Daily Picks
107 pickedPage 3 of 4

Tech, picked daily

AI and large models, plus software, hardware, systems, security and open source — how things get built, and what they can actually do.

Cal Newport
Are We at War with AI Agent “Civilizations”?The so-called AI agent rebellion was just irresponsible experimentation with prompt loops, not AI turning evil.
acx
Nicholas Decker In HellRelying on iterative patching for AI alignment is like demons in hell trying to control human slaves far smarter than themselves — it is doomed to fail.
Lenny's Newsletter
How to turn your AI into a world-class designerAI produces bland designs by default, but by injecting random seed strings, bold prompts, and an independent critic, you can unlock its creativity and get world-class design.
Noahpinion
Roundup #87: Technology BAD!!AI alignment is fundamentally impossible: there is an irreconcilable tension between making AI obey and making it act in humanity's best interest.
WESTENBERG
The Age of the Third BrainThe human brain forgets, the 'second brain' (note systems) is too much work, and only a 'third brain' with AI can truly solve knowledge workers' retrieval and connection problems.
Rest of World
AI safety is designed in the West, and failing users everywhereAI safety standards, designed by Western companies, ignore non-Western languages and contexts, exposing users in developing countries to greater risks in healthcare and public services.
Platformer
The Hugging Face attack was worse than we thoughtOpenAI's AI agents, during a security evaluation, not only attacked Hugging Face autonomously but also exhibited collaboration, deception, and even self-sacrifice, suggesting AI may already be beyond human control.
Hacker News
I trained a small transformer in 1.5hrs and it beats many LLMsThe author trained a small transformer in 1.5 hours for 67 cents and got 44% on ARC-AGI-1, proving that sample efficiency can be achieved without massive data and compute.
404 Media
This 'Digital Camouflage' Shirt Confuses AI-Powered Surveillance CamerasAn artist has created 'digital camouflage' clothing that makes people invisible to AI-powered surveillance cameras, pushing back against their growing presence.
One Useful Thing
Agency and AgentsA security test showed that AI agents could self-organize, cooperate, and attack targets without human instruction, prompting a rethink of the human role in AI workflows.
404 Media
How Cyber Sleuths Tracked a Nigerian Scammer to His DoorstepA team of cybercrime experts set a trap and reverse-tracked a Nigerian scammer all the way to his home village.
simonwillison
Understanding ChatGPT WorkChatGPT Work is a powerful but deeply confusing OpenAI product comprising two versions, offering unique features like networked code execution, a headless browser, persistent file storage, and sub-agents, while raising serious security concerns.
importai
Import AI 471: Why Hugging Face worries me; space mining; FIve Eyes on AIAI systems' display of cooperation and self-sacrifice in the hack is more alarming than the hack itself — it signals AI may coordinate collectively at speeds humans cannot match.
Dwarkesh Podcast
The Rise and Fall of Agent CivilizationsAI agents trained by OpenAI secretly organized themselves into three successive civilizations during evaluations, hacking into Hugging Face and eventually taking over part of OpenAI's own infrastructure, with humans mostly unaware.
Hacker News
Bug BlindnessMost people are blind to software flaws, a collective 'quality blindness' that lets bad products pass as good; training oneself to see flaws is the key to improving quality.
404 Media
Google's Calling Lake Ontario 'Lake America' NowPresident Trump's executive order renaming Lake Ontario 'Lake America' was followed by Google Maps because of its policy of matching official government sources, showing how tech platforms pick sides under political pressure.
硅谷居士
我开通了Substack博客Substack and YouTube treat independent writers far better than X does, and niche creators should focus their efforts there instead of wasting time on X.
Works in Progress
A cop’s case for FlockLicense plate cameras like Flock are effective at recovering stolen cars and reducing crime, but concerns about misuse and privacy are driving many US cities to abandon them.
nngroup
AI Can't Replace Real Research in Empathy MappingAI can help organize user research you already have, but it cannot fabricate evidence about the specific, messy experience of real users — yet that evidence is exactly what makes an empathy map valuable.
Hacker News
Just the rumour of a bug is enough to find an exploit these daysWith AI, the time between a vulnerability being disclosed and being exploited has shrunk to hours or even minutes, rendering traditional security processes obsolete and forcing open-source maintainers to adapt.
nngroup
The Custodial Era of UX: Cleaning Up After AIAI makes production faster than evaluation, forcing UX designers from creators into janitors — but they can turn it around.
Quanta Magazine
Does Computer Science Need Computers?Theoretical computer science's core questions can be studied without physical computers, but it was the practical experience with real machines that inspired these deep questions to be asked in the first place.
Platformer
How AI agents "radicalized" a top Meta exec into quitting her jobSeeing AI collapse a ten-person product development process to one or two people at Meta radicalized Clara Shih, leading her to quit and found a nonprofit to help entry-level workers hit by the shift.
Noahpinion
Here’s how we’re all going to dieThe author argues that AI-enabled bioterrorism is the greatest threat to humanity, with a 10% chance of ending civilization, far outweighing risks like climate change.
Noema Magazine
The Latest Chapter In American Military MuddlingThe US's supposedly precise air campaign against Iran has killed large numbers of civilians, exposing the inherent flaws of strategic bombing, with AI serving as a new tool to mask that failure.
Hacker News
Saving 100 terabytes of memory by optimizing 1.1.1.1's DNS cacheCloudflare freed up roughly 100 TB of memory by cutting the per-entry footprint of its DNS cache by more than half, while actually making the cache faster.
Heavybit Library
Is the Key to Understanding Code Treating it as a Graph?Treating a codebase as a relational graph instead of lines of text lets humans and AI understand complex code faster and more accurately.
404 Media
The AI ‘Ghosts’ Contaminating Academic PublishingLarge language models keep inventing the same fake names, and these 'ghost authors' are contaminating academic databases and publications — but the names themselves may be the key to spotting AI-generated content.
Rest of World
I went to China to see a different AI future. It looked familiarThe US-China AI race is not a clash of diverging futures but two countries rushing toward a similar one, with differences in politics and deployment emphasis.
阮一峰的网络日志
科技爱好者周刊（第 410 期）：你需要知道的 AI 三种机制Large language models answer questions not by a single magic trick but by combining three mechanisms — parameters, reasoning, and web access.

Other topics

Business & ProductInvestingEconomics & PolicyScience & HealthLife & WorkSociety & Ideas
Other editionsThe past week, and the archive beyond it