Warns large release candidates ‘are *not* conducive to long-term stability’

Linux kernel boss Linus Torvalds has signaled he’ll push back when he receives irrelevant pull requests, after complaining that developers are making badly timed and trivial submissions, sometimes after using AI to review code.

Torvalds foreshadowed changes in his weekly state of the kernel update , which on Sunday announced the release of a fifth release candidate for version 7.1 of the Linux kernel.

“To the surprise of absolutely nobody by now, rc5 is pretty big. Quite a bit bigger than rc5's have traditionally been,” Torvalds wrote, before revealing “I'm not entirely happy about it - most of this is totally trivial stuff to random drivers, which obviously makes it all less scary, but at the same time I'm really not convinced the churn is worth it at rc5 time.”

The Linux kernel development cycle usually sees Torvalds open a two-week window during which contributors submit code they hope will make it into the next release. Seven release candidates (rc1-7) follow, with each supposed to represent a step towards delivering a stable update. Revised code always arrives during that process. But high volumes of new contributions to rc5 add complexity at a time work on the new kernel is usually close to completion.

Linux kernel flaw opens root-only files to unprivileged users

When wolves bring down prey in Yellowstone National Park, ravens often appear almost immediately. Long before the predators finish feeding, the birds gather nearby to grab scraps of meat. Their ability to locate fresh kills so quickly has puzzled observers for years, leading many people to assume that ravens simply follow wolves across the landscape.

A new study suggests the real explanation is far more impressive. After tracking ravens and wolves in Yellowstone for more than two years, researchers discovered that ravens rely heavily on memory. Instead of shadowing wolf packs over long distances, the birds remember places where wolves frequently make kills and return to those areas later.

“They can fly six hours non-stop, straight to a kill site,” says Dr. Matthias Loretto, the study’s lead author.

The findings, published in Science, indicate that ravens use spatial memory and navigation skills to search for food spread across large areas. According to Loretto, ravens do not need to stay close to wolves all the time because they can recall where food is most likely to appear. “Ravens can cover large distances by flying, and they seem to have a good memory, so they don’t need to constantly follow wolves in order to profit from the predators,” he says.

The project was led by the Research Institute of Wildlife Ecology at the University of Veterinary Medicine Vienna and the Max Planck Institute of Animal Behavior (Germany), along with several international partners, including the Senckenberg Biodiversity and Climate Research Centre (Germany); School of Environmental and Forest Sciences at the University of Washington (USA); and Yellowstone National Park (USA).

Researchers carried out the study in Yellowstone National Park, where wolves were reintroduced in the mid-90s after being absent for 70 years. About one quarter of the park’s wolves wear tracking collars each year, allowing scientists to monitor their movements.

I just wanted to ship my app. Turns out that now costs $48 and a few guys in Afghanistan.

My second app was done. Tested, polished, ready. I figured shipping would be easy.

Google wanted 20 testers, for 14 consecutive days. 1 Four or five people, sure. But 20? For two weeks straight? That’s a part-time job on top of the app itself.

Googling the requirement only made it worse.

Hundreds of threads. Developers who met the requirement on paper and still got denied. Turns out there are hidden reasons Google won’t tell you upfront — testers weren’t “engaged enough,” you didn’t act on feedback, you didn’t push an update mid-period. None of it is in the official policy. It just appears in your rejection email. 2

And in Canada Apple controls 65% of the smartphone market, 3 probably 80% inside my friend group. I don’t think I know 20 people with an Android.

I was using Codex to build a side project over the weekend. A lazy vibe-coding session with no wireframes, Figma screens, or specs. Just an idea and an endless prompt thread with an AI. As I ran out of tokens at record speed, I realized that the way I think as a designer is fundamentally incompatible with how AI coding tools are meant to be used. The way designers think, aka the design proces s When I start designing something, I don't know what the final outcome will look like. I have a feeling. A loose sense of what the experience should be. The layout, style, flows, interactions, edge cases are details that emerge through making and exploring. Here's what this looks like translated into Codex: I give it a vague prompt. I see what it creates. I react. I give it more specific instructions. We go back and forth, reverting, detailing, testing and fixing. Iterations in, the product slowly starts to reveal itself. This is the process. This is design. It is through loops of this process that the core problem emerges. Tools like Codex and Claude Code are built for workflows where you know exactly what you want. An engineer using Claude Code starts with a clear description of what should be built and the AI implements it. The output is controlled and predetermined. The tool is a medium for execution, not exploration. But a designer is not doing that. I don't have the answer at the start. I'm finding the answer by building. The spectrum, as it stands

San Francisco immigration court shuts down after purge of judges, leaving asylum cases in chaos

Oliverio Mora Huerta, left, is interviewed with his family before going to Immigration Court Tuesday, May 12, 2026, in Concord, Calif. (AP Photo/Jeff Chiu)

Attorney Dr. Nidaa Pervaiz is interviewed outside of Immigration Court Tuesday, May 12, 2026, in Concord, Calif. (AP Photo/Jeff Chiu)

An interfaith community group holds signs while singing outside of Immigration Court Tuesday, May 12, 2026, in Concord, Calif. (AP Photo/Jeff Chiu)

SAN FRANCISCO (AP) — There are no immigrants waiting for rulings anymore at San Francisco’s main immigration court, no lawyers making arguments.

It's unclear right now how AI is going to play out for most companies, and I don't think anyone has a clean answer yet, including me. But there's a pattern I keep coming back to, and it has less to do with what AI eventually becomes and more to do with what it can already do.

I don't think the capability curve breaks at some single moment we'd call AGI. It just keeps climbing. Each release adds capability somewhere, and we don't need to reach the top of the curve for the bottom of it to start reshaping things.

This past Tuesday at Google I/O, Antigravity 2.0 built a functioning operating system from scratch in twelve hours. It spun up 93 sub-agents, processed 2.6 billion tokens, ran roughly 15,000 model requests, and cost less than a thousand dollars in API credits. Then they ran Doom on it live. When the keyboard drivers were missing, they asked the agent to write them on stage, and it did.

Take the staging with whatever grain of salt you want. The point underneath is what "good enough" looks like in mid 2026. Many small agents, running in parallel, cheaply, reliably enough to compose into something that actually boots. That's the engine worth paying attention to. Not because of where it ends up, but because of what it can already do.

A capable agent loop, called many times in parallel, with reasonable cost and reasonable latency, is enough to recreate most of what the application layer of software currently sells. The curve keeps going from here. The question that follows is which kinds of companies sit downstream of that engine and which don't.

The companies pulling ahead right now share a property that has very little to do with their marketing and not much to do with strategy either. Their products are increasingly being consumed by software, not by people.

Every cloud provider is racing to sell agents a VM. Freestyle.sh , Exe.dev , Vercel, Cloudflare, Google are all converging on the same primitive: spin up a sandbox, let the agent run a script, tear it down. For short-lived sessions, that’s the right primitive. For anything that needs to outlive the session, it breaks.

Agents are being asked to operate as if they’re running locally, but that’s not how production systems work. Developers reach for Docker Compose, Kubernetes, Terraform, and cloud services because a node isn’t enough.

We’ve seen this movie before. In the 2010s, companies outgrew VMs the same way agents are about to. They needed multi-service architectures, custom networking, stateful workloads, horizontal scale, cost control. Kubernetes won because it gave them the primitives of a cloud they could own and operate. Not a box, but a platform.

The base unit shouldn’t be a VM an agent rents. It should be an isolated cloud an agent owns. VMs, storage, networking, identity, DNS, all as composable resources inside that cloud. Basically the agent gets root over a mini-AWS, not SSH into a box.

The obvious move is to give each agent its own Kubernetes cluster. Clusters are the closest existing primitive to “a cloud you own”: namespaces, RBAC, networking, storage, a real API. But the moment you try to operate this at agent scale, every assumption in the stack pushes back.

A cluster takes minutes to provision, not seconds. A control plane costs more per month than most agent sessions will ever generate in value. Multi-tenancy inside a cluster is too soft for untrusted code; one cluster per tenant is too expensive to hand out by the million. The lifecycle is wrong too. Clusters are long-lived infrastructure managed by humans, but agent workloads are bursty, short, and created by software that doesn’t care about upgrade windows.

You can get new post notifications through RSS, email, Twitter or Facebook

I have just read a remarkably silly article in a popular news website. The apparently well educated author has decided her seven year old has to study towards job opportunities, and has reacted to AI by forcing the poor child to study “coding” (which AIs can do) instead of drama (which they cannot).

The reasoning in the article is a very confused mix of what 16 year olds think they have to study to get a job (for some reason they are treated as authorities on this subject) and the fear that “AI” will wipe out all non-STEM jobs. The author is educated in the humanities (a History degree from Oxford), has a PhD, and is an author. Rather than drawing on her own life experience, or researching the subject, or even discussing it with people who have some knowledge of the issues, she takes her 16 year old daughter’s pronouncement “best steer clear of anything that leads to a yak yak degree” (humanities) as unarguable.

I can understand teenagers who face the prospect of student debt and high tuition fees feeling pressured to do vocational subjects. The author had freedom of not having to pay tuition fees and receiving a government grant towards living expenses. Something young people no longer have. However, an adult with some life experience should be able to to be able to assess things more calmly, and an educated person should be capable of critical thinking and questioning the received wisdom. A 16 year old might not understand that it is harder to get a job when the economic cycle is against you (we have an asset price bubble looking like its about to pop, and multiple wars), but someone who has lived a few decades has seen this all before.

Education is not vocational training, and does not have to be. Plenty of people who study the humanities get good jobs. There is a lot to be said for developing your mind by spending three years studying something you love.

It is even worse with regard to her seven year old. The child must start studying vocational skills now. Broader mental development is a waste of time. The worst part of this is an utterly deluded decision to force a seven year old to study only useful skills, sacrificing drama for “coding”.

In December, Order of the Sinking Star was announced . Jonathan Blow's new game that's been a decade in the making. Understandably, the focus has been on the game and its insane amount of puzzles (apparently ~1400). But the engine itself is also planned to go open-source after release:

"...there’s a plan to release its custom engine for free as an open-source project shortly after the game ships."

Seems to have been flying under the radar. (Or alternatively I've been living under a rock.)

No information on the license of the engine yet. Fingers crossed for a permissive one. Either way, big news.

For years my side projects lived on Azure. App Services here, a Static Web App there, a managed SQL database, a sprinkle of Functions, some blob storage, a Communication Services resource for email. Each one made sense the day I created it. Added up over a dozen-plus projects, it became a sprawling estate with a monthly bill that kept quietly creeping — and a mental tax every time I opened the portal and saw resource groups I half-remembered.

This month I moved all of it onto a single VPS. Every API, every static site, every Blazor app, the databases, the object storage, even the transactional email. Azure now shows zero resource groups .

A year ago I wouldn’t have attempted this. It would have been weeks of fiddly, repetitive, error-prone work. But with an LLM coding agent doing the heavy lifting, it turned into something I could knock out in focused sessions between actual product work. That shift is the real story here, so let me walk through what I did, what it cost (and saved), the honest downsides, and why I think this is a no-brainer move for anyone bootstrapping.

The destination is deliberately boring: one solidly-specced VPS, plus a few free or near-free services stitched around it.

The whole public surface is fronted by Cloudflare, so I also get caching, DDoS protection, and analytics thrown in without lifting a finger.

The variety is the point — this wasn’t one app, it was a zoo:

We’re releasing a new paper that explains our views on the competition on AI between the US and China.

It’s essential that the US and its allies stay ahead of authoritarian governments like the Chinese Communist Party, or CCP. AI will soon become powerful enough to be used to repress citizens at unprecedented scale, and even to alter the balance of power among nations . And since AI is advancing more quickly by the day, we have only a limited period of time to set the conditions of the competition—and determine whether and how those threats materialize. It’s with this in mind that we outline what’s required to ensure America stays ahead.

The most important ingredient for developing AI is access to the computer chips on which the models are trained (or “compute”). Since the most capable chips are developed by American companies, the US government currently limits China’s supply by enforcing tight export controls on them. Recent history suggests these controls have been incredibly successful. In fact, AI labs in China have only built models close in intelligence to America’s because of their talent, their knack for exploiting loopholes around these export controls, and their large-scale distillation attacks that illicitly extract the innovations of American companies.

In this post, we present two scenarios for what the world might look like in 2028, when we expect transformative AI systems to have arrived.

In the first scenario, America has successfully defended its compute advantage. Policymakers have acted to tighten export controls further, disrupt China’s distillation attacks, and further accelerate democracies’ adoption of AI. In this world, democracies set the rules and norms around AI. It’s also in this scenario that we’re most likely to successfully engage with China on safety, which we’re supportive of to the extent this is possible.

In the second scenario, America has chosen not to act. Policymakers have not tightened loopholes on the CCP’s access to compute, and AI firms in China have quickly taken advantage—catching up to the frontier and even overtaking America. In this world, AI norms and rules are shaped by authoritarian regimes, and the best models enable automated repression at scale. It will be no solace that this authoritarian triumph has happened on the back of American compute.

In March, 2020, we published a book titled “Software Engineering at Google” curated by Titus Winters, Tom Manshreck and Hyrum Wright.

The Software Engineering at Google book (“SWE Book”) is not about programming, per se, but about the engineering practices utilized at Google to make their codebase sustainable and healthy. (These practices are paramount for common infrastructural code such as Abseil.)

We are happy to announce that we are providing a digital version of this book in HTML free of charge. Of course, we encourage you to get yourself a hard copy from O’Reilly if you wish.

Digital copy of Software Engineering at Google curated by Titus Winters, Tom Manshreck, and Hyrum Wright is made available on abseil.io under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Public License .

politics thoughts · 4 comments · 650 words · Viewed ~515 times

In the UK, it is illegal to buy alcohol if you are under 18 .

Similarly, in most countries, you cannot vote until you have reached a specific age.

These are age-gates. You do not need to prove your competence to drink, vote, smoke, or get married; you just need to be old enough.

Some things have skill-gates. If you want an amateur radio licence in the UK, you need to pass an exam. You can be any age 0 .

Similarly, most jurisdictions allow you to get a medical licence once you have passed the requisite tests 1 .

Reducio analyzes your prompt structure and strips redundant tokens without altering semantic meaning. Your model receives a leaner input and returns the same quality output.

Capture, extract, and search everything you see online — instantly recall any screenshot, document, or note.

You already have five ways to capture a screen. That was never the problem. The problem starts after capture.

SnapIndex picks up exactly where your screenshot tool drops off.

Extract → index → annotate → retrieve. One system, not seven.

Seven moves that turn captures into a system

Capture is step one. Retrieval is everything.