what cannot. The situation on the ground is defensive rather than offensive; it is a war-like scenario where we are under constant bombardment and must get everyone to safety.

\"We realise we are facing a situation where we are powerless. Yet, that powerlessness leaves no room for despair. It is certain that all the efforts being made will yield results. It is out of the question for the firefighters to feel that their work on the ground is in vain.

\"Every single action -- defending critical sites, rescuing animals -- breathes new life into their efforts, driven by their fierce determination. There is a support force ready to relieve them.

\"However, not all firefighters from across France can simply converge here, because the risk exists everywhere. We have to hold the line.\"

(Interviewed by AFP journalist Juliette Collen)

© 2026 AFP

\n\"\"","excerpt":"The gigantic wildfire ravaging southwest France has created an extremely dangerous \"pyrocumulonimbus\" -- a phenomenon never before seen in the country -- the spokesman for the country's National Firefighters Federation of France (FNSPF) told AFP.","image":"https://s.france24.com/media/display/28a20512-8901-11f1-af78-005056bfb2b6/w:1280/p:16x9/dafe7dea5fec7f3a48c4717459bcd02a6f29ce60.jpg","authors":[{"name":"FRANCE 24","url":"https://www.france24.com/en/live-news/20260726-french-firefighters-face-pyrocumulonimbus-for-first-time","avatar":"https://www.france24.com/favicon.ico"}],"id":"49060495","url":"https://www.france24.com/en/live-news/20260726-french-firefighters-face-pyrocumulonimbus-for-first-time","external_url":"https://news.ycombinator.com/item?id=49060495","date_published":"2026-07-26T17:49:44Z"},{"title":"GrapheneOS protections against data extraction from locked devices","content_html":"

GrapheneOS Discussion Forum

\n\"\"","excerpt":"GrapheneOS discussion forum","image":"https://discuss.grapheneos.org/opengraph.png","authors":[{"name":"GrapheneOS Discussion Forum","url":"https://discuss.grapheneos.org/d/40700-grapheneos-protections-against-data-extraction-from-locked-devices","avatar":"https://discuss.grapheneos.org/favicon.ico"}],"id":"49055169","url":"https://discuss.grapheneos.org/d/40700-grapheneos-protections-against-data-extraction-from-locked-devices","external_url":"https://news.ycombinator.com/item?id=49055169","date_published":"2026-07-26T05:57:22Z"},{"id":"49048034","url":"https://tobi.knaup.me/2026-07-25-open-weight-ai-is-having-its-kubernetes-moment/","external_url":"https://news.ycombinator.com/item?id=49048034","title":"Open-weight AI is having its Kubernetes moment","date_published":"2026-07-25T14:49:49Z"},{"title":"A shell colon does nothing. Use it anyway","content_html":"

I've written more shell scripts than I can count, but I still stumble upon tricks that honestly blow my mind far too often than I care to admit. Latest thing that blew my head clean off? The shell colon.

Published
Modified
Author
Filip Roséen
Tags

In a land far-far away..

... there was once a far too cold cup of coffee next to a freshly brewed far-too-hot one. Four different terminals where three could have been closed an hour ago, and a shell script which I really (really) did not want to write.

Who would have thought a single colon would be the one to save the day night?

Note: Want your mind blown straight away? See more colons in the limelight.

Checking for required arguments

This is a familiar dance, it's pretty much muscle memory by this point. You have a script, it takes a few arguments, and some of them are mandatory; alright, an if-statement like so many times before:

if [ -z \"$1\" ]; then echo \"missing argument, aborting.\" 1>&2 exit 1 fi echo \"Hello $1!\"

Though.. what if I told you the above four lines could be replaced by just... one?

: \"${1:?missing argument, aborting.}\" echo \"Hello $1!\" 
$ bash example.sh example.sh: line 1: 1: missing argument, aborting.  $ bash example.sh refp Hello refp!

And look what happens if we refer to a variable with a proper name — it's the same behavior as previously but easier to spot; the diagnostic includes the name of our variable!

: \"${GREET_NAME:?missing argument, aborting.}\" echo \"Hello $GREET_NAME!\" 
$ bash greet.sh greet.sh: line 1: GREET_NAME: missing argument, aborting.

Parameter expansion and the story of :?

There are two things going on in the previous snippet, and you are correct in identifying that one part is using parameter expansion:

That.. other colon

So that's one colon, but what about that other one, the one who sits alone at the beginning of the line?

More colons in the limelight

Perhaps we have already established that there is more to : than meets the eye, but to prove the real magic of the null-command — here are a few usages that blew my mind.

: \"${DATA_DIR:=/var/data}\" # set defaults, : swallows the result : \"${RETRIES:=3}\" # instead of running it as a command 
: > error.log # truncate error.log : > error.log > access.log # truncate both error.log and access.log 
( : < dataset.json ) && echo YES # is dataset.json readable? ( : >> result.json ) && echo YES # is result.json writable? 
trap : INT # trap requires a command sleep 60 # sleep is interruptible 
set -u # error on unset variables : \"$DEPLOY_ENV\" \"$HOST\" # check DEPLOY_ENV and HOST 
if some-command; then : # command required else echo \"command failed\" fi

Con-colon-sion

So, if you are like me and prefer less typing (gotta go fast) — the null-command and parameter expansion are a pair worth studying before your coffee goes cold.

And also.. isn't this — magic?

set : : : : : : : : : : : : : : : : : : : : : while : colons are more than \"${1:?magic}\"; do echo \"$*\" && shift done

Note: The above example is safe to run locally, try it!

Frequently Asked Questions

After reading a few comments online, it seems I skipped over some things worth explaining. I will keep this section updated as questions come up.

\n\"\"","excerpt":"I've written more shell scripts than I can count, but I still stumble upon tricks that honestly blow my mind far too often than I care to admit. Latest thing that blew my head clean off? The shell colon.","image":"https://refp.se/logo.png","authors":[{"name":null,"url":"https://refp.se/articles/your-shell-and-the-magic-colon","avatar":"https://refp.se/favicon.ico?v=1"}],"id":"49047453","url":"https://refp.se/articles/your-shell-and-the-magic-colon","external_url":"https://news.ycombinator.com/item?id=49047453","date_published":"2026-07-25T13:33:49Z"},{"title":"Kimi-K3 Technical Report [pdf]","content_html":"

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

Sign in

%2F%2Fblob%2Fshow&source=header-repo&source_repo=MoonshotAI%2FKimi-K3\" data-hydro-click=\"{"event_type":"authentication.click","payload":{"location_in_page":"site header menu","repository_id":null,"auth_type":"SIGN_UP","originating_url":"https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf","user_id":null}}\" data-hydro-click-hmac=\"8276f579d07b993e1aaed60c2cbb5e161acbaa64ea25edaba72004c266ab7571\" data-analytics-event=\"{"category":"Sign up","action":"click to sign up for account","label":"ref_page:///blob/show;ref_cta:Sign up;ref_loc:header logged out"}\"> Sign up

Appearance settings
\n\"\"","excerpt":"Open Frontier Intelligence. Contribute to MoonshotAI/Kimi-K3 development by creating an account on GitHub.","image":"https://opengraph.githubassets.com/7ba6413686b919b09f1b8e8c0c0f7636cd0f003430d3a8302d43829a3a21523e/MoonshotAI/Kimi-K3","authors":[{"name":"GitHub","url":"https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf","avatar":"https://github.githubassets.com/favicons/favicon.svg"}],"id":"49070985","url":"https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf","external_url":"https://news.ycombinator.com/item?id=49070985","date_published":"2026-07-27T15:23:45Z"},{"id":"49060856","url":"https://beyondloom.com/decker/","external_url":"https://news.ycombinator.com/item?id=49060856","title":"Decker, a platform that builds on the legacy of Hypercard and classic macOS","date_published":"2026-07-26T18:23:06Z"},{"title":"Ruff v0.16.0 – Significant new updates – 413 default rules up from 59","content_html":"

Ruff v0.16.0 is available now! Install it from PyPI, or with your package manager of choice:

uv tool install ruff@latest

As a reminder: Ruff is an extremely fast Python linter and formatter, written in Rust. Ruff can be used to replace Black, Flake8 (plus dozens of plugins), isort, pydocstyle, pyupgrade, and more, all while executing tens or hundreds of times faster than any individual tool.

Migrating to v0.16 #

Ruff v0.16 has a small number of breaking changes, allowing most users to update without significant changes to code or configuration. The main exception is described below.

Better default rule set #

Ruff now enables 413 rules by default, up from 59 in previous versions.

Since Ruff's default rule set was last modified in v0.1.0, the number of rules in Ruff has grown from 708 to 968. Many of these rules catch severe issues, including syntax errors and immediate runtime errors but were not previously enabled by default. With the new rule set, Ruff will bring these issues and many others to your attention without any Ruff configuration. Even if you're already using select or extend-select, we hope that this will draw your attention to helpful rules that you previously hadn't discovered.

The full listing of enabled rules is too long to include here, but you can find it on our new Default Rules page in the documentation. A few of the highlights include rules from the popular flake8-bugbear (B) and pyupgrade (UP) linters, as well as rules from our own RUF category.

Note that this is primarily an expansion, but 18 of the more opinionated pycodestyle (E) and pyflakes (F) rules have been removed from the default set: E401, E402, E701, E702, E703, E711, E712, E713, E714, E721, E731, E741, E742, E743, F403, F405, F406, and F722.

If you want to revert to the old default set, you can easily select the old rules with this configuration:

[lint] select = [\"E4\", \"E7\", \"E9\", \"F\"]

We view this work as closely tied to our longstanding goal of rule recategorization, so look forward to upcoming developments in this area.

New features in v0.16 #

Ruff v0.16 also includes several newly stabilized features that are highlighted below.

Markdown code block formatting #

Ruff can now format Python code blocks embedded in Markdown files.

In these files, Ruff v0.16 will format fenced code blocks with a python, py, python3, py3, pyi, or pycon info string. pyi blocks are formatted like stub files, pycon blocks as REPL sessions, and the others use normal Python file formatting. For example:

# README  Here's an example:  ```py import ruff  ruff_binary = (  ruff.find_ruff_bin() ) ```

will be reformatted when running ruff format:

\"Ruff

This can also be used to format Quarto notebooks because Ruff still recognizes the language when surrounded by curly braces (e.g. ```{python}). Note that you may need to configure your extension mapping if your Quarto files use the .qmd extension.

If you need to suppress formatting, you have several options. If you want the suppression comments to appear in the code block, you can use normal fmt: off and fmt: on comments within the code block itself, or you can use similar HTML comments to disable formatting for a whole region of the document:

# README  <!-- fmt: off -->  ```py x = \"this will be suppressed\" ```  <!-- fmt: on -->

To suppress Markdown formatting entirely, you can use the normal extend-exclude setting to exclude all Markdown files with a glob like *.md.

See the full documentation for more details.

ruff: ignore comments offer new suppression features #

Ruff now has its own suppression comment format that can be used on its own line.

In v0.15, the Ruff linter gained a range suppression mechanism through paired ruff: disable and ruff: enable comments, much like the fmt: off and fmt: on pair described above:

# ruff: disable[N803] def foo(  legacyArg1,  legacyArg2,  legacyArg3,  legacyArg4, ): ... # ruff: enable[N803]

Ruff v0.16 builds on this to add two additional ruff suppression comments. ruff: ignore can be used to suppress a diagnostic on the same line, like noqa, or on the following logical line:

import math # ruff: ignore[F401]  # ruff: ignore[N803] def foo(  legacyArg1,  legacyArg2,  legacyArg3,  legacyArg4, ): ...

In this case, the logical line spans the whole function header (from def to the colon), so the ruff: ignore suppresses all of the same N803 diagnostics as the disable/enable pair above.

ruff: file-ignore comments can be used to suppress diagnostics for the entire file, just like ruff: noqa comments:

# ruff: file-ignore[F401] Allow unused imports in this file  import foo import bar import baz

As this example also shows, each of these comment kinds can have an associated \"reason\" explaining why they were added, in this case Allow unused imports in this file.

ruff: ignore comments can be added automatically with the new --add-ignore CLI flag, and in preview, all of these ruff suppression comments support rule names instead of codes:

❯ echo 'import math' > try.py ❯ uvx ruff@latest check --preview --add-ignore try.py Added 1 ignore comment. ❯ cat try.py import math # ruff: ignore[unused-import]

You can find the full specification for all of these comments in the documentation.

Fixes are now shown in check and format --check output #

Ruff now shows the diff for linter and formatter fixes when rendering diagnostics.

Both the check and format subcommands have long supported the --diff flag for showing the changes introduced by applying fixes with check --fix or format, but this was separate from the normal output and suppressed the accompanying explanatory diagnostics. In v0.16, available fixes are now shown as part of the default full output format, rendered below the help subdiagnostic:

\"Ruff

The same is true for format --check. Given the following input:

# example.py  if True:  pass elif False:  pass

The formatter produces:

\"Ruff

format --check also now supports the full selection of output formats supported by the linter. You can use this to obtain machine-readable JSON output or produce the formats expected by GitHub and GitLab to render annotations in CI, as just a couple of examples. See the CLI help or documentation for the full list of supported formats.

One final note on output formats is that there is a small breaking change to the JSON output in v0.16. The filename, location, end_location, fix.edits[].location, and fix.edits[].end_location fields may now be null rather than defaulting to the empty string and row 1, column 1, respectively. This should affect very few of Ruff's existing diagnostics but better reflects the internal diagnostic representation and may become more common in future rules.

Rule stabilizations #

The following rules have been stabilized and are no longer in preview:

Other behavior stabilizations #

This release also stabilizes some additional behavior, previously only available in preview mode:

Thank you! #

Thank you to everyone who provided feedback regarding the changes included in Ruff's preview mode and to our contributors. It's an honor building Ruff with you!

View the full changelog on GitHub.

Read more about Astral — the company behind Ruff.

Thanks to Zanie Blue, David Peter, Micha Reiser, and Alex Waygood who contributed to this blog post.

\n\"\"","excerpt":"The next stable version of Ruff is out now.","image":"https://astral.sh/static/PNG/Ruff_v_0_16_0.png","authors":[{"name":null,"url":"https://astral.sh/blog/ruff-v0.16.0","avatar":"https://astral.sh/static/favicon-32x32.png"}],"id":"49056112","url":"https://astral.sh/blog/ruff-v0.16.0","external_url":"https://news.ycombinator.com/item?id=49056112","date_published":"2026-07-26T09:01:39Z"},{"title":"Netflix employee fired for sharing personal details in retreat trust exercise","content_html":"

A Netflix executive was fired from his $1.1 million a year job after revealing during a “trust exercise” at a work retreat that he had taken medically prescribed ketamine, a lawsuit has claimed.

Kevin Baillie, who was vice president and head of creative at Eyeline Studios, is suing the company after it launched an investigation into his comments that ultimately ended in his firing, the papers say. 

Baillie, who’s been on the visual effects team for “Pirates of the Caribbean” and the Harry Potter franchise, says he took the drug under medical supervision in October and November of 2022 at a Santa Barbara clinic.

\"Kevin

Kevin Baillie, ex-vice president and head of creative at Eyeline Studios, a division of Netflix. Instagram/fxnerd

\"Joe

Kevin Baillie (right) at the 24th Satellite Awards in Beverly Hills. I Hasegawa/HNW-Photo/Plux / Shutterstock

He sought the treatment for clinical depression after the death of his mother, according to the suit.

During what’s called a “Vulnerability-Trust exercise” at a January 2026 retreat at the exclusive Sendero Ranch, a Northern California property owned by Netflix, Baillie shared with his colleagues that he had undergone the treatment, the suit says.

Baillie claims he explained the reason why he had taken the drug but was investigated by Netflix. On March 18, 2026 a company investigator brought the incident up, “in a manner suggesting suspicion of recreational drug use,” the suit reads.

\"A

The retreat took place at the exclusive Sendero Ranch, a Northern California property owned by Netflix. Instagram/cecimeseeds

\"A

Baillie was fired in April, with Netflix’s attorney confirming, “The ketamine therapy issue has factored into the termination.” Instagram/fxnerd

The executive was fired in April with the company’s attorney confirming “the ketamine therapy issue has factored into the termination,” the papers say, and go on to suggest Baille was denied up to a year of severance pay.

Baillie says in the lawsuit the “scope” of the investigation related to alleged profanity and drinking. He had been warned during his performance review that he should “drop one or two less f-bombs but don’t stop entirely.”

It goes on to say that at the same retreat Baillie drank a Guinness standing on his head, after sharing that he had learned the trick from his former father in law during a conversation inspired by the trust session.

“His colleague immediately asked for a demonstration, rather than withhold the openness that the session had encouraged, he performed the trick,” the papers say.

\"Ribbon

Baille also suggests an alcohol-fueled company environment bolstered by Eyeline Studios CEO Jeff Shapiro (pictured). Netflix

Sign up for the California Morning Report newsletter

California's top news, sports and entertainment delivered to your inbox every day.

Thanks for signing up!

Baille also paints a picture of an allegedly alcohol-fueled company environment encouraged by the Eyeline Studios CEO Jeff Shapiro.

The lawsuit alleged “alcohol consumption was company-sponsored, leadership-modeled and condoned”, with multiple examples of how Shapiro “set the cultural tone concerning alcohol at the executive level”.

The documents claimed Shapiro on one occasion purchased beer at a corner store and brought it to a company car ride to the Visual Effects Society Awards for staff to share.

Baillie claims he also witnessed the CEO consume alcohol with Netflix and Eyeline employees at his own welcome dinner in September 2024, the Netflix Annual Business Review events in March 2025 and even a Lakers game attended by Netflix execs including Shapiro’s direct supervisor in February 2026.

All up, Baillie’s attorneys provided over half a dozen examples of the CEO being present at a work event with a drink in his hand, according to court papers.

Download The California Post App, follow us on social, and subscribe to our newsletters

California Post News: Facebook, Instagram, TikTok, X, YouTube, WhatsApp, LinkedIn
California Post Sports Facebook, Instagram, TikTok, YouTube, X
California Post Opinion
California Post Newsletters: Sign up here!
California Post App: Download here!
Home delivery: Sign up here!
Page Six Hollywood: Sign up here!

In addition to hosting multiple parties, Shapiro also had a personal bar in his office “from which he served alcohol (to Baille) including after a successful meeting with Netflix’s CEO Ted Sarandos,” according to the papers.

Baillie is asking for a jury trial, compensatory damages, lost wages, damages for emotional distress, and punitive damages.

Netflix and Eyeline were contacted for comment.

\n\"\"","excerpt":"He was fired from his $1.1 million a year job after revealing that he had taken medically prescribed Ketamine at a “Trust exercise”","image":"https://nypost.com/wp-content/uploads/sites/2/2026/07/netflix-exec-ketamine-firing-trust-exercise-comp.jpg?quality=75&strip=all&w=1200","authors":[{"name":"California Post","url":"https://nypost.com/2026/07/26/us-news/netflix-exec-goes-ballistic-after-being-fired-for-stunning-trust-exercise-confession-at-retreat-suit/","avatar":"https://nypost.com/wp-content/themes/nypost/static/images/favicons/favicon.ico"}],"id":"49076923","url":"https://nypost.com/2026/07/26/us-news/netflix-exec-goes-ballistic-after-being-fired-for-stunning-trust-exercise-confession-at-retreat-suit/","external_url":"https://news.ycombinator.com/item?id=49076923","date_published":"2026-07-27T23:21:37Z"},{"id":"49057574","url":"https://www.wsj.com/tech/google-discloses-94-1-billion-in-spacex-stock-marking-6-stake-91655d7c","external_url":"https://news.ycombinator.com/item?id=49057574","title":"Google Discloses $94.1B in SpaceX Stock, Marking 6% Stake","date_published":"2026-07-26T12:43:21Z"},{"title":"Decathlon Germany adds Wero payment option to decathlon.de website","content_html":"

Decathlon didn’t just add a new payment button to its German checkout page. It handed European banks their first big-box sporting goods test case at a moment when competition in digital payments is intensifying and the push for EU-made payment infrastructure is growing.

Decathlon switched on Wero on decathlon.de on July 20, making Germany the first market in the French group’s international footprint to carry the European payment scheme.

Wero is the retail payment product of the European Payments Initiative (EPI), a consortium of banks and payment providers backed by 18 shareholder institutions and more than 50 members across the region. It routes payments directly between bank accounts in real time, authorized in the customer’s banking app via fingerprint or face recognition, so the merchant never receives a card number.

Archive reporting shows the scale Wero is still building toward: 43 million active users across the region in August 2025, 46 million by November 2025 alongside the retail launch, and around 56 million by the time of the Decathlon rollout in July 2026, with German registrations rising from 1.3 million to 1.8 million over the same period.

Decathlon joins a merchant roster that already includes Eventim Live, with Lidl, Rossmann and Hornbach expected to follow.

The insider why

For Decathlon, the incentive is not just optics. Patrick Müller, chief digital and chief marketing officer at Decathlon Germany, tied the launch to the retailer’s membership program, saying the connection delivers “a genuine added value” on every purchase. That frames Wero as a loyalty and retention tool, not only a checkout option. Industry estimates commonly place card network processing costs at around 1 to 2 percent of transaction value in fees to Visa and Mastercard. An account to account rail that bypasses that layer can be a direct margin lever for a retailer operating at scale on thin sporting goods margins.

The rollout plan extends beyond the website.

Decathlon says it is preparing the technical integration needed to bring Wero to its roughly 110 physical stores in Germany, with the goal of a consistent payment experience across online and offline channels. A joint activation campaign with EPI is scheduled for October: a six week promotion offering a €10 voucher to customers who pay with Wero on decathlon.de, funded entirely by EPI rather than Decathlon.

The challenge and the promise

EPI’s willingness to fund customer incentives points to the network’s core problem. Payment schemes live or die on how many merchants and consumers show up at the same time, and Wero still lacks in store point of sale functionality. That capability is not expected until 2026 and 2027. Joachim Schmalzl, EPI’s supervisory board chairman, has acknowledged the initiative still faces real hurdles even as he called the ecommerce launch a milestone toward “a sovereign European payment system.”

Decathlon being among the first to implement the system can help test Wero’s promise: lower processing costs, tighter loyalty integration and one less transaction fee flowing to a US based network, provided Wero can close the gap between its ambitions and its still limited in store reach.

\n\"\"","excerpt":"Decathlon adds the European bank-backed Wero payment option to decathlon.de, with in-store rollout in preparation","image":"https://d1f00kj7ad54bu.cloudfront.net/Pictures/1024x536/9/6/4/60964_wero_807274_crop.jpg","authors":[{"name":"Sporting Goods Intelligence","url":"https://www.sgieurope.com/e-commerce/decathlon-germany-launches-wero-payment-on-its-website/122397.article","avatar":"https://www.sgieurope.com/magazine/dest/graphics/favicons/favicon-32x32.png"}],"id":"49072310","url":"https://www.sgieurope.com/e-commerce/decathlon-germany-launches-wero-payment-on-its-website/122397.article","external_url":"https://news.ycombinator.com/item?id=49072310","date_published":"2026-07-27T16:49:08Z"}]}