jlhawn

There aren't any _good_ elements that start with the letters D or E. Imagine if they name the next one "Dysprosium" where the Greek root translates to "hard to get at" or "difficult to access" reflecting how difficult it was for its discovers to isolate it in the late 19th century, but also sounds kind of untrustworthy to give an AI model such a name.

Bonjour!

True! there are lots of layers of abstraction written around that core `nsenter` piece of libcontainer. I just take issue with the "Go is kind of a terrible language for a container runtime" claim made by the OP. The maintainers' choice to use an expressive high-level language for most of the codebase is a practical one, I think. They also knew when it was not the right tool for the job and when to pull out the C bindings or how to work-around issues with forking the process. The OP mentions this, but for fork/re-exec helpers in the engine, this is where it lives now: [1]. I specifically remember needing to use something like this for the `docker-untar` helper during image pulls (I worked primarily on the container image filesystem-setup part of Docker) so that extraction of the image layer archives could be done in a chroot with fewer privileges. It was a security-sensitive part of the docker daemon and previously there were all sorts of malicious ways that a container image could be built so that contents are extracted _outside_ of the place they should be. This is the "images are just folders" part of OP's article. Another part I find funny is the "Containers really aren’t real". I think the best part of this is that when you use docker to "create" a container (`docker container create`), it mostly just creates the container's rootfs directory (it's actually multiple directories combined using an Overlay Filesystem) and its configuration file. The container doesn't actually "exist" until the process starts which is a separate command (`docker container start`) because Docker doesn't actually instantiate the namespaces or launch the container task until start time (and generally in Linux, namespaces are tied to living processes unless explicitly held open via bind mounts or another process reading a `/proc` file).

[1] https://github.com/moby/sys/tree/main/reexec

some people do all of their software development in a container. It really only works well if you are comfortable using TUIs instead of desktop apps though.

> Docker’s container runtime, runc, is written in Go.

I worked at Docker from 2014-2019. The most important part is actually written in C.

[I wrote a whole comment here about to explain how the hack works because I was too fixated on how technically incorrect "written in Go" is. Then I went back to read the rest of the article only to see that the paragraph immediately after this kind of explains it in its own way! Let that be a lesson to always finish reading the article before posting about it.]

the chain of thought is not auditable. The best you can do is have another model which later maps the intermediate latent states at every thinking iteration to text describing that latent state, but it's going to be lossy.

> without parentheses

> \x28

> \x29

does this not count as using parentheses?

I like how he just casually posts a photo of over $30k worth of equipment on his disk like "this is accessible to everyone!"

here you go: [1]

In 2022, I worked on a project to build a climate model that runs in your browser. I got as far as I could until the math and physics became too difficult. Now 4 years later I figured that if frontier models can solve open questions about Navier-Stokes equations then it can probably help me fix my climate model and it has been great! I also worked with it to create this 10-part series (still in progress) to explain how the climate (and the model) works.

[1]: https://webgcm.app/explain/air

I love this blog so much that I used it as a guide for building my own series of explainers for how a general circulation model of the earth's oceans and atmosphere works: https://webgcm.app/explain/air

I put nowhere near as much care into the interactives but the inspiration for the format comes 100% from the Bartos Ciechanowski blog.

> LLMs are powerful autocomplete with vast and broad memory, they do not think.

This is really only true insofar as they are designed such that they output language tokens and their intermediate "chain of thought" is also done in language tokens, but now that's really only done for auditability. There are much more abstract layers before the model output layer of a single token prediciton and they could be designed to to "think" using these more abstract layers as intermediates instead like how many people think in concepts rather than words. There are even some research models on this (see Next Latent Models)[1]. Humans need to communicate in language because we don't have telepathic abilities but these models need not have such a constraint.

[1] https://arxiv.org/abs/2511.05963

Just to clarify, neither of those projects (2006 and 2013) was done by Fervo. Fervo wasn't even founded until 2017.

The Basel project caused earthquakes because it used a single, vertical well to pump high-pressure water into dense granite all at once, which accidentally over-pressurized and lubricated a major existing fault line. In St. Gallen, the earthquake happened when drillers unexpectedly hit a high-pressure methane gas pocket. The rapid injection of heavy drilling mud used to control the gas kick built up intense, localized pressure that triggered a nearby fault.

Fervo's modern method avoids these issues by using horizontal drilling to distribute fluid pressure gently across miles of rock rather than spiking it in one spot. They also use advanced fiber-optic cables for real-time micro-seismic monitoring, allowing them to sense tiny rock movements and adjust fluid rates instantly before a felt earthquake can ever develop.

I have so far only completed about 14 minutes of the talk, but I would imagine it to be some sort of relatable analogy which foreshadows what the thesis of the talk is about. The "Christian Parents" are the authority figure and following their rules, lessons, and expectations for dating is a lot like doing only the software engineering tasks that management prioritizes for you. You end up unsatisfied with your life to the point where you become jaded about software (software development is like "dating" in this analogy) instead of learning the skills necessary to wield influence over what you get to work on. I'll watch the rest to see if I'm right.

edit: it's wasn't just about developing software but also using software. You don't have to accept the bad situation you're in (or the bad software you've been given or told to use). Finally, I'll say that while the speaker's comment about conservative christian upbringing isn't a universal one, it is a widely shared one. If you take offense by it maybe take a moment to pause and consider that maybe the way you were exposed to it was probably more balanced than what many others experience, or in the worst case that maybe your negative reaction to it is a sort of defense mechanism that you've learned to shield yourself from confronting the reality that your own conservative christian upbringing was not so good after all?.

y'know what's cooler than Perlin Noise? Simplex Noise [1]! This is what is hinted at at the end of this article. Also invented by the same person to address issues in the original Perlin Noise function. It's not as popularly known because it was restricted by patent issues for so long but as of 2022 the patent has expired!

1. https://en.wikipedia.org/wiki/Simplex_noise

the only thing I can imagine in the near-term are a subset of vehicles in the fleet which have carseats for infants and todlers which stay installed in the vehicle. We also already see vehicles from Volvo which have integrated/convertable booster seats for children 4+ and maybe the entire fleet could have these.

> The global leader in self driving cars has a headquarters thirty miles from here - the city could invite someone from Waymo to give a talk on the future of onsite parking and how our codes should adapt.

I agree with pretty much everything in your article, Kevin. But I wonder whether Waymo actually has a holistic vision for how families would use their service? Do they expect people with small children to haul around a carseat or booster seat (possibly multiple) for these trips until all their kids are old enough? And spend the extra couple of minutes double parked while they install/remove it?

Michael Crosby wrote this! He's a long-time maintainer of Docker, Containerd, and more! He was Docker's first to receive the 'Distinguished Engineer' Title. This means a lot coming from him.

they specifically said "Splotlight Semantic Indexing" which means they generate embeddings of all your "personal context" and store it in a local vector database so they can do on-device RAG.

that's interesting because my understanding was that a lot of models used pressure as the vertical (height) coordinate and with tracking temperature at any given coordinate lets you know the density of the air at any point.

Atmospheric circulation. hadley cell, polar cell, and mid-latitude cells. It would also be interesting to see how new bands would occur if we increased the rotational speed of the earth, thus increasing the strength of the coriolis effect. It would also be neat to draw your own continents and orography and see how that impacts climate. which places become more wet/dry, etc. Change how much long wave radiation is absorbed by the atmosphere as the composition of the atmosphere changes. I'm not interested in actually making weather or climate predictions but using it as a tool to educate people on how the climate works.

The title should be updated to note this is from 2006.

Has there been anything like this published in the past 20 years but for compressible fluids? I have wanted to make a simple atmospheric model for years but have been unable to because of the complexity and probably my lack of complete understanding.

I was waiting for him to make some point but then it ended up being that a wealth tax is like increasing regular income taxes by 20% which makes it seem like PG is trying to confuse people about what a wealth tax is designed to target. It's not targeting your wages or salary, or even your interest or dividends. It's primarily targeting unrealized gains on financial securities.

Maybe he should spend his time trying to work with these politicians to design something that is more fair? Like making it actually act as a tax on unrealized gains over $1B (so that it takes cost basis into consideration) OR make it so that if you need to sell some assets to pay the tax, you can writeoff the wealth tax you paid from your regular/capital-gains income so that you aren't taxed twice? There's a lot of actually useful stuff he could write about in this policy area instead of blogging about the financial equivalence between stocks and flows.

> Antigravity was the only autonomous agent that implemented the Pantheon’s signature interior ceiling pattern: repeated square coffers visible through the oculus.

That is seriously really impressive. I looked at the 3D model and didn't even thing to LOOK INSIDE the building before reading this.

Here's [1] the 3D model with `show_cutaway` enabled.

[1] https://modelrift.com/models/pantheon-benchmark-antigravity-...

while I understand that we are computing the tokens in parallel to get the "faster" result, is there a tradeoff where we're actually utilizing more compute resources by running multiple instances of the large model? That is, while it's faster, is it more efficient?

edit: doing some more of my own research, it sounds like the bottleneck in doing it sequentially is in shifting weights around in memory, so while it uses more compute it doesn't oversubscribe compute resources because the bottleneck is not in supply of compute but in supply and speed of memory. The GPU has a massive supply of compute but sequential decoding only demands a relatively small amount of it. Time is primarily spent waiting on loading values from vram.

This seems hyperbolic as it's actually a long path between age verification to full digital identity tracking. But I agree that pushing the burden of verification to websites is ridiculous. Like the GDPR requirements where every webpage has an annoying consent modal, the verification and preferences should be controlled on the device you use to access these digital services. My browser should know and enforce my cookie preferences in a way that has a uniform user experience. Likewise, if I am a minor, my parent should provide me with a device (or profile on a device) which knows my age and can use that to inform online services of the age of the user rather than needing to go through a separate process for each service.

they are saying that LA takes water from sources which would otherwise drain into the sacramento and san joaquin river delta. The video from this post mentions the California State Water Project which takes water from the Feather River (Oroville Dam) and distributes it along the Western edge of the central valley South to Bakersfield where it is then pumped over the mountains both towards Los Angeles and further East to San Bernardino and Riverside. It provides way more water to SoCal than the two Los Angeles-specific aqueducts from the Owens Valley on the Eastern side of the Sierras.

There were a few recent edits about this on Tony Hoar's Wikipedia page which were reverted because there was no substantial evidence: https://en.wikipedia.org/w/index.php?title=Tony_Hoare&action...

Here's the same post on Bluesky for those who prefer that platform: https://bsky.app/profile/josephpolitano.bsky.social/post/3mg...

3 things: Prop 13, Suburban Sprawl, and Bigger/Heavier Vehicles