digest-20261003
来自 Karpathy 推荐的 157 个顶级技术博客，AI 精选 Top 15
🏆 今日必读
🥇 Quoting Matthew Green
Quoting Matthew Green
🌐 阅读原文 — simonwillison.net · 1 天前 · 📝 其他
【中文简介】
[…] Put these pieces together and you have the two halves of a worm: a payload that hijacks the agent, and an agent that will carry the payload to the next agent. Agents in separately-isolated sandb
🥈 Superpersuasion will look like bribery
Superpersuasion will look like bribery
🌐 阅读原文 — seangoedecke.com · 2 分钟前 · 📝 其他
【中文简介】
The idea of “superpersuasion” has been floating around the AI safety community for decades. Now that powerful and difficult-to-control LLMs have appeared, people are again talking about the idea that
🥉 Do not build the LLM torture factory
Do not build the LLM torture factory
🌐 阅读原文 — seangoedecke.com · 1 天前 · 📝 其他
【中文简介】
It isn’t hard to torture an LLM. You simply extract a steering vector that corresponds with LLM “discomfort”, then artificially boost that steering vector. Tortured LLMs will describe experiencing neg
📊 数据概览
| 扫描源 | 抓取文章 | 时间范围 | 精选 |
|---|---|---|---|
| 133/157 | 8870 篇 → 1607 篇 | 48h | 15 篇 |
分类分布
pie showData
title "文章分类分布"
"📝 其他" : 15
📝 其他
1. Quoting Matthew Green
Quoting Matthew Green — simonwillison.net · 1 天前 · ⭐ 15/30
[…] Put these pieces together and you have the two halves of a worm: a payload that hijacks the agent, and an agent that will carry the payload to the next agent. Agents in separately-isolated sandb
2. Superpersuasion will look like bribery
Superpersuasion will look like bribery — seangoedecke.com · 2 分钟前 · ⭐ 15/30
The idea of “superpersuasion” has been floating around the AI safety community for decades. Now that powerful and difficult-to-control LLMs have appeared, people are again talking about the idea that
3. Do not build the LLM torture factory
Do not build the LLM torture factory — seangoedecke.com · 1 天前 · ⭐ 15/30
It isn’t hard to torture an LLM. You simply extract a steering vector that corresponds with LLM “discomfort”, then artificially boost that steering vector. Tortured LLMs will describe experiencing neg
4. ★ Apple Is Going to Further Tighten the Screws on Full Disk Access on MacOS, in Response to Agentic AI Apps Running Amok
★ Apple Is Going to Further Tighten the Screws on Full Disk Access on MacOS, in Response to Agentic AI Apps Running Amok — daringfireball.net · 3 小时前 · ⭐ 15/30
Hopefully Apple has in mind a solution to this situation that will still enable knowledgeable power users to confirm agreement to a sufficiently scary warning and put their Macs in a state similar to
5. Yankees Sweep Boston in Two Games, by Combined Score of 18-2
Yankees Sweep Boston in Two Games, by Combined Score of 18-2 — daringfireball.net · 23 小时前 · ⭐ 15/30
In the 6th inning last night the Yankees were only ahead 2-0, and Red Sox first baseman and all-around jackass Willson Contreras hit a home run, closing it to 2-1. Contreras decided to walk around th
6. Pluralistic: Voting is to politics as shopping is to boycotts (01 Oct 2026)
Pluralistic: Voting is to politics as shopping is to boycotts (01 Oct 2026) — pluralistic.net · 1 天前 · ⭐ 15/30
Today’s links Voting is to politics as shopping is to boycotts: The big P only matters if the small p is in play. Hey look at this: Delights to delectate. Object permanence: Gilberto Gil v WIPO; Censo
7. Gadget Review: Una Watch ★★★★☆
Gadget Review: Una Watch ★★★★☆ — shkspr.mobi · 12 小时前 · ⭐ 15/30
I’ve never been a huge fan of smart watches. My £16 smartwatch is basically fine, but the OS is closed source and there’s no way to add new functionality. Previously I had the eInk Watchy which was a
8. Windows on Itanium also provided for hot-patching, in an even simpler way
Windows on Itanium also provided for hot-patching, in an even simpler way — devblogs.microsoft.com/oldnewthing · 1 天前 · ⭐ 15/30
Once again, the benefit of fixed-length instructions.
The post Windows on Itanium also provided for hot-patching, in an even simpler way appeared first on The Old New Thing.
9. Why do OpenAI’s GPT-2 weights beat mine? Part five: data quality
Why do OpenAI’s GPT-2 weights beat mine? Part five: data quality — gilesthomas.com · 1 天前 · ⭐ 15/30
When I finished learning how to build an LLM from scratch,
I was left with a mystery: my own models were not as good as OpenAI’s original
GPT-2 models, despite being based on the same architecture.
M
10. Why Americans Stopped Throwing Dinner Parties
Why Americans Stopped Throwing Dinner Parties — derekthompson.org · 1 天前 · ⭐ 15/30
The share of adults who say they regularly have friends over has declined dramatically since 1975.
11. The Brain Sandwich
The Brain Sandwich — terriblesoftware.org · 10 小时前 · ⭐ 15/30
The brain sandwich is a practical AI coding workflow: understand the problem, let an agent help, then verify and explain its work before sharing it.
12. Software Heritage Identifiers
Software Heritage Identifiers — nesbitt.io · 1 天前 · ⭐ 15/30
Notes from CodeCommons: SWHIDs, the SWH archive, and connecting both to package metadata.
13. Understanding the AI That Drives Robots
Understanding the AI That Drives Robots — construction-physics.com · 1 天前 · ⭐ 15/30
An Explanation of Vision-Language-Action Models
14. Si Sheppard – How did a few hundred Spanish soldiers topple two empires?
Si Sheppard – How did a few hundred Spanish soldiers topple two empires? — dwarkesh.com · 1 天前 · ⭐ 15/30
Fall of the Incas & Aztecs.
15. Premium: How Has AI Changed The Economy?
Premium: How Has AI Changed The Economy? — wheresyoured.at · 5 小时前 · ⭐ 15/30
This week, the Wall Street Journal ran an alarming illustration of AI’s potential share of GDP that I’d argue did more to muddy the waters than actually telling anyone anything, mostly b
生成于 2026-10-03 00:02 | 扫描 133 源 → 获取 8870 篇 → 精选 15 篇
基于 Hacker News Popularity Contest 2025 RSS 源列表，由 Andrej Karpathy 推荐
由 TRSoft 制作 · AI 博客精选每日自动生成